David L. Shrier is a globally recognised expert on digital transformation.

He is Professor of Practice, AI & Innovation at Imperial College Business

School and co-Director of Trusted AI.
David helps public companies, private enterprise and more than 100
governments create ecosystems for new ventures in fields such as artificial
intelligence, data analytics and financial technology.
He and MIT Professor Alex Pentland personally helped revolutionise
how online classes are delivered by top universities, stimulating
entrepreneurial action across more than 150 countries and empowering over
20,000 innovators. David has fostered new business models for academia
that have generated over US$1 billion of financial support for Imperial
College London, MIT, Harvard and Oxford University.
David spends considerable time translating academic theory into
business practice, co-founding five MIT AI spinouts. His Visionary Future
venture studio both works with established companies to generate new
revenue, having advised on more than $10 billion of growth initiatives, and
launches new enterprises. His portfolio of university-related startups spans
artificial intelligence, new financial architectures and other disruptive
technologies.
His eight books since 2016 include, with Little, Brown: Basic Metaverse
(2023), Augmenting Your Career: How to Win at Work in the Age of AI
(2021) and Basic Blockchain (2020). He also has published multiple books
with MIT Press on advanced technology topics. David holds an Sc.B. from
Brown University.
‘David L. Shrier explains digital transformation in an extraordinarily
intelligent and easy to understand way. David combines his amazing
experience in both the private and the public sector, as well as in academia.
Basic AI is a must-read’
Marc Walder, CEO, Ringier, AG and founder, digitalswitzerland

‘David Shrier’s new book has come at just the right time. A thought-leader
in this space for decades, in Basic AI David provides business leaders with
the context and decision-making frameworks they need to help them
navigate the challenges, opportunities and risks of AI. This is an accessible
and actionable book that cuts through the hype to get to the heart of what
CEOs and their teams need to know to prepare for the disruption that is
already happening’
Margaret Franklin, President and CEO, CFA Institute

‘Want to get up to speed with AI? This is the book for you. Filled with clear
explanations, real-world examples and helpful chapter summaries, David
Shrier takes us on a whistle-stop tour to the very heart of the future now.
The author eloquently sets out the myriad of challenges but, ultimately,
rightly, leaves us with confidence in our human selves or, as he puts it,
“optimism tempered with caution”’
Baron Holmes of Richmond, House of Lords Science and Technology
Select Committee

‘An indispensable and excellent book that should be a must-read for anyone
who needs to get to grips with the AI revolution we are living in. David
Shrier is an expert guide to what has happened, what is happening and what
is likely to happen in future. His knowledge and experience make Shrier an
ideal person to make the case for the vast opportunities that AI will bring to
the world. A brilliant read that should become a contemporary classic – it
leaves the reader not only better informed but better able to comprehend our
future reality’
Rt. Hon. Chris Skidmore, MP, former UK Minister of State for Universities,
Science, Research and Innovation

‘Basic AI stands as a testament to the pressing need for understanding,

adaptation and strategic action in the burgeoning age of AI. This is not just
a book; it’s a compass for the AI-intertwined future. It beckons readers to
approach AI with a blend of optimism and vigilance, emphasising our
collective agency in sculpting a harmonious future where AI augments,
rather than supplants, human potential’
Cyrus Hodes, cofounder of Stability AI
 Also by David L. Shrier:

Basic Metaverse
Augmenting Your Career
Basic Blockchain

As editor:

Global Fintech
Trusted Data, Revised and Expanded
New Solutions for Cyber Security
Frontiers of Financial Technology
Basic AI
A Human Guide to
Artificial Intelligence

David L. Shrier

A How To Book
 Copyright

Published by Robinson

ISBN: 978-1-47214-897-1

Copyright © David L. Shrier, 2021, 2024

The moral right of the author has been asserted.

Illustrations © David L. Shrier

All rights reserved. No part of this publication may be reproduced,

stored in a retrieval system, or transmitted, in any form or by any
means, without the prior permission in writing of the publisher.

The publisher is not responsible for websites (or their content) that
are not owned by the publisher.

Robinson
Little, Brown Book Group
Carmelite House
50 Victoria Embankment
London EC4Y 0DZ

www.littlebrown.co.uk
www.hachette.co.uk

How To Books are published by Robinson,

an imprint of Little, Brown Book Group. We
welcome proposals from authors who have
first-hand experience of their subjects.
Please set out the aims of your book, its
target market and its suggested contents in
an email to [email protected]
To those who believe in a better tomorrow
Contents
About the Author
Praise
Also by David L. Shrier
Title Page
Copyright
Dedication
Introduction
Part I: AI Changes Everything
Chapter 1: Rise of the bots and AI hallucinations
Chapter 2: Defining AI: understanding terminology
Chapter 3: Evolution of AI jobs displacement: how did we get here?
Part II: Playing Defence
Chapter 4: Reskilling and developing cognitive flexibility
Chapter 5: Future-proofing: what to study and which industries are safer?
Chapter 6: What can Human+AI systems do? Adventures in collective
intelligence and more
Part III: Centaurs and the Future of Humanity
Chapter 7: Building Human+AI systems: from prompt engineering to
high-performing teams
Chapter 8: Alternative futures: Wall-E or Star Trek?
Chapter 9: The urgent need for policy: shaping a better tomorrow
Conclusion
Glossary
Endnotes
Acknowledgements
Index
‘Imagination is the only weapon in the war with reality’
— Lewis Carroll, Alice’s Adventures in Wonderland
Introduction

We live in interesting times.

Bear in mind that the period of the Black Death was an interesting time,
as was the Hundred Years War. Of course, the Italian Renaissance was also
an interesting time and so was the Golden Age of China.
What is clear is that we are in a period of historic change, as human
society faces a transformative force unlike any other it has previously faced,
a new technology that perhaps may more profoundly affect the course of
human history than the invention of the wheel, agriculture or the steam
engine.
I speak, of course, of artificial intelligence (AI).
Already, AI has changed the course of global politics and, for better or
worse, threatened US hegemony. It has introduced new uncertainty into
financial markets. It has suddenly become a topic of conversation in
boardrooms around the world.
Every week brings with it a new story of a major corporation or even
nation making or anticipating significant changes brought about by AI.
While I was writing this book, the CEO of BT Group (formerly British
Telecom) announced that he intended to replace 40 per cent of his
workforce with AI, reducing it by up to fifty-five thousand people by 2030.1
Discussing the final draft with Michael Foley of MAKR Venture Fund,
he revealed he and his colleagues were about to raise $1 billion (all
references to dollars are to US currency) for a new venture capital fund
focused on AI, and had already negotiated his first sixty investments.
As I finished a few final chapter summaries, news came out that a lawyer
with thirty years of experience was facing potential sanctions for having
submitted a brief created by the generative AI system ChatGPT — which
included fake citations about cases that never existed.
I wrote my first AI software program in 1991. It wasn’t terribly exciting;
I was using software in order to simulate what a particular type of neural
network computer would do if it, in fact, existed. Even twenty years after
my first stumble in the dark of AI code, in 2011 AI was interesting
technology but fairly arcane and challenging to navigate.
Two events occurred within the past few years that changed everything.
The first was largely unnoticed outside of technology circles: in 2015,
Google published a powerful library of AI technology known as
TensorFlow into the open source.* Now, large numbers of programmers had
access to sophisticated AI models and capabilities that previously were
restricted to big tech, big academia, some governments and a handful of
well-funded startups. This gave rise to a generation of startups whose
technology is now coming out from development.
The second event, which was a direct consequence of the first, has been
making regular headlines since December 2022. OpenAI, a company that
Elon Musk helped form with Y Combinator impresario Sam Altman and a
few others, released a public-facing interface to its large language model
(LLM) artificial intelligence. Known as ChatGPT, this system served as an
electric jolt to the nervous system of business and society at large. Rapidly
adopted by people numbering in the hundreds of millions, it revealed the
potential of AI in a way people had not previously been capable of
understanding. It brought the power of advanced AI to people in a fashion
that was as easy to use as a search engine.
Our world is changing as a result of this AI revolution. Speaking for my
own advisory practice at Visionary Future, the enquiries began coming
weekly, now daily, from all corners. Everyone from a top-ten investment
bank to one of the fastest-growing tech companies in Europe has reached
out to seek guidance on how to apply AI, what the new and emerging
developments mean in the context of their businesses and those of their
clients, and what new competitive threats might come about as a result of
this disruption.
The private sector is only part of my life. I’ve also dedicated a significant
portion of my life to public service, in my effort to apply technologies to
solve problems at scale.
In my work in academia, my colleagues at Imperial College London and
I, as well as those at the University of Oxford, the Massachusetts Institute
of Technology, the University of Edinburgh and elsewhere, share a common
concern that in the race to adopt AI, we need to ensure the direction we take
is good for humanity. We are working to bring about AI systems we can
rely on, that will improve the lot of everyone. We have formed the Trusted
AI Institute to promote this idea, an initiative spanning more than 3,000
researchers (to start with) who are ready to tackle this problem. It’s
imperative to the future of the human race because, if we get it wrong, we
could create catastrophic consequences for society. Worse than the Black
Death, worse than the Second World War, more far-reaching than the
Industrial Revolution. If we get it right, we have the potential to solve some
of the biggest and most intractable problems facing humanity.
With this in mind, I seek to provide you with a handy field guide to the
world of AI, so you can inform your own decisions about what to do, how
to go about it, and what the risks are of inaction.
This book is brief, of necessity, although I have provided some
guideposts if you wish to explore more deeply.
Let’s start with the basics …
 Part I
AI Changes Everything

Part I focuses on the fundamentals: what is AI, how has it impacted us thus
far, and what risks does it present in the workplace?
CHAPTER 1

Rise of the bots and AI hallucinations

CHAPTER 1: WHAT YOU NEED TO KNOW

♦ ChatGPT has upended the AI playing field, acquiring hundreds of

millions of users in a matter of months.
♦ AI systems have already had significant geopolitical impacts, such
as on Brexit and the 2016 US Presidential election.
♦ Caution is warranted with AI, since it can be prone to
‘hallucinations’.

‘What are you going to do about ChatGPT?’

The then-dean of Imperial College Business School, Francisco Veloso,
asked me the question almost casually when I stopped by his office. It was
January 2023. The storm clouds of artificial intelligence were breaking
open over the staid world of academia. Since my professorial title says ‘AI
and Innovation’, and my signature class that I created for Imperial teaches
students how to launch new AI startups, I was expected to have an answer.
Others in university circles were up in arms. Ban the technology! Use
plagiarism software to detect if students were ‘cheating’ with ChatGPT!
Reactions from professors worldwide were strong.
ChatGPT is a consumer-facing service that allows you to engage in
conversation with an AI and have it not only give you answers to questions,
but actually create content for you. Instead of the functional but still
somewhat awkward vernacular of search engine technology, where you
have to dive down a rabbit’s warren of links in order to extract meaning and
substance, ChatGPT does the heavy lifting for you. ChatGPT can write
software code or an essay in response to a prompt from a business school
professor in the blink of an eye.
Created by Silicon Valley phenomena OpenAI, ChatGPT has been the
most successful new consumer technology launch in history. As AI angel
investor Allie K. Miller noted in a LinkedIn post, it took Netflix about
three-and-a-half years to reach 1 million users, while ChatGPT got there in
five days1 after its launch in December 2022. The service reached 100
million users in under two months.2 It seems likely they will hit the 1
billion-user mark by the time this book is in your hands.
Back to our impromptu meeting in January 2023. Some professors
struggled with the implications of this AI tool. Test preparation company
Chegg saw its stock price plummet 48 per cent in a single day when the
CEO described the impact of ChatGPT on his forward projections. IBM
announced a halt on new hiring and Samsung stated it would ‘ban’ the
technology.3
The theory and practice of creating AI innovation, which is the substance
of the ‘Imperial AI Ventures’ class, suddenly became very personal, as
these disruptions began to roil higher education. My two faculty colleagues,
Chris Tucci and Yves-Alexandre de Montjoye, and I had constructed a
curriculum that provides carefully wrought scaffolding to help our mix of
business and computing students to architect a new enterprise in the AI
domain. A series of written assignments prompted students to assemble
components of their business plans, which they would then present at the
end of the module.
Normally this sort of work entails a reasonable amount of time to think
through, research competition and market, and articulate a cogent set of
activities around launching the new venture. Now, thanks to ChatGPT,
students can complete the assignment in five minutes instead of five weeks.
We had to come up with a new approach, since it appeared inevitable that
students would avail themselves of ChatGPT in completing their
assignments.
Dean Veloso suggested to me that we try the critical thinking approach to
how we design our assessments, and off we went: retooling an entire
curriculum in a week.
What we came out with at the end of this condensed process was a
decision to explicitly require the students to partner with their large
language model buddy, and tell it to write portions of the business plan. The
students then engage in critical thinking, examining the output from the
LLM (usually ChatGPT, but we suggested a dozen alternatives),
highlighting what it did well, and explaining what they could do differently
in order to make the output usable in their final presentation. Tempted
though we were, we did not make use of the generative AI tools for
university professors that we uncovered in our research, which can make
lesson plans, design assessments and conduct an array of other tedious,
time-consuming tasks associated with creating a detailed curriculum.
We were lucky enough to get an OpenAI executive to come speak to our
eighty highly engaged students. While the contents of the discussion are off
the record, I can share that it was exciting to be able to expose the class to
history as it was being created, a behind-the-scenes look at a new industry
being shaped in real time. I couldn’t have asked for a better case study. It
reminded me of the heady days a few years ago when MIT Professor Alex
(Sandy) Pentland, Joost Bonsen and I created the first graduate fintech class
at MIT, bringing in the CEOs who were shaping the headlines and allowing
our students direct Q&A about the new world that was opening before
them.
Something interesting happened through our learn-by-doing exploration
of generative artificial intelligence (a topic we will define further in
Chapter 2). While one student did unearth an important gap in their
business strategy using ChatGPT, our students discovered that it has
significant limitations. It isn’t creative. It tends to repeat itself. It is not, yet,
a complete substitute for a human entrepreneur.
Innovation economist Erik Brynjolfsson at Stanford’s Graduate School of
Business has argued that the AI revolution will not lead to wholesale jobs
displacement, but rather jobs substitution or jobs evolution.4 In 1965 (or
1865) we didn’t have a concept of an ‘IT Help Desk Specialist’ or a
‘Machine Learning Developer’. In this new age of AI new job
specialisations are emerging. Five years ago, we didn’t have a job called
‘Prompt Engineer’ (someone who knows how to write queries for LLM
systems like ChatGPT). Today, there are dozens of postings on employment
websites. The revolution is at hand.
Erik further argues that true economic expansion will occur when AI
systems augment or enhance humans, rather than replace human tasks. We
investigate this idea of Human+AI hybrid systems in Chapters 6 and 7.
WHEN BOTS ATTACK
Let me not paint an unduly rosy picture of the AI present and future. We’ve
already seen the adverse impact of unregulated AI on society.
Artificial intelligence in concert with human beings has changed the face
of the political landscape. But first let’s review recent history.
Robotic automation has been reshaping manufacturing for years. The late
1970s brought some of the first large-scale deployments of robots to the
factory floor. It contributed to the shift of the British and American
manufacturing economies into service economies. What’s interesting here is
how waves of AI and robotic automation created fundamental changes in
each country’s society.
First, the machines took jobs away, permanently. Whether in Yorkshire or
in the US Rust Belt, in places like Sheffield and Detroit, a generation of
workers was displaced by technology and by globalisation. These people
were, to all intents and purposes, made permanently unemployed, as their
skills weren’t transferrable and their employers, and society, failed to
provide sufficient retraining. For every robot, there is a robot repair person,
but those jobs need to be created and workers need to be trained in those
new skillsets. During the Brexit vote of 2016, Sheffield, originally expected
to vote Remain, shifted to Leave. How did this happen?
Brexiteers might tell you it’s because Nigel Farage convincingly
campaigned in the region.5 And without a doubt, he had some influence on
the outcome. But something else was going on as well, which helped propel
the momentum that he, Boris Johnson and others pushed.
More than a hundred and fifty thousand social media accounts engaged
in Brexit-related activity were linked to Russian intelligence and affiliated
entities,6 which together with Iranian agents generated millions of tweets in
support of Leave. It is believed one-third of all Leave tweets in the month
immediately prior to the election were generated by bots exhibiting
characteristics common of Russian troll-farm activity.7 Cyber firm F-Secure
has tracked continued activity years later in support of Brexit, identifying as
many as 18 million suspect tweets in less than two months in 2019.8 The
Russia Report on the attacks, finally released in July 2020 more than a year
and a half after it was completed, outlined indisputable evidence of Russian
meddling in UK politics, including specifically around the Brexit vote.9 AI,
albeit under the firm control of human hands, was directly responsible for
promoting messages that encouraged fear and xenophobia.
Would people have listened as intently if they hadn’t for decades felt ‘left
behind’? The anger expressed at the polls was genuine, and came, in part,
from people fed up with having no discernible future, many of whom were
unemployed and disenfranchised.
The bots would not have been able to swing the elections if it were not
for a disaffected, alienated electorate that made both the Brexit vote and the
2016 US Presidential election exceedingly close. Without a sense of
participation in the economic growth that followed the 2008 recession,
without the chance to enjoy wealth creation of the fruits of the European
Union, why would these left-behinders feel optimistic about the future and
positive about the promise of more of the same? Because more of the same
was, essentially, the core political platform of both the Remainers and the
Hillary Clinton campaign.
By themselves, these disaffected voters were statistically a minority of
the voting-age population. But what happened when they were combined
with outside influence that suppressed the mildly positive majority and
encouraged extremists in the minority? We have witnessed the results,
amplified perhaps by the COVID-19 crisis but apparent even without it.
Brexit has sometimes been referred to as the ‘dress rehearsal’ by Russian
intelligence for the US Presidential election. And indeed, we see
proportionate results: in the 2016 US Presidential election, more than 40
million ‘fake news’ tweets propagated on Twitter.10 A ‘significant portion’
of users were deemed to actually be bots, responsible for between 20 and 25
per cent of all of the content.11 Imagine if one out of every four or five
people you talked to was a robot. Imagine if you couldn’t tell the difference.
Researchers from the National Bureau of Economic Research estimated
that the Leave vote in the UK during the Brexit referendum was increased
1.76 per cent through the actions taken by bots, and Donald Trump’s votes
in the US Presidential election in 2016 were boosted 3.23 per cent. Leave
won 51.9 per cent to 48.11 per cent. Trump actually lost the popular vote in
the United States, but due to a quirk of how the US Presidential system
works, his wins in strategic locales meant he was able to take the White
House. The bot swing was within the margin of error of the estimation
techniques of the researchers.12 University of California Berkeley’s Yuriy
Gorodnichenko and Swansea University’s Tho Pham and Oleksandr
Talavera stated: ‘Our results suggest that, given narrow margins of victories
in each vote, bots’ effect was likely marginal but possibly large enough to
affect the outcomes.’
You’re not paranoid if they are actually out to get you. The bots attacked,
and they continue to do so.

AI HEAD-MEDDLING
How could we have arrived at a point where millions of people may have
had their votes swayed in national elections? It’s enough to make you
depressed. Let’s consult a psychiatrist.
‘I’ll be your therapist today.’
‘I feel sad.’
‘Tell me more about such feelings.’
‘My children never call me.’
‘What does that suggest to you?’
By now, you have probably figured out that this is a primitive AI. In the
mid-1960s, however, when MIT researcher Joseph Weizenbaum first put
the DOCTOR script into the ELIZA expert system, it was a revolutionary
moment in computer science. This was the first chatbot that successfully
impersonated a real person, passing what’s known as the ‘Turing Test’. If
you’re curious, you can play with ELIZA yourself at websites like this one:
https://web.njit.edu/~ronkowit/eliza.html.
Mathematician and pioneer of computing Alan Turing, who cracked the
Enigma code at Bletchley Park to help the Allied forces win the Second
World War and who was deemed the father of artificial intelligence,13 posed
a benchmark for the development of AI. He said that if a computer program
could be written that could interact in such a manner that someone couldn’t
tell if they were interacting with a machine or a person, it would pass the
first threshold of demonstrating that it’s a machine that thinks like a human
being. It would, definitionally, be classified as ‘artificial intelligence’. We’ll
investigate more about the different types of artificial intelligences in the
next chapter, but it’s worth considering the point that the dawn of AI began
with a chatbot.
In modern times, we see chatbots providing customer service on
websites, tirelessly, benignly, without ever getting annoyed at dumb
questions. We also see them invading dating apps, fronted by repurposed
pictures of models or simply stolen images of attractive people, luring us
into subscribing to adult websites or, more malignly, conning lonely hearts
out of hundreds of thousands in savings. The invasion is so prolific it has
prompted numerous articles with titles like ‘How to Tell If You’re Talking
to a Bot: The Complete Guide to Chatbots’ and ‘Spot the Bot: Keep Bots
from Taking Over on Dating Sites’, and even a private investigation service
offering a piece on ‘How to Spot Scams & Bots on Tinder and OkCupid’.
The bots are coming after our bedrooms. The Nigerian prince scam is now
interactive.
The origin story of all of these scams was a humble parody of Rogerian
therapy (in which the therapist is effectively passive, asking questions that
reflect what the client has already said rather initiating new areas of
discussion). The first bot was intended to illustrate how inane (in the view
of the authors) this particular brand of psychiatry was. Users were
absolutely convinced that ELIZA had feelings and could think. Our
tendency as a species to anthropomorphise inanimate objects found a new
model of expression that synchronised with technology engineered to
emulate human beings. Perhaps it was only a matter of time before we
would go from pouring our hearts out to an unthinking robot (one that was
actually written as a parody of psychiatrists, dumbly parroting back what
we said in the form of a series of prompt questions) to simply mistaking
political chatbots for actual people. People who think just like we do, even
down to the darkest corners of our psyches. But rather than soothe our souls
by ‘listening’ to our troubles, these chatbots encouraged polarisation to such
a degree that society fractured. The foundations of Western democracy
grew a bit shakier with these Manchurian chatbots, albeit ones that weren’t
fully autonomous, but were tuned and contextualised with the aid of human
intervention.
Chatbot-enabled electoral polarisation didn’t happen in a vacuum.
Another, more subtle, form of artificial intelligence systems had already
been working to push people apart in the form of the Facebook feed. People
like people who think like them. If you promote this positive feedback loop,
you can create an information or behavioural cascade that suddenly gets
large numbers of people moving in a certain direction. Demagogues have
known this for centuries. With the advent of AI, with the ability to make
chatbots that attack, this knowledge can now be made predictable and
scalable, with terrible implications for society.
 Witness the polarisation of the American electorate over the course of
about twenty years. Donald Trump did not spring fully formed from the
brow of Ronald Reagan. The Trumpists of today would have found Reagan
to be a lefty, hopelessly liberal in their eyes. The Republican Party of the
1980s proudly talked about the ‘Big Tent’ that could encompass many
views. They had to – at the time, that was how you won elections, by
appealing to the centre. What happened to set up the current dynamic,
which plays itself out as much in the streets of Birmingham as it does in the
hallways of Washington, DC?
Let’s look at some numbers that reveal what polarisation looks like in the
American electorate:

Source: Pew Research Center 2014

As you can see from the chart, the political leanings of Americans
gradually shifted from having a fairly cohesive central mass to separating
into increasingly polarised ‘liberal’ and ‘conservative’ camps. From 1994 to
2004, despite the fragmentation of media, rancorous political debate and a
war in Iraq, Democrats and Republicans were relatively convergent and
there was a political centre. The blame could not be laid on Fox News,
which was founded in 1996. By 2015, the centre could not hold and things
flew apart. Arguably things have become even more extreme in the United
States since 2015, to such a degree that Democratic mayors in large cities
were fighting armed invasion and abduction of citizens off the streets by
unmarked government vehicles at the instruction of a Republican federal
administration.14
 What happened in 2004, you might wonder? Facebook was founded.
There are some who may criticise me for confusing correlation with
causation. I am not alone in my contention that, by legitimising fringe
extremist sources, AI ‘newsfeed’ algorithms and unconstrained chatbots
were directly responsible for electoral polarisation.15
Facebook makes money through selling advertising, which is driven by
how many people look at the website and interact with its content. Through
AI analysis of our behaviour, Facebook discovered that more extreme
headlines give us a little thrill. When we get that charge, we stay on the site
longer and we click more – particularly on extreme articles that agree with
our intrinsic biases. Generally speaking, Facebook doesn’t care about
Democrat or Republican, liberal or conservative.* Facebook cares about
making money. More page views and more time on site means more profit.
Chatbots were weaponised16 and gleefully deployed into the fertile
environment created by Facebook and other receptive social media, in both
the UK and US (and to a lesser degree in France and other places). They
supported far-right extremists and far-left extremists. Anything to break the
political discourse. Intelligence agencies and government committees have
pointed to Russia, Iran and North Korea as sources of a number of attacks,
including use of weaponised chatbots and other attack vectors,17 up to and
including a series of hacks surrounding the 2020 US Presidential election.18
One analyst group called these ‘Weapons of Mass Distraction’.19
Chatbots do strange things when they are left unsupervised. Technology
titan Microsoft suffered a bit of a public relations disaster in 2016 when it
launched AI chatbot Tay. Digesting masses of data from Twitter, Tay
rapidly taught itself to be both racist and sexist. Microsoft pulled it down
within sixteen hours of launch.20 If a simple chatbot can so quickly go off
course in an unsupervised learning environment, what might happen with a
more sophisticated artificial intelligence in a mere few years, one that is tied
to critical systems? What if our self-driving car decided that it would be
happier without some smelly human ordering it around?
In a way, we are fortunate that chatbots are so dumb. Yes, they were used
to attack democratic systems. They were noticeably machinelike, however,
and so researchers and security professionals have been able to identify
them. Twitter (now called ‘X’) was until recently a progressive company
that proactively sought to neutralise these AI-fueled bots. Under current
ownership, those effects have been abated or eliminated. We may soon
reach a point in artificial intelligence systems development where we see –
or rather don’t see – imperceptible AI, invisible AI, systems that
convincingly create video and audio simulations of people that are so
accurate they are impossible to tell apart from the real thing. Then we can
really start worrying about what happens when bots attack.
Perhaps I ought to be more worried about the simple threat of the AI
author and thought leader that can read more than I can, write better than I
can, and produce work at a greater speed than I could ever hope to. Is there
anything we can do to avert a world of digital twins that displace us from
our jobs? Or are we destined to become obsolete, a faint memory in the
minds of the beings who created the machines that will replace us as the
dominant life form on the planet? How far away is this grim vision of evil
bots from the mirror universe, coming to take over our lives, steal our
money and ruin our reputations?
Thanks to generative AI, we now can easily produce highly realistic
‘deep fakes’, video and audio that seem as if they were coming from a real
person, but actually are manufactured by a machine. US President Joe
Biden can be made to chant an amusing rap or, more ominously, warn of
impending nuclear attack. The risks to society, to the foundations of
democracy, have never been greater.
In Chapter 9, we’ll discuss some of the emerging policy responses to AI
threats.

THE FEAR INDEX

Robert Harris wrote a 2011 novel, The Fear Index, that played with the idea
of AI actively promoting a feedback loop to spiral human fear out of
control, creating market disruption that the ‘evil hedge fund’ can take
advantage of. The HBO TV series Westworld envisions a war between
humans and humanoid machines. Part of the terror arises from machines
that are made to look like people, passing imperceptibly for human beings
to better infiltrate society, which then is subverted in service of the goals of
the machines. The Matrix movies, likewise, envision technology-gone-
wrong, as humanity is turned into mere batteries to power the machine
empire, notwithstanding the faulty physics needed to make that possible.*
 Fear of science and technology is nothing new. When hooligans tear
down 5G towers, they act out the same mindless savagery embodied in
Mary Shelley’s Frankenstein, where pitchfork-wielding villagers chase a
biological artificial intelligence to its death. Shelley based part of her story
on the legend of an alchemist who conducted research at Frankenstein
Castle centuries before she travelled through the region along the Rhine that
likely helped inspire her to write her opus.21 One could trace influences on
her, in turn, back to the legends of the Golem predating 1100 and forward to
the Isaac Asimov robot novels of the 1950s where he envisioned a set of
‘laws’ programmed into machines so they wouldn’t harm humanity. This
utopian vision of robotic safety was undermined in more recent films such
as Her and Ex Machina. Boy meets girl robot. Boy falls in love with girl
robot. Girl robot abandons boy and terrorises society.
AI, among all technologies, seems to have a special ability to inspire
fear, particularly in Western Europe and the United States. We rarely, if
ever, see movies or books about the dangers of quantum teleportation, or
headlines about politicians and workers railing about lithium batteries
exploding. Our imaginations are captured by images of machines running
amok, of killer robots and disposable humanity.
Partially, this fear may come from the way in which we have welcomed
AI into our lives. We have taken AI on to our phones, and use it to navigate
or to shop. We have taken AI into our homes, and use it to provide
entertainment – either explicitly, when we ask Alexa to play us music, or
implicitly, when we allow Netflix to recommend a movie to us. It’s
everywhere. Some people have taken to putting sexy or soothing voices on
to their in-home AI interfaces, and have a machine read poetry or erotic
literature to them. Or a bedtime story.
And we rely on AI, increasingly, to tell us what is what. This embeds a
new kind of danger, when the AI begins deviating from normative reality.

HALLUCINATIONS
An AI hallucination is an expression of artificial intelligence going ‘off the
rails’. Specifically, it means that the AI is doing things that it was not
trained to do — so much so that it enters into a kind of delusional state, and
begins making things up.
 For example, if you teach an artificial intelligence to speak colloquial
language, and it starts spewing nonsense. Meta describes it as when ‘a bot
confidently says something that is not true’.22
This reveals an important and fundamental characteristic of artificial
intelligence: its lack of reliability.
Thanks to popular culture, we are given to thinking that artificial
intelligence is either predictable or in its own mind rational. So, for
example, the artificial intelligence in 2001: A Space Odyssey attacking the
crew because it was given bad programming by humans. There’s a rational
explanation for why the AI stopped doing what it was supposed to be doing,
and started killing people. It did so because people messed it up with
conflicting instructions. The AI was trained to pursue truth and honesty, and
then people told it to lie.
Artificial intelligences are actually a great deal more complicated and
less predictable than how they appear in media. Some of the most powerful
approaches to artificial intelligence deliberately and directly design systems
based on the architecture of the human brain. The AI has layers upon layers
of linked networks, similar to how the human brain is composed of linked
networks of neurons. This layered approach of neural networks can make
AI extremely powerful. This is how ChatGPT works – what’s referred to as
a large language model built in a deep learning system.
It shouldn’t surprise anyone, then, that the AI misbehaving looks a bit
like the human brain misbehaving. And so we use a psychological term to
describe this erroneous behaviour: hallucinations.
Because ChatGPT and similar large language systems build sentences
based on relationships of prior words, the longer the piece you ask them to
write, the greater the chance you spiral off into some really odd directions.
However, if you iterate with the AI, a domain space sometimes called
‘prompt engineering’ (as we will discuss further in Chapter 7), you can
refine the outputs and get significantly better answers.
We continue to research, and train our AI researchers, on dealing with
model management, model drift and systems design. For example, there
was a major paper a few years ago on the concept of ‘under specification’23
– essentially, you train AI on one set of data and it gives you good results,
so you deploy it, but then it encounters the real world and you discover that
the training data wasn’t representative of real-world conditions, so the
model ‘breaks’, producing bad results.
 AI hallucinations are another source of AI error, one that is still poorly
understood by expert AI researchers and remains an area of intense study –
one that needs more funding; AI is still very much an emergent technology,
eighty years after we built the first AI during the Second World War. With
AI hallucinations, the AI will not only make up fake ‘facts’, it will even
fabricate citations (fake scientific articles) in order to support its delusion.
This might include clever fakes: actual authors and real scientific journals,
but fictitious articles. Caveat emptor, indeed.
What this means, in practical terms, is that we should be very wary of
undue reliance on generative AI to the exclusion of human involvement.
Let’s take a minute to look at what the different kinds of artificial
intelligence are, and where the state of the art stands today. That’s the
subject of our next chapter.

CHAPTER 1: THE RUNDOWN

♦ Developing a strategy to understand and respond to AI has never

been more urgent due to the rise of generative AIs like ChatGPT.
♦ AI has already had global scale impact on society in the past
decade, and more change is coming.
♦ The first AI to fool humans into believing it could think (passing
the ‘Turing Test’) was invented in 1965. They have only become
more sophisticated since then.
♦ It’s important to understand artificial intelligence, because it is
more complicated and less predictable than how it is depicted in
the media.
CHAPTER 2

Defining AI: understanding terminology

CHAPTER 2: WHAT YOU NEED TO KNOW

♦ Categories of AI include: expert systems, machine learning, deep

learning, generative AI and artificial general intelligence (AGI).
♦ Dimensions to consider include: structured vs unstructured data,
supervised vs unsupervised learning.
♦ Human+AI systems open up new potential for societal and
economic advancement.

It can seem, to the layperson, that there’s a bewildering array of

terminology that gets used in the description of artificial intelligence. It’s
important to understand the different kinds of AI because that can help
reveal the threats, and opportunities, that AI systems present.
We’re going to understand in this chapter, in relatively non-technical
language, the bestiary of artificial intelligence. Better to think of it as an
unarmed safari than a trip to the zoo, because some of these AIs pose a very
real threat to your future livelihood. Others, on the other hand, could
represent utopia. While the early chapters of this book, including this one,
generally paint a grim view of the next ten to twenty years, I enjoin you to
recall that once Pandora loosed all of the ills on the world, what remained
was Hope.

EXPERT SYSTEMS
Rules-based expert systems and other kinds of rules-based computer
systems were some of the earliest AIs. ‘If A, then B’ is the central
programming concept. You get a smart human to imagine all of the possible
answers to a set of questions or circumstances, or perhaps you create clever
mimicry in the case of ELIZA and newer generation rules-based chatbots.
The machine is following a big set of rules that deterministically drives its
actions: if a certain circumstance is presented, take one action; if another
circumstance is presented, take another, and so on.
The reason most chatbots today seem dumb is because they are. They
qualify as artificial intelligence by the most basic instance of the definition.
An important myth to puncture is the notion that ‘artificial intelligence’ is
automatically smarter than humans. Chatbots are typically following a
variation of a table of rules, sometimes randomly accessed (such as in the
case of some Tinder scambots), sometimes tied directly to a discrete set of
answers like in an online help system. Often minimally programmed, they
are as limited as their list of answers to specific phrases or words. I would
argue that the reason ELIZA worked better than expected by its creator is
that it was put into a very specific context, the therapy conversation, and it
was mimicking a very specific type of therapy, Rogerian, which consists of
asking questions back to the patient following a near-formulistic model.
ELIZA has been the inspiration for many of today’s modern chatbots, at
least in spirit, but chatbots that use this model are only as good as the
particular programming (and programmers) involved. It is possible to
manufacture smarter chatbots, as we will learn in Chapter 8, but many of
today’s chatbots are primitive and rules-based.
More serious expert systems have considerably more investment placed
into them, nonetheless following the same principles of a programmed set
of rules that trigger specific actions or results based on specific inputs. The
catch is that programmers have to spend quite a bit of time in a structured
capture process, gleaning information from experts in order to program an
expert system. This, in turn, limits the applicability of these types of AI.
Research on the performance of expert systems shows that their
effectiveness is influenced by such factors as the narrowness of the
questions being answered, the manner in which the information architecture
is designed, and the people involved (both expert and programmer). The
type of knowledge itself limits the viability of the expert system.1
What would be a viable expert system, versus a novelty? One example
would be configuring computer hardware and software.2 This used to be a
task that needed extensive manual labour by a highly trained IT
professional. It is, however, a constrained system: there are only so many
types of computers, and only so many possible combinations of hardware
and software. This is a dynamic that lends itself to codifying in a taxonomy
of rules and specific scenarios; if faced with system X, then make software
perform function Y.
While posing a modest threat to the livelihood of the IT department, a
second-order effect of removing the direct intervention of a highly trained,
on-site IT professional was that it made it more feasible to move IT
functions offsite. Eventually, even offshore: AI-enabled labour disruption
twinned with the trend of globalisation. The downstream effects have been
profound in many ways. In 2017, I was one of the twenty-thousand
passengers caught up in the mess when the entire British Airways computer
system went down across the globe. Stranded at Heathrow, I was fighting to
get across the Atlantic so I could give a speech to the central bank of Chile.
Air Canada eventually rescued me. It turned out that outsourcing BA’s IT
function was not perhaps the best systems reliability strategy.3 I was given
to understand that a poorly trained outsourced systems engineer mishandled
some critical power systems, and since they had made redundant the
internal BA IT people who knew how to handle the problem, there was no
one to call. They essentially broke the on/off switch. Prophetically, British
Airways had another notable IT failure while I was attempting to finish this
book, cancelling my flight out of Mumbai in the middle of the night.
What if you run into complex systems? What if you aren’t sure what you
need to do with a different set of circumstances? Expert systems
definitionally rely on frameworks that can fit within the scope of the human
mind. What if you are faced with gigantic amounts of data that defy the
span of human cognition? The limitations of expert systems, coupled with
growing compute power, the plummeting cost of memory and the
accelerating pace of data created or captured by sensors and sensing
platforms large and small, led to the emergence of another form of artificial
intelligence: machine learning.

MACHINE LEARNING
Machine learning (ML) is a more sophisticated form of artificial
intelligence than expert systems. Machine learning programs learn from
data and tend to get more accurate as you feed them more data. When I say
‘more data’, I mean very large quantities of data. It is perhaps no
coincidence that machine learning systems really began to come into their
own within the past ten to fifteen years, as the quantities of data generated
from internet usage – from satellites, from wearable computers, from
mobile communication networks and other sources – started to achieve
truly massive volumes.
Let’s pretend it’s 2014, and you are building a new facial recognition
system using machine learning. With a limited data set, you can get to about
96 per cent accuracy. Great. It still means that one time out of twenty, it’s
going to get it wrong. Thanks to the masses of data being pummelled
through facial recognition systems, by 2023 the accuracy rate is now closer
to 99.9 per cent.4
One of the issues is that the system needs to be able to deal with data
variety. The difference between 2014 and 2023 is that enough images have
been fed through ML algorithms that they are now much better at dealing
with so-called ‘edge conditions’ – those rare events that only show up
occasionally but are enough to mess up your entire model. For example, if
you have trained your model only on people without beards, or only white
people with beards, your system might struggle with someone who has a
different skin tone or facial hair. If your model never encountered close
family members that resembled one another, it might never learn how to tell
siblings apart.
This concept of edge conditions runs throughout computer modelling; it
is not limited to facial recognition. Steve Wozniak is reportedly worth more
than $100 million, but when his wife applied for a credit card, she was
purportedly given a credit limit ten times lower than his, as the system had
no concept of how to contemplate a billionaire’s wife.5 In the 2008 financial
crisis, and when the Long Term Capital Management hedge fund collapsed
a decade earlier, computer models were confronted with so-called ‘Black
Swan’ events that cascaded through markets and triggered catastrophic
losses.
This same concept holds true for teaching machines to translate words
that humans express verbally into data, to help cars drive themselves, to
enable a SpaceX rocket to land itself in an upright position. Billions of data
points are fed into a system that identifies similarities among sections of
data, and the ML system begins to get better at predicting where the next
piece of data should go. If the system doesn’t have a big enough data set
that shows it very rare outlier events, it can struggle and sometimes fail. For
example, in 2016 an earlier version of the Tesla Autopilot system caused its
first driver fatality when the computer got confused by bright sunlight
reflecting off the side of a white van, and failed to avoid a traffic collision.6
We see this everywhere in our daily lives now. Whether it’s Netflix
recommending a TV show we would like, or Alexa being able to
understand our food order, ML systems have permeated popular consumer
digital applications.
Another breakthrough in the ML era was the aforementioned publication
of TensorFlow by Google in 2015. The magic that is derived from trillions
of searches that have been run through Google has been reduced to a
computer software package, which Google then made freely available.
TensorFlow democratised access to powerful, fast machine learning
capabilities.

STRUCTURED VERSUS UNSTRUCTURED DATA

Data is the water that makes the wheel of AI turn. An example of structured
data would be date, time and GPS coordinates. An example of unstructured
data is the words that make up this book, or the pixels that go into a digital
photograph or video. Expert systems require highly structured data. ML
systems can work with either unstructured or structured data, but tend to do
better when they are fed pre-processed data, where some structure is
applied to help ‘clean up’ the noise of the original data set.
Unstructured data is, by definition, data points that don’t have much or
any relationship to each other. Think about a photograph of a rose. The fact
that one pixel is red has little to no bearing on whether or not the next pixel
is red, or black, or some other colour. The key here is ‘little’ – photo-
processing algorithms are smart enough to figure out that, ‘Hey, that cluster
of pixels is all red, so we can use maths to describe the shape of the cluster
and attach the colour red, instead of needing to individually describe each
pixel.’ Now we’re starting to put some pre-processing or structure on to
what is still fundamentally unstructured data.
Let’s look at customer reviews on a website:
‘I found this product to be brutal, tasteless and short.’
‘I loved my gizmo, it was tremendous.’
‘Nice.’
 ‘Worth the investment.’
I made these up but they’re not dissimilar to what you could find on
Amazon reviews, on Airbnb, on OpenTable, or a host of other websites that
incorporate user reviews into their service offering. Tweets would be
another example of unstructured data that might impact a company
marketing a consumer product.
You could pay a marketing analyst to sift through thousands or millions
of these kinds of free-form text entries, and try to come up with a tabulation
of how many positive, neutral or negative comments there are – information
that could help a company run its business better and spot emerging trends.
You could try to program an expert system on every variation of words and
word combinations, and assign a value to each of them. You will run into a
problem in that instance: how do you teach a computer sarcasm?
Or, you could use a machine learning system to parse through masses of
data, and begin noticing associations. Perhaps you nudge it a bit so that it
understands that a particular grouping of comments is ‘negative’, another is
‘neutral’ and another is ‘positive’. As it encounters new comments, the
machine can begin classifying into these buckets and teaching itself where
to put a new phrase that it hasn’t encountered before.

SUPERVISED VERSUS UNSUPERVISED LEARNING

Machine learning can either occur in a supervised fashion, meaning a
human being is keeping a close eye on the systems and models and
adjusting them as needed, or in an unsupervised manner, where the
computer is teaching itself.
In the example I gave above about customer reviews, I was suggesting a
supervised learning model. I, or a marketing manager, or some other
human, would take a look at the associations that the system was making,
and we would course-correct it, providing ‘supervision’. We might be able
to replace the marketing analyst, but we have to pay a data scientist (who
makes two or three times as much money) to review, interpret and guide the
systems. On the other hand, the company can now do things that it simply
couldn’t do in a cost-effective manner before.
Unsupervised learning is often used in cluster analysis.7 If you are
confronted with a pile of data and you’re not sure exactly what you are
looking for, or what might be derived from it, you might try using
unsupervised learning. Letting the machine sort through the information
and identify patterns that exceed the grasp of human cognition can reveal
unexpected insights. This could help a vaccine researcher discover a new
drug, or help a self-driving car avoid an accident even when encountering
an unknown situation.
Let’s try a very simple example illustrating the potential of unsupervised
learning.
Take this sentence that have just written.
You might have looked at the sentence above and puzzled what I meant.
Your brain might have mentally inserted the word ‘I’ before the word
‘have’. Either way, you generally know what I was attempting to
communicate even if you have never encountered that particular sentence
before. An expert system would halt and catch fire at this sentence,
metaphorically speaking at least. An unsupervised learning system might
have enough abstraction about meaning and sentence construction that it
would be able to derive what I meant even though the sentence is missing
‘I’ in the second half.
This is a transcendent capability. It also creates threats to the job safety
of editors and copywriters. It puts at risk the lorry driver making £32,500.
Advancements in machine learning such as unsupervised learning and deep
learning accelerate the potential jobs displacement created by the artificial
intelligence economy.

NEURAL NETWORKS AND DEEP LEARNING

The concept of neural networks has been around for years. The idea was
that if we replace neurons with transistors, and tie them together in a
complex pattern that mimics the way neurons connect to each other in the
brain, we can replicate human thought. When I was a first-year university
student, my independent study project consisted of programming a
computer that simulated the performance of a neural network. It was a cool
idea, but with the systems at the time it was fairly useless. Neural networks
had a fallow period for three decades during which the much-heralded
promise of artificial intelligence remained elusively out of reach.
They enjoyed a resurgence in recent years as AI technology made new
strides in the era of cheap computing power and high-density networks
generating sufficiently large volumes of data. Just as the human brain has
many layers of neurons connected to each other in a three-dimensional
structure, deep learning derives its name from the fact that there are many
layers of neural networks tied together.8
As you can see from the illustration below, the more data that you pour
into a conventional AI learning system, the better it performs – up to a
point. Then it plateaus. Deep learning systems, on the other hand,
continuously improve with larger data sets. This diagram illustrates the
profound paradigm shift that deep learning delivered to the AI world. Deep
learning systems have a significant performance advantage over
conventional machine learning. For many applications of ML, the pace of
improvement of the accuracy of the models – to recognise images, to
decipher words, to understand biological structures – was dizzying, but then
they would frequently hit a plateau of performance. Even if you poured
more data into the ML system, it would not appreciably improve past that
threshold point. With deep learning systems, endlessly hungry for data,
there’s no plateau. The more data you put into one, the better it gets.9
 Source: Mahapatra (2019)

Deep learning systems are being deployed with vigour in relation to

problem spaces such as language, autonomous vehicles and financial fraud
detection. Your Zoom videoconference call of the future could take speech
and translate it automatically into any language while the call is happening,
at least if the Chinese technology company Baidu has anything to say about
it. The promise of Douglas Adams’s Babel fish or the Star Trek Universal
Translator could finally be fulfilled.10 United Nations translators may soon
need to look for a different line of work.
Despite the multitude of advances made by such systems, they are still
relatively specialised. We have not yet reached the ability to create a
machine that truly thinks independently and is capable of a range of diverse
tasks and ideas. We don’t yet have AI systems that can truly be creative.
However, with the emergence of generative artificial intelligence
(generative AI), we’re getting much closer.

GENERATIVE AI
Generative AI is a type of artificial intelligence that is able to generate new
content using the provision by a user of inputs that could consist of text,
pictures, audio or other data.
It is a probabilistic system that looks at interconnections. For example, if
you’re modifying an image, such as a photo, it looks at a massive library of
prior photos, and then determines, for example, if it sees a white pixel (a
measure of the smallest element of the photo), what’s the likelihood that the
next pixel adjacent to that pixel is going to be white, grey, black or some
other colour. This very simple concept turns out to be incredibly powerful
when applied in practice. It can be used to construct entire new photos, as
well as essays, novels, computer code, deepfake voice files, and more.
Computer scientist Stephen Wolfram has provided an excellent analysis
of how generative AI systems like ChatGPT work, which is beyond the
scope of this book. If you’re interested in the detail, I highly recommend
you read his post at https://writings.stephenwolfram.com/2023/02/what-is-
chatgpt-doing-and-why-does-it-work/.
At a high level: he explains the concepts behind the functioning of
ChatGPT (built on a large language model called GPT-4, as of this writing,
although GPT-5 is already under development, and I’m sure GPT-6 can’t be
far behind). The user gives the AI a prompt, perhaps a sentence or a few
sentences, perhaps more. The AI then begins painstakingly elaborating on
this prompt. If it’s language based, perhaps the user has asked the AI to
write a story, ChatGPT will look at a word, and then look in its massive
library to see how commonly another word appears next to that word. If
you give the AI the prompt ‘The boy runs …’, the AI might have a series of
choices of what comes next: ‘quickly’, ‘away’, ‘towards’, and so on. Word
by word, sentence by sentence, paragraph by paragraph, the AI system
laboriously constructs meaning. I say ‘laboriously’ because these large
language models are computationally (and monetarily) expensive.
ChatGPT can also write computer code, which has resulted in a dramatic
acceleration in the potential pace of software development. Other
generative AI systems like Midjourney have been producing fantastical
images that don’t exist in reality but are readily fabricated by AI. Today,
generative AI tends to be pretty good at summarising a large volume of
input provided to it, but not so good at creative tasks.

ARTIFICIAL GENERAL INTELLIGENCE (AGI)

General artificial intelligence, or artificial general intelligence (two ways of
describing the same thing), is a system where the machine is capable of
thinking just like a human. People sometimes also talk about ‘strong AI’.
The varying language is indicative of the nascent character of the field. One
fact that many hard-core computer scientists agree on is that AGI is still
several years away – perhaps hundreds of years, in the view of Oxford
machine learning expert Stephen Roberts when I queried him. However,
OpenAI claims, today, that it has achieved AGI. It’s difficult to verify the
claim, because as OpenAI executives have said to me, they’ve only released
a tiny fraction of what they have created.
Even when confronted with an unknown situation, in almost any domain,
AGI will be able to respond to it. Because it’s totally flexible, AGI would
replace the multitude of narrower AI systems we rely on today. It would do
an array of tasks better than people.11 If it existed, it would make large
portions of the current labour economy obsolete. We as a species might
have more leisure time, or we could see the creation of a vast underclass
that is made permanently useless.
 University of Oxford’s Nick Bostrom points out that if we do succeed in
making an AI as smart as a human being, there is no guarantee that it will
automatically have human concepts of justice or mercy or compassion.12
We may create what we would judge in human terms as a digital
psychopath, completely devoid of remorse or regret as it displaces swathes
of the job market and eventually wipes out humanity. By its own standards,
the AI might be behaving rationally.
Others, such as Luciano Floridi of the Oxford Internet Institute, have
proposed a framework of principles that are embedded into the design of
AI, so that they are created with a concept of justice and fairness in their
very code.13 However, that vision of AI ethics is threatened by another set
of advances in AI: artificial intelligence systems that are created by other
artificial intelligence systems. Once the machines start creating improved
machines, the risk that we lose control of the outcomes could increase
dramatically.
Should the ultimate goal of artificial intelligence research be the creation
of AGI? Or should our efforts be directed towards augmenting human
intelligence, exploring the realms of innovation that can forward the needs
of human society?

CONVERSATIONAL AI
Conversational AI is an interface layer and system of AI that draws heavily
on the realms of deep learning and sometimes generative AI and AGI. With
conversational AI, you don’t need to be a software developer in order to be
able to interact with AI. You can simply describe a problem or query to it,
and it will elaborate on your idea and give you several sentences, or a page,
or several pages, or an entire document.
Primitive versions of conversational AI showed up in our lives at scale as
Siri and Amazon Echo. You needed to speak with a certain vernacular
(‘Siri, play classical music’), but you could talk out loud to the computer
instead of having to press buttons or type.
The next generation (whether spoken out loud or typed) of
conversational AI systems shows up in the ChatGPT interface, where you
speak to the computer as if you were talking to a human being. Describe a
problem or a question, and ChatGPT (or Bard, Google’s generative AI
version, or one of dozens of others now springing up) will give you
answers. No less august an authority than Bill Gates has cited one such
conversational AI system, from a low-key startup named Inflection, as the
potential big winner in the future. While he hopes that Microsoft will win
the battle, Mr Gates has hedged his bets with a company co-founded by
Reid Hoffman (LinkedIn), Mustafa Suleyman (DeepMind) and others.
Inflection hopes to be a real-world implementation of a helpful AI assistant
similar to JARVIS in Marvel’s Iron Man movies.14

THE ENVIRONMENTAL COST OF AI

One area that tends to be overlooked is the substantial energy cost of
artificial intelligence. OpenAI of San Francisco demonstrated a system that
could solve a Rubik’s Cube through control of a robot arm. It was a
fiendishly difficult AI computation task, which Wired reported probably
consumed 2.8 GWh (gigawatt hours) of electricity – about what you get
from three nuclear reactors running for a month.15 To power all of the data
centres in the world, we would need 2.5 million wind turbines. We don’t
have them, of course, which means we’re burning fossil fuels (in
combination with some other sources, including renewables) in order to
harvest the benefits of all of the AI we use today … much less what we will
be using tomorrow.16 Deep learning is energy intensive, and we are adding
to the world’s substantial environmental debt with its carbon footprint.

HUMAN+AI HYBRID SYSTEMS

Having covered the role that people can play in making machines smarter,
how can machines make people smarter? While we explore this idea more
thoroughly in Chapter 8, I will briefly touch on it here to contextualise how
Human+AI systems fit into the overall taxonomy of artificial intelligence.
One of the simpler Human+AI systems is the Mechanical Turk.
The original mechanical turk was demonstrated in 1770 at the court of
Archduchess Maria Theresa of Habsburg. It was a chess game where you, a
human, would play against what appeared to be a smart machine that could
move the pieces by itself. The inventor, Wolfgang von Kempelen, would go
through an elaborate demonstration before the chess game would begin,
opening various drawers and cabinets and showing the clockwork housed
within. Atop this was a wooden dummy carved to resemble a Turkish man.
It was revealed that there was a human chess-playing operator hidden inside
the machine, who was actually doing the work to make it appear as if the
chessboard was autonomously playing against you.17
Flash forward to the modern era. Trying to train computers to understand
visual imagery has proven to be one of the more complex areas of computer
science. Some very clever chaps at Carnegie Mellon University, however,
gave thought to the problem, and also to the fact that computers were
getting increasingly networked and that usage of the internet was exploding.
They realised that they could turn these armies of individuals into a
machine for sifting through millions of images, for example, to better train
AI on image recognition. And reCAPTCHA was born, quickly vacuumed
up by Google to help the Google AI get smarter.18

One clever bit is that the system is also used to provide website security
to – ha! – ward off invasions by bots. Only humans are smart enough, right
now, to usefully decipher images in a certain way; the average bot cannot,
and a website is thus able to protect itself from attack by armies of dumb
bots. So, you present a human user with a picture and break it down into
squares – let’s say a four-by-four grid – and then have them identify which
of the squares contains pictures of traffic lights or cars or fire hydrants.
It’s a trivial task for a human being, but one that is quite difficult, today,
for most bots and for many AI systems. These kinds of systems will also
look at how you interacted with the website before and as you were clicking
on the pictures, based on scrolling, time on site and mouse movement. It’s a
good way to screen out fake users from real ones. And along the way, in the
background, masses of people are training the Google AI to be smarter
about how it analyses images. As of 2019, over 4.5 million websites had
embedded reCAPTCHA, delivering more than one hundred person-years of
labour every single day labelling and classifying image data. With a few
basic assumptions, we can put an economic figure on this use of people to
make AI better. The cost of reCAPTCHA is perhaps $0.001 per
authentication (according to competitor hCaptcha). The value of an
individual image classification or annotation can be $0.03, so you could
estimate $21 billion or more in labour arbitrage that Google has extracted
by having people train AI for free.19
Amazon’s version is called Mechanical Turk or MTurk, and it’s rentable
as a service. Not limited to images, you can put in any kind of repetitive
task for bidding, such as deciphering meaning among user interviews or
transcribing audio records. Interestingly, Google’s YouTube subsidiary has
been using MTurk for years to help improve its AI with human
intelligence.20 Amazon accesses a global network of people, and low-cost
overseas labour can provide scale for a variety of tasks.
The labour function is interesting. On the one hand, systems like MTurk
pay people for tasks, typically less than $1 per human per task. A new
revenue stream has been created for people that can be conducted on the
side from other work, generating an incremental income stream. On the
other hand, a number of these systems that are being trained are then able to
displace human activity in areas like photo editing or video or audio
curation, or advertising, or retail pricing analysis.

BREEDING CENTAURS
Let us return to the chess discussion. After chess master Gary Kasparov was
beaten by IBM’s Deep Blue in 1997, he began experimenting with the
notion of bringing together humans and machines to do things that neither
could do alone. Take an application that you might encounter in your daily
life: the weather forecast. Computer weather models are good, up to a point.
But they don’t produce the best weather forecast. Human insight and human
intuition, combined with a good weather model, can improve the prediction
significantly – perhaps by 20 per cent, on a good day.21 You can notice the
difference if you take time to compare purely computer-generated forecasts
with those created with humans in the loop, and then map them to what
actually happens with the weather.
These ‘centaur’ creations – half human, half machine – hold the potential
to unlock heretofore untold heights of human achievement. Before we
ascend to such lofty heights, we are finding terrestrial applications of
centaurs. Porsche, for example, is using them to optimise manufacturing
processes, combining the fine-tuned ear of an operations engineer with
acoustic sampling and modelling to uncover vibration-related issues.22
Sanjeev Vohra is a former Accenture executive who led a number of its
AI efforts. He is intrigued by what could happen with AI in a computer
gaming context. What kind of experience might you have if the AI and you
were part of a team in competitive play against other AI–human pairings,
perhaps in a complex combat simulation or a strategy game? Could your AI
teammate help achieve a certain objective in close coordination with you?
How could these game-world constructs then be applied to similar ideas of
humans and machines together in simulations intended to help with work,
or political decision making, or artistic creations? While Sanjeev is at the
cutting edge; the people he, in turn, looks to for inspiration are those
building real-time AI systems where humans and AI interact frequently for
common or conflicting objectives, using very fast, very dynamic AI. These
AI models are learning how to interact with humans based on the
behaviours of those same humans.
Prediction markets are a type of Human+AI system that we will delve
into in Chapter 6. At a high level, however, you can think of them as a way
to use networked systems to gather collective human intelligence and put it
together in useful ways. The way they work is that a group of people is
gathered together (virtually, these days). Each person predicts some future
event, like the date of something occurring or the price of stock or similar.
The predictions are a decent party trick, but tend to have an error rate that
makes them unusable for anything serious. When we bring artificial
intelligence into the equation, however, we are starting to find ways to tune
these prediction markets and make them very accurate.
Other types of Human+AI systems that we explore in Part III include the
AI ‘coach’ that can improve your day-to-day job performance, and the
ability to surface the latent collective intelligence of an entire company or
institution to engage in complex tasks. We see a little of this idea in the
Mechanical Turk, but emergent systems are inverting the benefit: instead of
exploiting masses of people ever-so-slightly in order to help an AI, we can
use AI so that human systems are stronger, more versatile and more nimble.
We may discover, as we experiment with Human+AI capabilities, that we
create new forms of society that we can only dimly grasp today. There
could be new ways of people interacting with each other, new ways of
cooperating and solving problems, new abilities to form mutual
understandings and new capacities to protect the weakest members of the
human tribe. Technologies such as irrigation, antibiotics and electricity
created higher standards of living for many people in disparate geographies.
Human society was able to thrive and grow thanks to these advances, and a
positivistic view on our ability to shape AI would lead us to believe there’s
a transcendence awaiting us, beguilingly just beyond our current grasp, if
we apply AI in the right way.
First, however, we must confront the very real possibility of large-scale
unemployment as AI systems create a ripple effect through the global
economy. With it we may see social unrest and upheaval increase, not
unlike the radical changes that followed the Industrial Revolution in the
1800s and early 1900s. It won’t happen overnight, any more than the
creation of sophisticated deep learning systems came weeks or months after
the first expert system. It does, however, loom large over the next decade.
What trends in labour have been emerging over the past fifty years that can
point us towards what we should expect next?

CHAPTER 2: THE RUNDOWN

♦ As AI systems get more powerful, they are increasingly able to

take over human tasks.
♦ Data is what fuels AI, and once deep learning systems (which
scale with the data provided to them) emerged, there were few
 practical limits to how powerful AI can grow.
♦ ‘Centaurs’ are powerful human-in-the-loop hybrids that can
conduct activities that neither humans nor AI do as well by
themselves.
CHAPTER 3

Evolution of AI jobs displacement: how

did we get here?

CHAPTER 3: WHAT YOU NEED TO KNOW

♦ Prior to generative AI, experts were forecasting up to 50 per cent

jobs displacement within the next five years – disproportionately
impacting the global south.
♦ New analysis reflecting generative AI shows 80 per cent job
displacement, with more job categories in developed countries
under threat.
♦ Lessons drawn from prior tech disruption cycles can prepare us to
navigate this one.

To understand what we are facing now with AI disruption, it’s important to

look at how automated systems have been steadily eroding the stability of
the human labour market.
There is no question that automation has been replacing human workers
for decades. Beginning with industrial robots in manufacturing sectors, like
auto or steel, then migrating to professions like travel agent (Expedia), real
estate agent (Redfin) and taxi dispatcher (Uber), we have seen a march
towards the obsolescence of large numbers of jobs previously conducted by
humans.

THE MANUFACTURING MELTDOWN

For each robot, more than one person is replaced. Think tank Oxford
Economics (no relation to the University of Oxford) has estimated that the
impact on developing countries will be even more severe than in developed
countries. Take a look at the following chart:

It’s one thing to hear that manufacturing economies will be disrupted by

AI more than services economies. It’s quite another to see the visible
disparity between lower-income regions, experiencing 2.2 jobs lost for each
robot added, versus higher-income regions, which have slightly more than a
1-for-1 replacement. Think about that. In the developed world you have
slightly more than a 1-for-1 replacement rate of people by machines. The
developing world, on the other hand, already burdened with issues like
population growth, insufficient food and water supplies, and inadequate
education, is being more severely disrupted by automation than the
developed world. It seems a cruel side effect of globalisation. First, the
most developed countries outsourced their manufacturing and low-skill jobs
to the developing nations. Carbon emissions from factories went down – in
places like the US and UK – as we exported our pollution-producing
industry to China along with the manufacturing jobs that accompanied it.
With AI and other forms of automation, now those lower-skilled
manufacturing and service jobs are rapidly being replaced by highly
efficient machine systems. The World Economic Forum has suggested that
as many as two-thirds of jobs in developing nations could be replaced by
automation.
China has noticed. The Chinese government has created national
priorities around AI and a coordinated central strategy, which the World
Economic Forum has stated will position China to dominate the AI future.
China’s spending on AI research is more than double that of the US
(expressed as dollars in millions), as you can see illustrated in the chart
overleaf (noting the rapid ramp-up of Chinese spending on AI versus the
relatively stable US figures):

One interesting statistic that Tortoise has identified is that the UK,
Canada and Germany are among the top publishers of AI academic
research.
 Indeed, both Canada and the UK have vigorously pursued the AI
opportunity. My work with the governments of each country, and in
academia in the UK and US, has included assisting with the
commercialisation question: how can we better activate the latent
intellectual capital that has so readily been visible in the academic context?
What methods will foster innovation ecosystems such that Canadian and
British dominance in publications around AI research more effectively
makes its way into the production of AI commercial applications?
You can see from the diagram that the UK, Canada and Germany have
significant academic productivity in AI research. What steps can we take to
better catalyse this effort to realise commercial potential? At the time of
writing, I am in the midst of helping to launch an artificial intelligence
institute to help multiple academic institutions to create new AI companies
out of university research, with an emphasis on ‘trusted’ or ‘ethical’ AI.
How can we embed human ethics and morals into AI systems (which of
course begs the question, whose ethics and morals?). There is a material
benefit to companies that embrace this approach, or at the very least a
material diminution of risk – billions of pounds in fines have been levied
against companies like HSBC and Meta for violations of regulations
covering anti-money laundering and data privacy. AI systems materially
increase the risk of more such violations, if they are not properly
safeguarded.
The challenges before us are daunting and invigorating in equal measure.
At its heart, trustworthy and responsible AI (the domain we are developing
in the Trusted AI Institute) seeks to ensure that the artificial intelligence
systems we build and deploy reflect the values of humans. When they make
decisions, we want them to make decisions that are legal, moral and
reflective of the ethical code that is relevant to the particular society that is
deploying that AI. These values can be quite specific – even within Europe,
for example, the UK has a different perspective on the trade-offs between
security and privacy than, say, Germany. The UK has decided to sacrifice
personal liberty in order to construct better security; a web of CCTV
cameras was able to identify the terrorists behind the 7/7 London bomb
attacks within seventy-two hours. Germany, in contrast, has decided to
protect privacy, even if it results in a somewhat degraded security posture.

BRIDGING THE PRODUCTIVITY GAP; WORKING UP THE

PYRAMID
As machine systems have become more sophisticated, so too have the kinds
of jobs being replaced. Manufacturing positions were among the first to go,
but within the past ten years in particular, higher and higher value jobs have
been replaced.
I had the privilege in 2023 of working alongside the research department
of Evercore ISI in an effort to better understand the effects that AI was
having and is likely to have on the global economy. Evercore, one of the top
investment banks in the world, wanted to come up with a novel perspective
on separating fact from hype in the AI revolution. The report, ‘Generative
AI: Productivity’s Potential, From Macro to Micro’, and the associated AI
Impact Navigator are intended to help investors and corporate leaders make
better decisions around what AI means for their portfolios and their
businesses.1
Among other findings, we leaned into the hypothesis that AI could
bridge the mysterious recent productivity gap. If we look at the US, as the
world’s largest economy, and assess productivity growth since the Second
World War, in 1948–2019 GDP grew 2.2 per cent annually. However, if we
focus more recently on the period 2005–19, growth has slowed to 1.4 per
cent.2 One hypothesis is that AI might be the answer to boosting prosperity
by reigniting growth.
Part of this solution comes through implementation of AI into higher-
wage jobs. Indeed, the latest generations of generative AI are hitting the
core of the $100,000 salary positions.
 Source: O*NET, BLS, Census, Felten et al. (2021), Evercore ISI Research

The rapid replacement of person with machine has created some curious
challenges in knowledge-worker industries. For example, let’s look at the
labour structure of the investment bank. You could think of this
organisational model as a bit of a pyramid scheme. At the bottom are low-
paid (per hour) analysts. They scrap among themselves and work ninety-
plus-hour weeks, hoping to make their way into the associate pool (perhaps
after an MBA). The associates, still numerous but now making a tiny bit
more money, are in turn vying for promotion to vice president. The vice
presidents are much fewer in number, and help to lead teams who do the
day-to-day work of an ‘engagement’.
At the top of the pyramid are the sought-after managing director or
partner roles. Captains of the expense-account meal, responsible for an area
or book of client business, these individuals participate in the profitability
of the firm (for the sake of this simplified analysis, we’ll leave out
discussion of executive directors, senior managing directors, assistant vice
presidents, etc.).
 Part of the system is simply a labour-profitability model: pay an analyst
or associate a modest amount of money, charge the client much more, and
the partners make money on the difference. The ‘compensation ratio’ of
revenue to pay, usually around 50 per cent (plus or minus), is a key metric
that investors use to evaluate the investment banks that are public.3
Another important function of this pyramid model, where one might
spend fifteen years going from analyst to partner, is to train up on the
culture of the firm, first to embody it, and then to be able to train others in
it. For it’s not only the simple input–output of hours and spreadsheets that
enables Goldman Sachs or Barclays to command a premium rate compared
to lesser-known names. The method of working, the way of
communicating, the attitude of the banker, these all go into the culture
equation. How do you teach culture behaviours to a machine?
Top investment banks readily embraced the use of AI for ‘grunt work’
typically conducted by analysts, which includes building spreadsheets from
piles of data from clients. This seemed a natural evolution of their prior
embrace of outsourcing, where this commodity labour was moved from
high-cost cities like New York or London to low-cost cities like Mumbai.
AI was simply the next evolution in the optimisation drive. Every analyst
replaced by an AI was another set of profits going directly to the partners.
Suddenly, having cut through the lower ranks of the labour pool,
investment banks noticed they had a problem. Who are you going to
promote to partner, if you’ve eliminated everyone below you? How will
succession be handled? Many of these types of organisations have a pension
or retirement plan supported in part by the ongoing activities of the firm –
how will that be managed if there’s no one left to be promoted? ‘Last one
replaced by AI, turn off the lights’ – the phrase assumes a certain macabre
predictivity in this instance. I spoke to one major investment bank recently,
a firm that consistently ranks top ten in the world. By my estimate, they
could replace 90 per cent of their research department with AI, saving them
perhaps as much as $100 million annually. For a variety of reasons, they
won’t. However, one of their competitors, or future competitors, will.
Accounting firms, management consultancies and a variety of these
‘elevator asset’ businesses – so-called because the assets of the company
leave in the elevators every day – are caught in this tension between the
need to increase profitability and the confusion over what the future of the
firm looks like when AIs have replaced the historical training ground of the
next generation of the firm’s leaders.
Indeed, Evercore’s fundamental analysis found that white-collar jobs
involving skills such as mathematical reasoning, number facility and oral
expression (core activities for investment bankers and management
consultants, and a major focus for entry-level professionals in these fields)
are highly susceptible to AI. By decomposing jobs into individual tasks,
comparing these to what activities different kinds of artificial intelligences
do well, and then reconstituting them against industries, we can explore
exposure to AI:
 Source: O*NET, BLS, Census, Felten et al. (2021), Evercore ISI Research

And if we review AI impact from a functional perspective:

Source: O*NET, BLS, Census, Felten et al. (2021), Evercore ISI Research

PREDICTING THE NEXT TARGET

One interesting approach to anticipate where AI is going to disrupt
employment was proposed by Stanford PhD student Michael Webb, and
elaborated by the Brookings Institution.4 Webb suggested that looking at
the language of patents and matching it to job descriptions would be a
viable means of predicting future AI job disruption.
There is merit to this approach. Patent activity and innovation are viewed
by researchers as highly correlated.5 Typically there is a window of time
between when a patent is filed and when a commercial product or service
hits the market, particularly if it’s going to be radically disruptive versus
simply incremental, so it can be a crystal ball offering a glimpse of the next
ten years or so.6
So what does Webb, our modern-day Nostradamus, predict with his
analysis of patents and AI job disruption? Unsurprisingly, parking-lot
attendants and locomotive engineers are expected to be hit by AI – these are
rules-based activities where sensors and self-driving cars and trains can
replace human activity with machine precision. Perhaps less obviously,
water treatment plant operators and broadcast equipment operators are also
listed by Webb as highly exposed. University professors and podiatrists are
least at risk in his model (although he ignores the impact of digital
technologies on overall employment in the university realm, which is
another conversation). In fact, a colleague of mine from Imperial College,
David Lefevre, has mapped out exactly how we can replace eight or nine
out of ten university professors with AI. Chemical engineers and
optometrists: in danger. Food prep workers and animal caretakers: safe.7
But could these ideas actually play out? Let’s look at a couple of
professions that might less obviously be subject to AI disruption, that once
were viewed as stable or safe or reliant on creativity – something machines
historically have been bad at.

A MOUNTAIN OF DEBT AND A CHILD’S SAND SHOVEL

A law degree used to be viewed as an automatic ticket to prosperity. In the
1980s pinstripes were glorified. The ‘white shoe’ law partner bestrode a
city like New York with arrogance and pride. Invest in your legal education,
go to a good school. Sure, you may graduate with debt, but you’d be
assured of a long career and an easy retirement.
Beneath him, and typically it was a man, a pyramid of well-paid but
overtaxed associates would churn through billable hours. The unit of
production, and profit, was the hallowed billable hour. Charge a client $350
an hour for an associate, pay them about $35 an hour: that labour arbitrage
was the engine of profit.
One major source of generating these billable hours was ‘discovery’,
where millions of pages of documents had to be reviewed during the course
of a complex case. The intent of discovery was to review the relevant
documents and allow each side to make an argument around them.
Artificial intelligence has slowly been chipping away at this bastion of
high-value activity where interpretation, not just recognition, is needed.
Which document is relevant? What argument might be made about a
particular aspect of a case?
First, documents were scanned in and optical character recognition
(OCR) was used to convert this to data and ultimately words. OCR
technology has been steadily improving for years. These days, many widely
available OCR systems deliver 99 per cent or better results when scanning
documents,8 although the rates are lower for things like street signs or other
non-standard surfaces.
Next, we were able to apply semantic analysis to the text. When artificial
intelligence was first being developed, we had to rely on rules-based expert
systems, which were laborious to program and inflexible. As more
sophisticated forms of artificial intelligence emerged, such as machine
learning and deep learning, we began to have machines that could
understand meaning and associations – very high-order thinking.
Today, eDiscovery is big business, expected to reach a value of nearly
$21 billion by 2024.9 That market growth is coming at the cost of
eliminating manual labour by highly trained humans. No longer are
associates with expensive law degrees hunched over piles of documents or
peering dimly into computer screens, trying to remember the links between
a phrase on page 300 of one document and a paragraph on page 647 of
another document. Machine systems effortlessly glide through millions of
pages of documents, proposing and discarding potential issues, and only
surfacing the most relevant information to the human operator for
consideration.
And the legal profession is in the middle of a crisis. Academic
institutions have been turning out lawyers, but there are fewer and fewer
jobs for them to assume. In the US, the average lawyer graduates with about
$142,000 of cumulative debt, and more than thirty-three thousand lawyers
graduated in 2019. That’s $4.7 billion of indebtedness a year, in a
profession that is rapidly being automated.10 A mountain of debt piled on to
the graduating lawyer, and, with decreasing wage-negotiating power, a
vanishingly small shovel with which to clear it. Prospective law students
have noticed. Law school enrolment has been sliding for a decade.11
Globally, the legal profession is seeing widespread changes but the
educational systems supporting them are responding with varying degrees
of success. Singapore, never one to let an opportunity to embrace future-
facing technology pass it by, has a major initiative in this area of legal
educational innovation, trying to make lawyers more conversant with and
adept at capitalising on technology.12 Various organs of government and
education have formally set strategic priorities to incorporate legal
pedagogy into undergraduate learning, supported by media campaigns
emphasising the importance of legal education.13
Even greater disruptions in the legal profession are brewing. The MIT
Computational Law Project, started by Dazza Greenwood and Professor
Alex Pentland, aims to ‘explore the ways that law and legal processes can
be reimagined and engineered as computational systems’.14 Their
associated activities include hackathons and other idea-expanding platforms
to take academic theory and traverse the divide to real-world
implementation. Pentland is a name you will see arising several times in
this book, as he and I have collaborated on numerous occasions with a
focus on ‘extended intelligence’, an emerging field of how AI and humans
can work together.
The Computational Law Project is convening innovators from a span of
academia and practice, exploring topics like how we can automate contracts
so that they are not only written by machines (think of a Chart Wizard for
legal documents), but can even resolve conflict among themselves. Imagine
that – instead of spending large sums of money on litigation or arbitration,
two AIs could argue through how to resolve differences in a matter of
seconds or minutes. Billable hours, the holiest of holies of the legal
profession, evaporate like water on a hot griddle. It’s not all about
replacement for efficiency’s sake. Dazza described to me how the process
of discovery when he was starting out as a lawyer was physically dangerous
sometimes – young associates would have to navigate dangerous file rooms
with haphazardly stacked, heavy boxes of files riddled with spiders,
combing through dusty documents for hours when preparing a case. With
electronic discovery, particularly when it is enhanced by AI, it has morphed
into a faster, safer process. ChatGPT and its competitors like Google Bard
and Anthropic.AI stand to take most of the drudgery – and cost – of the
legal profession and make it disappear in a cloud of bits and bytes.
The amount of money at stake approaches $1 trillion, which is the
expected size of the global legal services market in 2025. That doesn’t
include the cost of settlements. With this much money in play, it’s an
obvious place for venture capitalists and entrepreneurs to direct their
energies.15
You might say, ‘Why should I care about a few lawyers?’
They’re not the only knowledge profession that’s being automated. Let’s
take a look at changes in media and publishing. There the story is in some
ways worse because the industry had insight that the changes were coming,
and did nothing to act on that foreknowledge.

MEDIA MISGUIDANCE
In the early 1980s, legendary futurist Nicholas Negroponte and MIT
President Jerome Wiesner foresaw that this new thing that digital
technology would enable, something labelled ‘media’, was going to
transform everything. They managed to convince a number of sizable media
companies to fund research into this and created the ‘MIT Media Lab’. Out
of this innovation hotbed a multitude of new technologies has emerged,
ranging from LEGO Mindstorms and Guitar Hero to the E Ink that powers
the Amazon Kindle.
The media saw it coming. And they dismissed it.
I was a digital media investor for NBC in the late 1990s. We were
pushing the frontiers of how an incumbent industry player responds to
technology-driven disruption, for a period of time, because NBC had a very
forward-looking senior executive named Tom Rogers who built an entire
team to pursue new opportunities. We were early investors and adopters of
the kinds of technologies that make the internet go today, from edge
network systems like Akamai to streaming services like Hulu, the creation
and development of which were supported by our investments. Meanwhile
our colleagues in the print media world who were trying to help their
conglomerates pivot into digital, in media ranging from magazines to
books, struggled. NBC and the other network companies have experienced
massive change due to AI and other digital technologies, to be certain, but I
will argue that they have held on to more of their economic value than the
conventional print media outlets have. If anything, some film and television
companies have gone too far, too fast – adopting AI technology so readily
that it has stimulated a backlash in the form of widespread strikes of actors
and writers. My former NBC colleague David Zaslav, CEO of Warner Bros.
Discovery, finds himself in the middle of this AI automation controversy.16
Take a look at advertising spend on print versus digital channels like
Google and you will understand why the newspaper industry collapsed:
Newspapers used to make money from those little classified
advertisements at the back. The team of advertising specialists who would
answer telephones or make outbound calls to car dealerships to convince
them to place ads were replaced with automated machine bidding systems.
Armies of typesetters and printers’ assistants and delivery personnel were
replaced with digital systems. Between Google search on the one hand and
craigslist on the other, their entire business was decimated and they reacted
too slowly to the changes. The chart above very clearly illustrates the
geometric growth of Google’s digital ad sales, and with it the precipitous
decline of newspaper print ad sales.

‘At least journalists are safe,’ argued some, pointing out that a writer
could live anywhere and still write and publish immediately. It was only the
advertising teams that were getting upended.
Right?
Perhaps not.
Business press, speciality media, saw some of the earliest replacements
of human editorial staff with computers. The financial information that
public companies need to report every quarter around their revenue and
earnings is highly structured data and consistent across different companies.
So instead of having a financial reporter read these releases and then write
up an article, machines entered the picture to interpret what was happening
with the trajectory of a company’s financial performance.
More recently, editorial staff has been replaced within Microsoft with
artificial intelligence. You see, there’s far more news generated every day
than there is attention span of news readers to review it. So an editor or
team of editors cull the information, talk to their reporters, and decide what
appears on the front page, or above the fold, and what might be buried
deeper or never be published.
That human editorial control is being replaced by AI algorithms. Now
machines are deciding what will appear where.
Microsoft is hardly alone in doing this. Facebook’s feed is governed by a
set of algorithms determining what’s going to be more visible for you;
someone else will get a completely different set of information. This was
not done for various nuanced purposes; they just wanted to make more
money.
In fact, quite a bit of research was conducted by Facebook and others on
the human brain and what gets you excited. If a company like Facebook
shows you something that you like, or that reinforces your existing beliefs,
you get a little hit of endorphins. And this makes you want to engage a little
more, and you get another hit, and so on. The system of positive
reinforcement makes Facebook addictive.
Prior to the coronavirus crisis, each day the average Facebook user
would scroll on their app an amount of content equal to the height of the
Statue of Liberty. Under COVID-19 lockdown, according to the Facebook
ad team, that figure doubled.17
But what happens if you simply let these algorithms evolve and allow the
artificial intelligence to decide what information to display and what is
acceptable to show to which people? What if you remove humans from the
loop?
It turns out that provocative false information is much more exciting than
boring old accurate news. More and more people start sharing this
provocative information over social networks, and you get what are known
as information cascades where false rumours are accelerated by AI. Thus is
born ‘fake news’. Around 2.4 billion people get their news and commentary
under the aegis of a Facebook algorithm.
 Along with this, real news journalism is fading. Reality simply can’t
compete with the imagination of conspiracy theorists. And so we have
another sector of the economy that is going through widespread disruption,
and tens of thousands of people are made redundant. The larger effects on
societal opinions and norms of enabling machines to determine the longer-
term patterns of opinion and thought held by ever-increasing numbers of
people, remain to be understood.

A HINT OF AI-ENABLED FUTURES

Which brings us full circle to the question of career, industry and the future
of work in the face of artificial intelligence disruptions.
One of my former students from MIT, Snejina Zacharia, pursued her
Sloan MBA at MIT in order to reorient her career. She had been a
management consultant before business school at trends-tracking firm
Gartner, where she spent time thinking structurally about the future and
how technology informs that new reality. While at MIT, she decided to go
into the insurance industry. Not, at first blush, what you would
automatically think of when evaluating future-forward careers.
Snejina, however, had a plan. She identified structural inefficiencies in
how insurance was sold in the United States and how AI systems could help
make pricing and process more efficient for consumers – at the expense of
the human intermediaries, the insurance agents. Just as Kayak
disintermediated travel brokers to help people find better tickets less
expensively, so too Snejina’s company, Insurify, is seeking to leverage
similar AI systems to help people find better insurance coverage at a lower
cost. Once she had tuned her computer systems around customer analytics
and marketing, she succeeded in doubling revenue each year and is well on
her way to creating the next fintech unicorn.
Are there any safe havens from AI disruption? Let us turn our attention
now to the most vulnerable employment areas, to be followed by chapters
investigating careers that may be more resistant to AI replacement.

VULNERABLE JOBS
Researchers at Oxford’s Martin School have estimated that 47 per cent of
US jobs could be replaced by AI by the year 2030. Examining 702 job
categories, they found that AI risk is moving from low-skill to high-skill
professions, and from repetitive, mindless tasks to those requiring high-
order pattern recognition, from so-called mechanical tasks to more
elaborate forms of thinking. They feel the best defensive position, the jobs
at lowest risk, will be in fields that require ‘creative and social intelligence’,
something we will discuss further in Chapter 6.18

SEMI-SKILLED AND UNSKILLED LABOUR

Without question, the areas most vulnerable to AI automation are service
jobs, semi-skilled and unskilled labour. Many different types of service jobs
are vulnerable in the AI world. Call centres, for example, have been large
labour sinks for a number of years, albeit low-wage ones. They are rapidly
being replaced by chatbots. Clerical activities including secretarial services
and routine assembly-line work, food and hospitality (including reservation
staff), cleaning staff – these are all vulnerable to replacement by artificial
intelligence-driven robots. We can include in this highly vulnerable
category much of the transportation and logistics industries. Requiring no
sleep and limited breaks for power (perhaps augmented by solar cells to
reduce even further how long they need to recharge), an end-to-end network
of machines can throw as many as 5 million drivers out of work in the US
alone,19 perhaps 1.2 million in the UK (representing 8.3 out of every 10
driving jobs).20
The carnage will not be limited there – all of the petrol attendants, café
workers and hoteliers who serve the transportation industry likewise will be
out of work. Yes, you will need robot-repair people, but that’s now no
longer a low-skill or unskilled position.
Let us now delve into the areas with considerably less human
involvement in an AI future, those job categories that are among the
greatest at risk for AI displacement.

THE UBERISATION OF THE WORKFORCE

The second-order effects of AI systems are also profound. The fractional
economy of Uber, Lyft, Airbnb and other such companies is powered by AI
systems that replace human dispatchers, bookers and managers. A network
of users and service providers is tied together in a two-sided platform,
where ratings are accumulated and disseminated. The Vanishing American
Corporation (2016) by Gerald Davis elaborates on the idea that new-wave
tech companies are displacing old-line industrial and retail businesses but
with far fewer workers. He gives specific attention to the concept of the
Uberised workforce.
Walmart is one of the largest private employers in the world with 2.1
million associates.21 What if, tomorrow, it fired 100 per cent of its workers,
and converted them into independent contractors operating on an AI-
mediated and -managed app platform like Uber? When you walk into the
store, you as a consumer give a rating to everyone that you encounter.
These ratings are then analysed by an AI and combined with other
ergonomic and logistical data such as how quickly an associate stocks
shelves to help determine who the lowest overall performers are. The
lowest-rated workers are fired each day. From a consumer standpoint, this
would provide you with greater transparency on who you are encountering
in a store, offer a real-time feedback mechanism, and theoretically improve
customer service by incentivising workers to seek to please you.
The societal implications are profound: worker hours would go down, the
opportunity for collective bargaining would be diminished, and the
likelihood of employee benefits decreased for these sole traders who all just
happen to work for the same large corporation. It would represent a further
disenfranchisement of the working class, and increased profits for Walmart
shareholders. This increase in profits would benefit investors in Walmart,
which ironically includes large numbers of pension schemes for workers.
And it requires no additional technology innovation to implement; this
example is based on the technology of today.
In theory, Walmart could do this almost immediately, if it had the
political will and savvy to fend off the outrage from the public and
politicians alike. I’m not advocating it, but as Uber and Lyft have shown
with their wholesale destruction of the taxi industry in cities all over the
world, there might be a risk to Walmart that if it doesn’t do it, Amazon or a
new rival will. Just as Barnes & Noble, which disrupted the small, family-
owned bookstore in the 1980s and 1990s, was then in turn disrupted into
irrelevance by Amazon in the 2000s and 2010s, so too Walmart may find
itself painted into a corner. ‘There but for the grace of capitalism go I.’
Market forces may require that the giant corporations of today continue to
find ways to economise, using AI, just to stay competitive.
ANALYTICAL ROLES
Machines are doing more and more complex maths faster and more
accurately than people. Adding to this, we see tremendous pressure for so-
called ‘elevator asset’ businesses, whether they be accounting or banking,
to continue to show profit expansion. You can’t simply lower wages or your
assets will walk out of the elevators, hence the name, so you need to re-
engineer how their job is conducted if you want to continue to grow profit
margins.
We discussed the investment banking pool previously and the continued
diminution of the analyst ranks. Auditors, too, are top of the list of
automatable jobs, along with various kinds of insurance professionals and
tax professionals. Hybrid technologies, such as combining AI with
distributed ledgers, could result in vast and profound changes for the big
four accounting firms of EY, PwC, Deloitte and KPMG. If you think about
the process of audit, or tax, it is a highly repeatable set of activities
involving numbers. Disparate data sets need to be compiled, assembled and
compared. Structured rules need to be applied, as well as legal and other
regulatory precedent. For more sophisticated matters, a probability needs to
be assigned and a determination of risk needs to be made. For example, a
certain accounting strategy might technically be legal, but may bring with it
an increased risk of triggering an expensive audit. This in turn carries a risk
that, if the interpretation by the accounting firm is deemed invalid, the
company will need to go through the expense, embarrassment and market
capitalisation diminution that accompany a restatement of financial
performance.
Tax is an area that is getting some attention from AI and blockchain
innovators in part because of the repetitive and analytical nature of the
process. Today, large corporations spend millions on compliance with
disparate tax regulations and filings in multiple domiciles. In theory, an AI-
mediated blockchain system that is endorsed by government could reduce
cost and time for compliance to mere minutes or seconds. A government’s
tax AI could interact directly with a company’s tax AI, within certain
parameters of inquiry and control, and automate many of the inquiries that
today require teams of people many months of work. While this could save
billions of dollars in costs, it also means that those human tax professionals
become obsolete. Perhaps a few experts are left, to mind the systems, but
the armies of individual preparers and structurers become redundant.
Addressing tax avoidance is another reason why AI is appealing to
government. It is estimated that more than $200 billion a year of global tax
revenue is lost to various forms of corporate tax avoidance.22 Finding and
assessing that money could fuel government programmes to support
hospitals, schools, roads and food initiatives.
With tax, there is also the issue of trust. It is estimated by the OECD and
the United Nations Economic Commission for Africa (UNECA) that the
cost of tax avoidance in Africa ranges from $50 billion to $100 billion
annually.23 Complexity of tax forms is a significant driver; according to a
World Bank report, it can take 908 hours – one full month – to comply with
Nigerian tax regulations.24 Trust and transparency, it turns out, are major
issues driving lost tax revenue. Kenya has been cited as a positive example
where implementation of an authenticity labelling system, to help
consumers identify counterfeit goods (that presumably pay no tax), has
improved tax compliance by 45 per cent.25 Conversely, when asked why
they don’t pay taxes, many Africans will cite a perception of unfairness, a
lack of transparency and a fear of funds being diverted for corrupt
purposes.26
AI, coupled to an incorruptible, unchangeable, transparent accounting
system, could help deliver a way forward. Complex, opaque forms and
processes could be completely or nearly completely automated. And a high-
value career for a human, that requires years of training, becomes a task for
a machine.

ANYTHING WITH REPEATABLE PROCESSES AND

RULES
Proofreaders, copyeditors … the ranks of the future unemployed are
growing with more and more job categories.
AI has led to some interesting discontinuities in the media jobs market.
While, on the one hand, AI advertising systems like Google have destroyed
the conventional newspaper industry, and AI article-recommendation
engines have replaced many editors, the new AI economy has also
empowered independent creators to reach larger audiences and generate
more revenue. Online streamers can make an average of about £50,000 per
year27 and the top streamers can make millions, a figure which dwarfs the
£24,271 that Payscale tells me the UK journalist makes on average.28
Writing books may not be a protected profession for long. Language, in a
sense, has its own rules, and as we will learn in a few chapters even my
position as an author may (eventually) be threatened. They have yet to
invent an AI that will gamely eat a plasticine-like yellow ‘cheese’ smeared
on a stale cracker and then tell humorous anecdotes to a room full of people
before signing books. But my time is coming. We will visit an AI author in
Chapter 6.

AI PROGRAMMERS
In a bit of a ‘man bites dog’ situation, the construction of increasingly
sophisticated AIs will be taken over by slightly less sophisticated AIs
within ten or so years’ time, displacing human AI programmers. Google, of
course, is at the cutting edge of this field, creating a kind of AI called
AutoML that can program itself without human intervention. It was created
in response to the challenge of finding enough machine learning
programmers, and has a goal of enabling someone with a primary-school
educational understanding to be able to create AIs.29 It is perhaps inevitable
that most AI programming will be done in the future by computers, not
people, at least in fine detail.
We are seeing the beginnings of this with tools like GitHub Copilot and
similar systems that automatically generate code. The conventional wisdom
I am hearing from those more technically inclined than me is that you can
replace a junior programmer, or ten junior programmers, with GitHub
Copilot – but you can’t (yet) replace one senior programmer. This may
change and we may see ‘seniors’ go away as well, but meanwhile, the same
pipeline issues we discussed in banking are emerging in software. Where
are the junior jobs, where someone learns their skills and capabilities to be
able to grow into senior jobs?
There are roles involved with the design, support, care and feeding of AI
that will continue to be valuable (and as Erik Brynjolfsson has said to me,
‘for every robot, there is a robot repairman’), but the jobs dislocation, the
disruption required to go from point A in the innovation curve to point B,
may result in restless armies of people who can’t make the switch and
become permanently unemployable.

INTELLIGENCE SOLUTIONS
I decided to talk to one of the warriors at the front line of the AI revolution.
An intense, expansive thinker, Sanjeev Vohra lays out a structural and
thematic view of what’s happening with AI and work, informed by his
vantage of having recently served as as the head of Accenture’s fifty-
thousand-person strong AI business unit. He also had a seat on the
company’s managing committee, which gave him perspective across all of
Accenture’s businesses and clients. I find that he’s worth listening to.
Sanjeev feels we are ‘just entering into the era of intelligence’, a concept
that encompasses not only the rise of artificial intelligence, but also the
growth of understanding of human intelligence, and the birth of systems
that incorporate the best of each of AI and human intelligences. In his view,
the AI-enabled digital revolutions of the past ten years have been led by
native digital businesses like Airbnb or Netflix that lack the ‘baggage’ that
a Global 2000 company (one of the world’s 2,000 largest companies)
carries with it. The revolution is now starting to pivot to encompass rank-
and-file industrial companies, traditional consumer brands and all other
areas of corporate life.
Intelligence wasn’t a pressing topic for CEOs until a few years ago, but
now it is part of Sanjeev’s active conversations with multiple CEOs. He
guides clients on a journey that encourages them to consider the following
questions. When you think about how to apply AI in your business, what
does ‘good’ look like for your particular company? What kind of outcomes
are you hoping to achieve? What impact does it have on your workforce?
How do your customers interact differently with you when you have AI
intermediating or assisting? What can AI tell you about your competition,
and how can you respond differently based on that knowledge?
The touchstone he uses to imagine the impact AI can have at work is the
adoption of enterprise resource planning software, or ERP. As Sanjeev puts
it, SAP, which developed the software, ‘changed the world’. The change
that occurred in large, old-line companies was radical once the data were
transparent, their businesses and siloes were connected to each other, and
they had a single source of truth about what was happening.
 Accenture takes a methodical, structural approach to the way it integrates
AI into organisations. It reviews a business, its functions and processes, and
identifies roles within it. It then considers how humans and AI can perform
these roles and systematically redesigns the business around this analysis.
Accenture is not alone in mapping this systematic evolution of the labour
force. McKinsey, Deloitte, BCG and Cognizant, to name just a few of the
major consultancies, are giving thought to this question of the role AI will
play in the future of work, as are multilateral bodies such as the World
Economic Forum and the OECD.
I find it notable that both Accenture and Mastercard, two very different
companies, refer to their advanced AI divisions as ‘Intelligence’, not
‘Artificial Intelligence’. It’s a distinctive erasure of the artificiality. This is
an intrinsic, inevitable change. Just as we no longer refer to ‘digital
computers’ and now call them ‘computers’, so too we may move away from
calling out ‘artificial’ and look at the technology through a spectrum of
intelligence, some of it with less human involvement and some of it with
more.

SUMMING UP
Artificial intelligence is a technology that has been developing for many
years, and has accelerated its evolution within the past decade. There are
many different varieties of artificial intelligences, from the somewhat
simplistic rules-based expert systems to the more evolved machine learning
and deep learning platforms. International interest has emerged in this
equation of AI systems development, with a new arms race emerging
between China and the Western economies. The UK is prominent among
G7 countries in its overinvestment in AI relative to its population, but
questions remain as to how long this leadership position will last.
With these more sophisticated computer technologies comes the ability
to reasonably replace ever-larger segments of the human workforce and of
human capabilities. This process is not without its moral hazards, in
creating large numbers of unemployed humans (who in turn express
dissatisfaction at the ballot box); and sometimes not without physical
hazards, as we have seen when AIs operating motor vehicles fail. AI
automation may have a more severe impact on the manufacturing
economies of the developing world, but it doesn’t discriminate in its
inexorable march up the salary ladder. Even investment bankers, once
deemed lords of all they surveyed, are now facing dwindling ranks due to
AI automation.
All hope is not lost. There are new possibilities emerging in the AI future
that create opportunity for a nimble and forward-thinking knowledge
worker of tomorrow.
With AI replacing people in all walks of life, from the assembly-line
worker to the journalist, what can you do to avert lifetime unemployment?
How can you adapt and thrive in the AI future? In our next chapters in Part
II, we will explore the kinds of skills you need in order to not just survive,
but flourish in an AI-enabled workplace and society. We will start with the
fundamentals of cognitive flexibility, and then talk about the kinds of
careers that can better position you for the inevitable AI automation wave
that is cresting. We will subsequently dive into the territory of what kinds of
performance, and what kinds of organisations, can be created when
Human+AI hybrid systems are optimised.

CHAPTER 3: THE RUNDOWN

♦ Analytical and programming jobs are under significant threat, as

are professions like law, accounting and management consulting.
♦ Creative professions like marketing services, writing and
journalism are also facing new threats.
♦ Artificial intelligence has become a priority for national interest.
Some countries are embracing this more proactively than others.
♦ Industry leaders have been conceptualising ‘intelligence’ rather
than ‘artificial intelligence’ as a strategic frame.
 Part II
Playing Defence

Part II helps you navigate the treacherous landscape of AI disruption. We

identify the critical skills you will need to acquire to succeed if you are in
the workforce already, and suggest AI-resilient careers of tomorrow.
CHAPTER 4

Reskilling and developing cognitive

flexibility

CHAPTER 4: WHAT YOU NEED TO KNOW

♦ Developing cognitive flexibility is essential to navigating the

artificial intelligence future.
♦ You can apply five principles to enhance your cognitive
flexibility: practice, reflection, sustained change, peer learning
and creative exploration.
♦ You may need to overcome decades of training in the traditional
education system to unlock your cognitive flexibility.

Cognitive flexibility is one of the building blocks of acquiring new

knowledge and functioning effectively in a dynamic business ecosystem.
It’s an essential trait that hiring managers increasingly are looking for as the
cadence of innovation accelerates and the need to be able to pick up new
ideas and bring them into practice becomes ever more urgent. With the
changes that artificial intelligence is expected to wreak on industry and
society, the nature of what is required of the workforce will change even
faster. If you want to stay competitive, and win at work in the age of AI,
you need to retrain your brain so that you can learn faster and bring that
learning to bear in a context relevant to your workplace.
Unfortunately, the conventional manners in which we are taught are ill-
suited to keep up with the accelerating pace at which we are expected to
learn new facts and absorb new ideas. Many aspects of the educational
system have not changed in substance since about AD 1100.
 You have a wise person at the front of the room, delivering information
from a prepared text. You have rows and rows of students, sitting on
uncomfortable, creaky wooden benches, trying to take this word stream and
turn it into something they can use. This scene would look no more out of
place in a medieval church than it would in a twenty-first-century university
auditorium at most of the world’s twenty-six thousand or so higher
education institutions. Interestingly enough, the University of Oxford was
born out of a centre of ecclesiastical teaching on the same site about 850
years ago, but has since become renowned for its ‘tutorial’ system, where
two or three students engage in active dialogue with a professor about work
they have prepared beforehand.
The way the tutorial works is that you complete an assignment and show
up to meet with your tutorial partners and your fellow (your professor). You
get feedback on your assignment, discuss a topic and lead up to the next
assignment, which you then work on independently. The tutorial system,
also used at Cambridge, incorporates some of the techniques of what is
considered ‘best practice’ based on study of cognitive and neuroscience, but
was developed more than 130 years ago.1 Unfortunately, most universities
don’t operate that way. It’s more cost-efficient to jam five hundred students
into a large lecture hall and have one instructor at the front broadcasting
material. It’s more efficient to grade multiple-choice exams than to have the
student assessed based on their ability to speak and communicate
effectively about their ideas.
The COVID-19 crisis laid bare the limitations of the conventional
learning model when, overnight, universities worldwide instantly had to
convert into delivering over videoconference. One educator shared with me
that she was counselled simply to deliver her standard three-hour lectures
on Zoom, and that no other preparation was required. Spoiler alert: it’s not
going to go well. Another educator shared with me that he feels that
students in the back half of the semester, the virtual part, got about 20 per
cent of the learning value that they should have.
A better way is needed. This chapter will help you begin to understand
what is wrong with the old way of doing things, and what the new way
looks like. Later on in the book, we’ll talk about how artificial intelligence
is beginning to be used to improve the ability of students to learn and work
in small groups, including (in a nifty sleight of meta hand) to learn about
artificial intelligence.
THE PROBLEM WITH EDUCATION
Our current educational system does a good job of training creativity and
mental agility out of us, by punishing those who ‘colour outside the lines’
and rewarding conformity. When we are children, we possess a great deal of
cognitive flexibility. We are able to invent and play, creating stories with
our friends and imagining new worlds. Then, as we enter conventional
education systems, we are trained out of our cognitive flexibility. Almost
every major educational system currently extant puts people into rules-
driven environments dictating what you learn, how long you learn, how
quickly you learn it and how you spend your day.
To a greater or lesser degree, this extends all the way through our
undergraduate educations. It also lines up with graduate education, to such a
degree that I often say it takes a couple of years to retrain graduate students
for useful employment. I say this only half-facetiously; I have taught a
number of graduate students, particularly MBA students, who are
exceptional executives and who use the university environment as a
carefully calibrated tool to achieve the next step in their careers. I also see
many students who get lost in the constructs of academia, who overestimate
their market value based on the nature of positive reinforcement they’ve
obtained in a top graduate programme, and who then flounder for a couple
of years after graduating until they find their footing. All too often I feel
that we do our students a disservice by providing theoretical constructs and
intellectually interesting ideas, but then failing to connect those to
management practice. We reward those students who agree with our
epistemologies and leave them with an abstraction of work rather than the
ability to work. For my part, I try to make sure my classes bridge between
seriousness of intent of academic respectability, providing perspectives and
facts grounded in rigorous academic research, and useful applicability in
real settings, delivering practical tools and real-world examples so that my
students may take theory into practice.
Now let’s think of the successful professional in today’s competitive jobs
environment. Twenty-five years ago, you needed to understand what the
internet was and what it meant. Ten years ago, you needed to understand the
cloud and why it was changing networked systems. Today, you have to
assimilate information about technologies like blockchain and artificial
intelligence, as well as the business process changes necessitated by the
pandemic. In five years, it’ll be some other disruption, perhaps quantum or
nano or something else. Most of my students have been working
professionals who need to keep pace with disruptive, technology-driven
change, but who also need to excel at their day jobs while doing so. They
aren’t taking two years off to pursue knowledge. They are trying to pick up
innovative skills and capabilities while still employed and building their
careers.
How can you keep working and also keep up?
You need to develop better skills for acquiring new information. You
need to acquire it faster, you need to understand it more deeply, and you
need to stay on top of what is relevant so you can pick up the next set of
skills. Education is no longer a fixed duration event (perhaps four years of
undergraduate and two years of graduate studies). In the AI-enabled future,
you need to be able to acquire new knowledge every six to twelve months.
Sounds impossible? It isn’t. But it may require retraining your brain.
Computer systems can help and we will discuss this later in the book. In
order truly to augment your career, however, you need to begin by
augmenting the human computer: your mind.
I have some good news for you: there are certain cognitive skills that you
can develop that will help you in your everyday work, and also help you
navigate and optimise how you work with artificial intelligence. Preparing
your brain can help you prepare for the AI future.

REMAKING YOUR BRAIN

I will suggest five principles that can help you develop greater cognitive
flexibility and ability to acquire knowledge, and once you’ve mastered part
of these there are others you can begin to consider:

1. The importance of practice

2. The benefits of reflection
3. Sustained and gradual change
4. Peer learning
5. Creative exploration
1. The importance of practice
I feel good about the value of the knowledge in this book, and I feel that
I’m sharing important information with you. It’s only going to be relevant
to you, however, if you try to apply the lessons. If you simply read the book
and stick it on your shelf, while it might provide you with fodder for some
cocktail-party conversation, it’s not going to deliver career transformation
or position you for better outcomes as more automation enters the
workforce. For that matter, definitionally ‘innovation’ is something new
and useful that is put into practice. If it’s just new, but no one uses it, it’s an
‘invention’. And it is on this dimension of innovation that human beings
still hold a clear and sustaining edge over machine systems.
When I teach with my colleagues at MIT, University of Oxford or
Imperial College London, the architecture of the classes is designed to
weave together frameworks (‘scaffolding’, in learning science terms) that
help you construct mental models with examples, so you begin to
understand how the idea looks in a real-world setting, and activities where
the students try out their own interpretations of the ideas. My frequent
collaborator, Professor Alex Pentland, likes to say that the meaningful
learning in his classes happens outside the classroom: it’s happening in
small group discussions among teams of students who are taking the
academic frameworks and converting them into something they intend to
apply in a tangible setting.
Building a mental model will help your practice convert into longer-term
knowledge. Mental models are where you abstract a set of facts or ideas
into a framework, which then helps you address a new situation that might
not be exactly identical to what you studied. It relates to how your brain is
more like a machine learning system than it is an expert system. Peter
Senge describes mental models as an internal picture of how the world
works.2
If you train and practise exhaustively throwing a hard, solid cricket ball a
fixed distance in a windless environment like a closed room, you might
achieve adequate performance at throwing that one type of ball in that one
setting. You won’t necessarily have a construct of different types of balls or
the effects of wind and gravity on throwing different distances. Faced with
an air-filled beach ball, you would perform miserably at trying to get it to
hit a target if there’s a light breeze. Varied practice can help you construct
mental models of different activities as you put theory into action, versus
rote memorisation of repetitive activity.
As a human being, you are still better at doing a number of tasks than
even a very good AI. An autonomous car might have sensors and LIDAR
(measuring distance through use of lasers) and lightning reflexes, but it
struggles to distinguish stationary and moving objects at times. Tesla’s
Autopilot, a precursor to truly autonomous vehicles, had its first human
fatality in 2016 when it failed to distinguish between a white tractor-trailer
against a bright sky, and when the trailer crossed in front of the car, it drove
right underneath it, crushing the windshield into the human operator.3 Your
human brain is able to classify ‘lorries’ as a subset of ‘moving vehicles to
watch out for when driving’ as well as ‘sky’ and ‘road’, and notice that if a
vehicle is suddenly veering in front of you, you should apply the brakes.
Your mental models of driving enable you to deal with a variety of
circumstances, even if you’ve only seen blue tractor-trailers or only white
cars prior to an encounter like the one that the Tesla failed to handle in
2016. Since then, Tesla has made several advances, of course – its auto
safety report in 2023, not without controversy, claimed that with Tesla
Autopilot, they have only had one accident for every 6.26 million miles
driven, making it significantly safer than human drivers.4 Most AIs don’t
have the same amount of data as Tesla does, and there are numerous AI
applications that are more complicated than driving, so – for now – you
have your ability to make mental models working strongly in your favour.

2. The benefits of reflection

Your brain can only absorb so much knowledge so quickly. ‘Wait a minute,’
you might say to me, ‘you just told me I need to hurry up and learn
something new.’ This is also true. Don’t mistake the exhortation for regular
acquisition of new skills with a false claim that you can jam the information
into your brain quickly. The reason that swotting up on some facts the night
before an exam doesn’t create durable knowledge is that your mind is pretty
good at getting rid of information it doesn’t need to hold on to any more.
Once you’ve sprayed the facts on to the test paper, the memories quickly
degrade.
Professors and non-university training professionals love pushing ‘block
classes’ or ‘block training’. The reason is not because it’s the best way to
teach you; it’s because it’s more convenient for the professor from a
schedule perspective. For that matter, the education industry has taught
educational participants to think that training is something that you clear out
a few days for, go on a retreat over a weekend or for a set of several days in
a row, and they can ‘get it out of the way’. It runs contrary to how our
brains truly function.
What happens when you first encounter new information is that it’s kept
in short-term memory. Your brain takes in a massive amount of information
every day. Most of it is unimportant – do we really need to retain the price
of petrol that we saw on a sign that we passed while driving on the
motorway? We probably remember too many advertisements and internet
memes but that’s a different discussion. Most of the masses of data that our
eyes, ears and other sense organs process is evaluated in short-term storage,
determined not to be salient, and discarded.
Some data, however, is deemed to be worthy of holding on to. Perhaps
you have disciplined yourself, as you attempt to learn a new subject, to
classify something as ‘important’. Maybe you try repeating the fact over
and over. By doing so, you are actually strengthening the connections
between the neurons that store that data. Your ever-efficient brain will then
engage in a process known as ‘consolidation’, where the chemical and
physical structures of the brain move around to shift that particular piece of
data from short-term to long-term storage (from the hippocampus to the
neocortex, if you want the structural definition).5
Sleep is an important part of this process on several levels. It allows for
your neurotransmitter reservoirs to replenish, which lets you think more
clearly. It lowers your cortisol levels (a stress hormone), which helps you
better remember things (high levels of stress have been shown to have a
deleterious effect on memory and recall). And one of the major cognitive
theories of memory suggests that your brain is performing a form of
‘housekeeping’ on your memories, moving information around in your
brain and putting the important bits into long-term storage.6
One tool I like to use when I am forced to teach block classes is to
provide an exercise or activity on one day, particularly where the students
talk to each other, let the students sleep on it, have them journal and reflect
on the prior day’s work first thing the next morning, and then discuss the
ideas again in an iterative fashion. I find people are able to harvest new
insights through this process of engagement, debate, consideration and
revisiting. Professor Marc Ventresca and I used a variation of this technique
in a block class we taught at Saïd Business School called Oxford Fintech
Lab, which helped propel forward several startups and, according to the
students, was a life-changing experience (our last iteration of Oxford
Fintech Lab had a net promoter score of 100, which is a measure of client
satisfaction – it’s impossible to do better than 100).
Your brain does better when it has time to mull over an idea, to reflect on
it, and to figure out where that particular piece of information fits within
your overall store of knowledge and your array of mental models. It needs
to engage in ‘reflection’. Ideally, this means you take classes or learn new
materials in other ways by spacing out your learning over a period of
several weeks.

3. Sustained and gradual change

The only way you can build a better way of continuous learning is to
engage in sustained change and to allow yourself permission to have it take
time. Altering your learning habits is not going to be an instant experience,
so you need to be both patient with yourself and disciplined. It’s better to
set aside thirty minutes every day to work on this instead of trying to do
four hours all at once on the weekend. You want to try out new habits a
little bit every day, and cumulatively build on those activities.
Think about training for a marathon or even a 10K: you don’t wake up
one morning, run several hours a day, and five days later run the race.
Instead, what you do is start by stretching. You run a few kilometres. You
rest for a day. You run a few more kilometres. You rest. Gradually over a
period of weeks or months, you prepare your body for the race.
Modifying your learning habits to create an ability to continuously
acquire new knowledge is a patient process for which you need to allow
time. Learning is intentional. By this I mean you need to set your intention
to learn, and very deliberately maintain your discipline of building new
memories and new understandings. It’s not something that can simply
happen to you, and it doesn’t come from passively consuming an audiobook
or a TED Talk, as intellectually stimulating as those activities can be. They
may be part of a path to learning, but they’re not sufficient on their own.
For more than fifteen years I had wanted to write books. I had pages of
notebooks filled with various scribblings, but the last time I had seriously
sat down and written anything of any length was when I took playwriting
classes in my undergraduate studies. To migrate from aspirational to actual
writer, I began by drafting op-ed pieces and LinkedIn articles about various
topics I found interesting (I’m now up to a cumulative fifty-five thousand
words on LinkedIn and counting, which is about a full-length business book
by itself). I started compiling white papers with co-authors, and eventually
edited volumes where I contributed chapters. By the time I sat down to
write Basic Blockchain (2020), I had a coach lined up to help me crank out
the manuscript (we had a very tight deadline), and the comfort of knowing
that I could push past creative inertia even if I felt ‘blocked’ and could sit
down and write every day. I didn’t always write equally well each day, and
editing was required, but I was able to crank out the first draft of the
manuscript quickly because I had built up the discipline of writing over a
two-year period.
Any new discipline, whether it’s the discipline of writing or the
discipline of learning or something else altogether, requires both the
introduction of new habits and behaviours (which will take some time for
you to get used to and begin to engage in fluidly), and the repeated focus to
ensure that you stay on your course. Keeping track of your progress,
celebrating important milestones and sharing your achievements with your
friends will help you provide yourself with positive reinforcement to keep
going.

4. Peer learning
One of the best ways to learn is through peer education. Medical education,
for example, runs on the principle of ‘Watch one, do one, teach one.’ It’s
very high touch, insofar as the ratio of instructor to student is one-to-one or
perhaps at worst one-to-four – this is not something that scales. It’s lengthy
as well: four years of medical school followed by several years of advanced
training. When we want to instil mission-critical learning, however, where
lives are literally at stake, we rely on this peer learning-based model.
You’ll notice I use words like ‘learning’ a great deal, and am disdainful
of words like ‘education’. My former business partner Beth Porter and I like
to say, ‘Education is what happens to you, learning is what you do.’ And
instructor-led training is one of the least effective forms of education, the
‘sage on a stage’ model where a learned individual spews forth facts at you
and you are supposed to write them down and figure out what they mean.
 The best way I can coach your learning is to provide a little bit of
information, then you do something to apply it, ideally in a group, then I
provide a little more information, we discuss it as a group, then you go off
and work on your own trying to apply the information and elaborate on
what I’ve said. Finally we come back together as a group and we (not just
me, but all of us) critique the various projects to provide further insight.
I have run into impatient students who resist the group-project model.
‘Why should I have to be stuck with a bunch of people who aren’t as smart
as I am?’ I have been asked, in one form or another, over the years. I have
observed that some of the students who protest most vociferously about
being allocated to a team are among those who benefit the most from the
team activities.
One of the benefits of peer learning is that if you are forced to explain a
subject to someone else, you tend to understand it better yourself. Your
intuitive grasp of a particular topic, which might be half-organised in your
mind, acquires greater structure and solidity when you are required to
reduce its principles to an explainable form that someone else can absorb.
Here’s some interesting research out of MIT from a few years ago: the
scientists took 699 individuals, divided them up into teams of between two
and five, and set them to work on complex tasks. The participants were
evaluated on measures such as intelligence quotient (IQ). The most
successful problem solvers were not the teams with the highest average IQ,
nor were they the teams that had an individual with the highest IQ (the
‘surround a genius with minions’ strategy). Instead, the teams that
performed the best, that had the highest collective intelligence, were those
with an exceptionally high emotional intelligence quotient or EQ.7
Part of what’s going on is the diversity function: having many different
perspectives on a problem increases the probability of finding a solution to
it. In fact, Northwestern University research conducted on over 3.5 million
workers in 154 public companies validates that high levels of diversity,
coupled with a disciplined and scaled innovation function, produce the best
financial outcomes for corporations.8
What is also happening in these dynamic group interactions is that
instead of ruminating on the problem inside their own heads, the team
dynamic forces individuals to begin to trade concepts back and forth,
strengthening the best solutions and discarding the flawed ones.
 Over the past five years, Professor Pentland, one of the authors of the
original collective intelligence study, has been working with Beth Porter
and me to create a repeatable way to measure these interactions (known as
‘instrumenting’ them); to begin to stage interventions on them; and over
time to optimise team performance, whether that’s in a learning setting or
other kinds of activities. We created a spinout company around this called
Riff Analytics that conducted US National Science Foundation-funded
studies in these areas.9 As you will learn in Chapter 9, we have discovered
you can obtain better learning outcomes or craft new knowledge if you
bring together the right mix of people in the right kind of structured
environment, and have AI help you work better together.

5. Creative exploration
Mastering peer learning can also open up more possibilities of creative
exploration. Creative exploration (or ‘serious play’ as the LEGO
corporation would have it) is a building block that leads to other, more
advanced, concepts such as gamification and complex problem solving. In
the AI future, any analytic or quantitative task is going to be done faster,
more accurately and with greater scalability by a machine. The role of
humans will be focused on emotional intelligence-dependent functions
where creativity and inventiveness are required. So far we haven’t
succeeded in creating an artificial intelligence system that is creative by the
definition that we use in human systems.
How can we make ourselves better at creative exploration?
Someone might derisively describe an individual as having ‘the mind of
a child’. When I hear this expression, I associate it with being open to new
experience, neuroplasticity (meaning your brain can readily adapt to new
ideas and is better at learning) and being endlessly inventive. The rigid
ordering of conventional educational systems takes that creativity and
playful spirit, and typically shaves off any bits that don’t conform to a
dogmatic adherence to memorisation and regurgitation of fact. We then
expect these people to go into the world and build new businesses or solve
problems for society. Some of us are lucky: we have exposure to a well-
designed undergraduate experience that focuses on critical thinking instead
of storing facts. Storing facts becomes increasingly useless as the rate of
change of knowledge accelerates. Critical thinking is timeless.
 Children engage in creative exploration all of the time. They do so solo,
making up imaginary friends or scenarios. They do so in groups,
collectively envisioning heroic settings or strange new worlds. Listening
and talking to each other, they trade ideas back and forth, experimenting
with concepts, throwing them away effortlessly and trying out others as
they endlessly create.
A famous creative collaboration experiment known as the Marshmallow
Challenge has been conducted globally with many different types of groups.
In eighteen minutes, with limited resources (a marshmallow, some string
and tape, twenty pieces of uncooked spaghetti), competing teams work to
build the tallest tower. It opens a fascinating window into problem solving
and group dynamics. Tom Wujec has a wonderful TED Talk explaining the
key insights, which is worth reviewing at:
www.ted.com/talks/tom_wujec_build_a_tower_build_a_team.
Here are some of the findings:
Five-year-olds are among the top-performing groups. MBA and law
students are among the worst-performing team configurations, as they
waste precious minutes navigating status and planning. They are searching
for the one best answer, rather than experimenting and discarding several
ideas rapidly. Five-year-olds jump in, immediately start trying out different
configurations, with subtle social signalling as they grab pieces and interact
with each other. What’s notable is that not only do the teams of young
children perform well, but that it’s a team activity, rather than a single
designer coming up with the idea. And the children often produce the most
interesting structures.
The most vibrant, scalable and repeatable innovations tend to come from
creative collaborations, not solo genius. Diversity, as it turns out, is a
necessary input to effective ideation. Social scientists may refer to this as
the ‘law of weak ties’ but the core concept is that many different
perspectives being brought to bear on a problem can often reveal
unexpected avenues of solutions, or even open up new solution spaces.
People who all come from the same backgrounds and cultures tend to bring
the same collection of mental models, fact bases and modes of thought. The
more different perspectives you can introduce into a brainstorming
discussion, the greater the likelihood you can fabricate creative collisions
that produce truly breakthrough thinking.
 Accordingly, when you need to solve a complex problem at work, think
about who you solve that problem with. Who can you enlist to assist you
with the analysis and generation of potential solutions? How can you
diversify your creative inputs to the process? Perhaps you can ‘opinion
shop’ ideas with a variety of colleagues rather than simply attempting to
resolve the issue alone. You could bring several colleagues from different
functional areas or cultures together, possibly augmented by external
experts, and conduct a focused brainstorming session around the problem.
One of the crucial tasks in unleashing your creative exploration
capabilities is getting in touch with the imaginative part of your brain,
which the educational system has worked to diminish. You can do this
through a variety of individual exercises, such as writing or drawing. You
can engage in meditation, to clear distracting thoughts, and mindfulness, to
gain heightened awareness of yourself and your surroundings. And if you
are looking to excel, you will find ways that you can participate in creative
play with others.

Now that we’ve taken the time to understand what good learning looks like,
I’m going to talk about which subjects are best positioned in the AI future.
This conversation isn’t limited to what subject you ought to study if you are
university age, because with the rate of technology evolution increasing, we
are entering an era where you will want to acquire new skills every few
years, and eventually every year, to remain relevant.
Although education revolution is needed, for now most of the world’s
employers expect you to show up with a credential from a recognised
institution, at the very least an undergraduate degree. Even if you are past
school age, as you probably are, you may have children, cousins, nieces or
nephews who are thinking about what to study. My children will be
graduating from school into the face of society’s AI max Q (what the
aeronautical engineers call peak dynamic pressure, or the point in a rocket
launch when the body of the rocket is undergoing maximum stress). As a
university teacher, what would I tell my own children to study?

READING LIST

Out of necessity for brevity we have applied a superficial lens

on cognitive flexibility. If you want to dive more deeply into how
the mind actually works, Daniel Kahneman’s Thinking, Fast and
Slow (2011) is a good place to start, or if you want a less-dense
treatise, Chip and Dan Heath’s Decisive (2013). Other books,
like Mark McDaniel, Peter Brown and Henry Roediger’s Make It
Stick (2014), are specifically focused on learning. Sanjay Sarma
and Luke Yoquinto’s Grasp (2020)* makes a valiant effort to
summarise and explain an entire body of cognitive and
neuroscience research around learning. James Clear’s Atomic
Habits (2018) is a useful guide to shifting your behaviour a little
bit each day, which can help you instil new learning capabilities.
For a bit of fun, Austin Klein’s Steal Like an Artist: 10 Things
Nobody Told You About Being Creative (2012) can help you
appreciate the difference between being creative – our goal
here – and being original – not necessarily needed.

CHAPTER 4: THE RUNDOWN

♦ Untraining and retraining your way of thinking may better prepare

you for the AI future.
♦ Engaging in regular application of the five principles of cognitive
flexibility can deliver transformational results to you.
♦ Creative exploration may be the most powerful of these
principles, and is sometimes best performed as a team versus
individual activity.
CHAPTER 5

Future-proofing: what to study and

which industries are safer?

CHAPTER 5: WHAT YOU NEED TO KNOW

♦ Creative professions and philosophy may be most resistant to AI

disruption – and may have unexpected job utility.
♦ Professions with sustained human contact, like health services,
psychotherapy and hospitality, will be more resistant to AI
displacement.
♦ Emotional intelligence and other ‘soft skills’ represent key job
capabilities that AI will have difficulty replicating.

When you think about job security, and what you might want to pursue in
undergraduate studies today, you would probably imagine that computer
programming would be a good field. After all, if AI is everywhere, don’t
you want to develop an understanding of the technology that runs our lives?
Computer programmers will be replaced first by systems that are ‘no code’,
where a human can simply tell the computer what it wants to happen, and
eventually by AIs that design themselves. Perhaps the last human
programmers to go will be those whose focus is on the theoretical design
and architecture behind systems, but there exists the very real possibility
that twenty years from now we will be looking at a world where ‘computer
architects’ will be fairly abstract conceptualists, or possibly the software
engineering function will consist of a single administrator who provides
broad guidance to an array of machines. While not an immediate-term
worry for next year or the year after, the matter of job security becomes
more pressing as we look ahead ten or twenty years and think about career
longevity.
The answers to the question, ‘Which subject areas are best suited to the
AI future?’ might surprise you.

A BRIEF ASIDE ON WHAT NOT TO STUDY

Programming languages, a degree in blockchain, any applied field that is
heavy on analytics and computation – these are all likely to be displaced by
smart machines or made obsolete by technology innovation. On multiple
occasions in recent years I have had to dissuade people from pursuing a
‘PhD in fintech’ and encourage them to think about foundation-based
degrees such as ‘financial services’ or ‘technology innovation’ with a thesis
that focuses on fintech, but that provides them with longer-term benefit than
a five- to ten-year tech cycle. Automotive engineering, too, may be a labour
of love more than profit: we have very likely passed peak auto ownership,
between the effects of technology-enabled cars, ride-sharing systems and
the rise of autonomous vehicles.1
The accounting professions, once considered among the safest career
paths, actuarial students and even pre-law, as we reviewed earlier, are all at
meaningful risk. We’re setting aside the trades, since this discussion is
focused on university education, but it’s worth mentioning that everything
from welding to window cleaning faces a good likelihood of robotic
automation.2

CREDO QUIA ABSURDUM EST

As I mentioned at the outset of our voyage together, philosophy may
provide an intriguing platform from which to enter the field of artificial
intelligence. The future ‘AI programmer’ may be more in dialogue with the
machine than giving orders to it, and will need a grounding in critical
reasoning and formal logic. Conversational AI, and AIs that don’t require
computer programming in order to get them to do something, are frontier
fields that are rapidly maturing into market realities. Indeed, nearly ten
years ago Oxford physicist Professor David Deutsch postulated that
philosophy could be the key to enabling artificial general intelligence,3 and
his statement holds as much validity today as it did back then.
 Philosophy, however, offers a grander field of view than just AI
development. As tech billionaire Mark Cuban noted at the South by
Southwest festival in 2017, a philosophy degree may be able to help you
assess facts and circumstances with a keener eye, and take in the bigger
picture.4 He should know – his fortune was created by launching
Broadcast.com. Not only did Cuban spot the trend of rich media over the
internet, he understood the importance of intellectual property rights to
airing professional sports over digital media, then was not only
perspicacious enough to identify the gap in the market, but skilful enough to
execute with his partners on securing those rights. The big picture led to the
big payoff, not (directly) a better technology.
Credo quia absurdum est. I believe because it is absurd.

SPARK OF CREATIVITY
The creative arts remain a uniquely human endeavour. While you may see
AIs essaying digital quill and brush to create ‘art’, I find the images
(however exciting they may be from a conceptual standpoint) lack empathy.
There is an indescribable emotional affinity conveyed by a human artist that
machines still lack, whether that be in fine arts or in performance.
Design, too, is a field where machine inventions remain a novelty and a
bit limited, and at a minimum a human mind directing a machine
implementer seems more likely than wholly replacing humans throughout
the aesthetic process. Notwithstanding advances in artificial intelligence
systems, the human touch is still essential to marrying meat-based cognition
(i.e., Homo sapiens) with silicon-based technology.
I ask Sanjeev Vohra, who recently managed Accenture’s fifty thousand
AI developers, what advice he would give his own children, or my children,
for a stable career when they graduate. Sanjeev has a son who has just
finished university who wants to go into a new speciality area at the
interface of people and machine. As he puts it, ‘designing AI to meet the
human’. Just as the digital media era brought User Experience (UX) design
to the fore, the artificial intelligence era is bringing forward this AI
interface design role. Sanjeev’s tantalising hint of an answer lies in the field
of bringing human ingenuity to smoothing the interface between human and
machine.
THE HUMAN TOUCH
Healthcare remains a field where in the foreseeable future – ten, twenty or
possibly thirty years – humans are still needed front and centre in the
experience. Not only doctors, but nurses, pharmacists, physicians’
assistants, physical therapists and other healthcare professionals. While AI
is beginning to show up in various areas of medicine in a supporting role,
such as cancer diagnostics and treatment plan recommendation,5 it is
ultimately human decision makers who are accountable, and human
medical professionals working with the patients. Dentists and dental
technicians represent a skilled profession that will be difficult any time in
the next twenty to thirty years to replace with AI-powered robots.
It’s not only moral and ethical judgement made by humans that
accompanies deciding a course of treatment. If a medical professional
interacts positively with a patient, they are more likely to follow care
instructions and even engage in spontaneous activity to help their health
journey. When patients develop an emotional connection with their health
professionals centred on trust, it leads to better health outcomes.6 This is not
yet something that can be readily replicated by a machine dispensing care
instructions.

THE ART OF THE MIND

Psychology, neuroscience and cognitive psychology are all fields that
benefit from AI tools and enhancements, but there remains a need for
human invention, interpretation and interaction. Notwithstanding our good
friend ELIZA, effective therapy requires human-to-human connection.
While some organisations are experimenting with creating emotional robot
companions, some of which show early promise, we are still years away
from fully replacing the functions of a trained therapist, or the nuance and
finesse exhibited by an industrial psychologist employed within a large
organisation. Teaching a computer to understand the intricacies of the
human brain to help a patient become more high-functioning and self-
actualised is beyond the grasp of today’s AI systems, and probably
tomorrow’s AI systems. The day after tomorrow may see a different kind of
AI emerging, which we will address later.
PREPARATION FOR THE UNKNOWN
Knowledge becomes commodified in a world where we all may have access
to brain implants that feed us data in real time. Facts and figures are at
everyone’s top of mind, on demand. It’s the inverse of today’s ‘Google
brain’ where people have grown forgetful because all knowledge is but a
search bar away. Instead, we may have a world of machine enablers tied to
our consciousness, constantly feeding us what we need to know, when we
need to know it, and in the process accelerating our throughput of
information. This is, perhaps, a glimpse of what Accenture’s Vohra is
talking about when he describes the dawning age of intelligence.
The concept of a tighter linkage between people and how they obtain and
use information through machine filters also introduces the risk of greater
fragmentation of society and greater bias. If the machine assistants we
employ to feed us data apply screens that have been improperly designed or
perhaps improperly evolved, we can find ourselves down a very narrow
path of thought. If your brain implant discovers you are uncomfortable with
certain areas of thought – perhaps you chafe at certain information that
disagrees with what you currently believe – will it helpfully try to make
your life better by not showing you information it perceives you don’t like?
This is, in essence, what has already happened with Facebook’s algorithms,
to deleterious effect as people are polarised into an array of fragmented
echo chambers where they only interact with other people who agree with
them. It does not take a great imaginative leap to perceive a future where
this ‘Facebook Effect’ becomes commonplace across all modes of assisted
access to knowledge, whether sold by Mr Zuckerberg or by another
technology titan.
Ahead of that more dangerous vision, we continue to see encroachments
by AI into more and more professions, coupled to an increasing rate of
technology innovation across society. Therefore the undergraduate
education that can prepare one for the AI future may be universally
grounded in critical thinking, cognitive flexibility and studies that foster
emotional intelligence. In a world where AIs assume repetitious, analytical
tasks, soft skills assume greater importance. This reverses an educational
trend of the past thirty or so years favouring preprofessional tracks like
finance or business or computer programming, increasingly steering
students into semi-vocational training for perceived ‘safe’ jobs. Engineering
schools rose to pre-eminence, while small liberal arts colleges withered.
Now, however, the need may grow for liberal arts education.
Since no one gets a degree in ‘soft skills’, what are the courses of study
that can help you prepare for an AI future?
There are certain disciplines and associated co-curricular activities that
actively develop soft skills. Philosophy, creative writing, dramatics, all can
lead to interactions with others that hone both critical thinking and problem
solving in groups. An English language student who is the editor of the
university newspaper learns to deal not only with language, but also with
motivating and managing teams of people who aren’t working for money, in
an all-volunteer effort, but instead take pride in the work that is being done.
Getting a group of lightly organised undergraduate students to engage,
focus and meet deadlines with, at best, remuneration comprised of pizza
and beer requires nothing short of a virtuoso performance in leadership and
emotional intelligence.

EMOTIONAL INTELLIGENCE
The ‘soft skills’ ingredients for emotional intelligence were for many years
dismissed in education in favour of teaching more ‘tangible’ facts, figures
and capabilities. What’s interesting to observe, as awareness of the
relationship between soft skills and corporate success rises, is the growing
focus of institutions on identifying, hiring for and cultivating emotional
intelligence or ‘EQ’ in their workforce. The field is still in its infancy;
managers trained in traditional methods will pay lip service to the need for
EQ and then promote for, and incentivise around, quantitative skills
(because they are easier to measure) – at least until they make entire
departments redundant by replacing them with AI. Emotional intelligence is
harder to outsource to a machine, although newer forms of AI include those
that actually help measure and develop EQ, as we will discuss in Part III.
Before we do, however, it would be beneficial to extend our discussion
around areas of study into a strategic lens on defensible industries. Where
do you want to develop your career path, in a manner that is resilient to
potential automation from AI? Which industries will prove to be the most
resilient to the AI future, or at a minimum are best positioned to benefit
from AI technologies?
 With these questions in mind, I have to admit it is a struggle to argue
compelling answers. There is no one industry that is entirely ‘AI proof’, no
magic sinecure that will provide thirty or forty years of untouched income.
Where there are people, there is labour. Where there is labour, there is cost,
and with cost, the inexorable forces of capitalism and the public markets
will drive towards creating market efficiencies, i.e., reducing labour. Some
countries, such as Germany, have involved labour unions in the governance
of the company, but protecting high-wage jobs will create natural
inefficiencies when competing against companies domiciled elsewhere.
Gerard Verweij, Global Data & Analytics Leader for consultancy PwC,
said, ‘No sector or business is in any way immune from the impact of AI.’7
I am inclined to agree. You can run, but you cannot hide. You can, however,
focus on industries that are more resilient to AI disruption, and apply
yourself within those industries in a way to extend your lead over the
machines.
The answers to AI adaptability – career areas that help you evolve along
with the evolution of AI – will align with the selection of the best
undergraduate studies I described previously, but also include blue-collar
professions. The industries most protected require creativity, emotional
intelligence and other ‘soft skills’, or they require capabilities for which
robots are impractical at our current state of the art.

HEALTH
Health services will remain the province of human practitioners, for the
most part, for some time. Select areas of medicine such as radiology or
diagnostics are showing suggestive results where AIs equal or outperform
human doctors,8 but current standards of care in medicine require a human
decision maker with human accountability. As well, patients prefer
interacting with people, particularly in areas like physical care. We have yet
to create an AI-powered robot that can give a bed bath that people feel
comfortable with. Not because we can’t automate robotic medical bathing –
more than a decade ago, researchers at Georgia Tech created a robot that
demonstrated it could give sponge baths9 – but rather because most people
would rather receive empathy and care from another human being.
In the more distant future, when truly humanoid robots emerge that are
indistinguishable from people, or if a generation raised on Roombas reaches
retirement age and is fine with being touched by machines for intimate
personal care, this may change, but until then the human health worker is
secure.

LIVE PERFORMING ARTS

Live performances, whether theatre, music or dance, remain a domain that
benefits from the human touch. Human beings are tribal and pack-like, and
have an intrinsic need for connection with each other that can be modulated
through live performance.10 Watching a pre-recorded performance isn’t the
same; it loses the feedback loop that a performer creates with an audience.11
Until we can create human-like robots that very closely resemble human
performers and are able to interact with an audience dynamically to create a
connection with the audience, we will need human performers. The electric
connection that is created by a crowd that is going through a shared
experience is difficult to replicate in a digital environment. For pre-recorded
music, AI-generated product has started appearing – but not yet at a quality
where it would create a ‘hit’. Legal issues have begun proliferating as well,
around a host of copyright-related matters.12

RETAIL, HOSPITALITY AND OTHER PERSONAL

SERVICES
I recall when the eCommerce boom was first gaining momentum in the
1990s that people were worried Amazon and eBay would spell the end of
in-person retail. To a degree, this prognostication was correct: Amazon is
exploring converting abandoned J. C. Penney and Sears stores into
warehouse facilities inside shopping malls.13 At the same time, people still
prefer to see, touch and try on items where they can. Even mighty Amazon
has been exploring a high-tech variant of real-world shopping, albeit with a
high-tech twist: there’s no checkout, you simply walk out of the store and
wireless transmitters on your phone and on the products you are buying
take care of your purchase automatically.14 And, of course, they made a
$13.7 billion bet on acquiring gourmet grocery Whole Foods.15 Restaurants,
likewise, are not a robot-friendly environment. People prefer human servers
in most countries.
 Other personal services remain stubbornly personal, and likely will for
some time. Massage therapy and physical therapy remain hands-on,
personally delivered services, in spite of various massage chairs and other
machines that can move and stretch the body. Hair stylists and nail
technicians are likely to be difficult to replace by machines, as much for the
personal interaction in the salon as for the technical intricacies of the work.
It’s hard to feel pampered by a robot.
We may, however, end up seeing a class-based bifurcation. As robot
labour gets cheaper and cheaper, we may see lower-end delivery by
machines, with the human touch reserved for the ‘posh experience’. It’s
difficult to argue that the McDonald’s restaurant worker adds significant
value over a machine. You could more quickly and more cheaply get your
gut-busting fast food from a robotic system, and you wouldn’t need a very
smart AI to run it. Conversely, it is easy to imagine that twenty years from
now you will still talk to the (human) maître d’ at Chiltern Firehouse in
London or Waku Ghin in Singapore, or take a recommendation from the
sommelier at Le Bernardin in New York.

ENERGY SYSTEMS
Installing wind farms or solar panels will still require people in the next ten
or twenty years. Further out, it may be possibly for robotic labour to replace
humans; for now, we are still reliant on human hands to erect wind towers
or place panels. The power industry, more broadly, relies on human
implementation and deployment of power systems, and at least some human
intervention on the maintenance of them. Partly, there are mechanisms that
still benefit from the human touch. Partly, there is the conscious decision
not to put an AI completely in control of, say, a nuclear power plant, but
instead to require human-in-the-loop technology. Extractives such as oil and
coal, despite various bits of automation, still rely on a significant amount of
human labour both upstream and downstream.

CONSTRUCTION
Although I am not as sanguine as some with respect to the ability of the
construction industry to resist automation, there is an argument that says it
will be a very long time before AI displaces human. This is to some extent
due to the nature of the work, given that the ground-level hand labour in
home construction occurs in a wide variety of uncontrolled environments
with numerous variables, which a moderately trained twenty-five-year-old
human can easily handle but a very expensive and sophisticated robot today
struggles with. To some extent, it is because of the industry structure of
construction, with numerous specialised subcontractors working under the
direction of a general contractor. Each element of home building has an
array of complex choices and these different choices – ranging from what
kind of fixtures you choose to the kind of flooring you put in – need to
seamlessly integrate with each other.16 While in theory robots can do parts
of these tasks, and in theory you can assemble prefabricated homes with the
design selected from a catalogue and the parts snapped together out of a kit,
most people don’t. New home construction is the purview of builders
operating in a complex ecosystem of contractors.

GOVERNMENT
The industry of government will be among the last to accept widespread
adoption of AI systems. I don’t mean selective use of AI for certain
purposes – the company Palantir has reached over $2 billion of revenue in
2023 from a few handfuls of government contracts – but rather widespread
replacement of human government workers with AI systems. I refer
specifically to the staff professionals in government versus politicians.
While Japan may have had an AI mayoral candidate,17 that is an outlier
event rather than an immediate trend. The rank-and-file government
workers who manage various government departments and programmes are
likely to remain in place for years to come.

AGRICULTURE AND ANIMAL HUSBANDRY

AI is slowly trying to work its way into the agricultural industries, but
human labour will be required for the care, feeding and harvesting of plants
and animals. Even with innovations such as vertical farming and farm
automation systems, we still have people involved at numerous steps in the
farming process. We will need much more powerful and simultaneously
less expensive robotic systems, which are perhaps thirty to fifty years away,
in order to have widescale replacement of people with machines,
particularly in developing countries.

CHECKLIST TO ADAPT YOUR FUTURE

The industries that will likely prove best suited for a significant proportion
of human involvement, not unlike the individual professions described in
Chapter 6, rely on one or more of the following characteristics. When you
think about career and industry strategy, see how many of these attributes
you can attach to a particular area:

♦ requires a high degree of empathy or emotional intelligence;

♦ involves repeated physical contact with human beings, living
animals or plants;
♦ entails a high degree of judgement in the face of rapidly changing,
variable or novel circumstances;
♦ benefits from a significant level of creativity or artistic expression;
♦ persists in requiring intuition for competitive advantage or greater
success.

AND YET …
Thinking of an industry sector that will forever be completely untouched by
AI automation is difficult in the extreme. As artificial intelligences get more
sophisticated, the range and nature of fields in which they can be deployed
are increasing.
Hiding out in a particular field isn’t the answer. If you want to win at
work in the age of AI, you need to find ways to make AI work for you,
rather than against you, augmenting your capabilities, rather than seeking to
replicate them.
You will need AI and digital literacy to succeed, even in primarily
human-centric professions like design or health services.
The next section will explain what this emerging model of Human+AI
hybrids looks like, and how you can bring artificial intelligence into your
competitive toolkit.
 CHAPTER 5: THE RUNDOWN

♦ Analytical and quantitative professions are under greatest threat

from AI displacement.
♦ Careers that require human-to-human contact, such as retail, as
well as careers that involve complex physical tasks, like energy
systems, and construction, will take longer to disrupt than other
domains.
♦ Ultimately, most professions will face some degree of AI
disruption over time. The key to success will be learning how to
work in collaboration with AI versus in competition with it.
CHAPTER 6

What can Human+AI systems do?

Adventures in collective intelligence and
more

CHAPTER 6: WHAT YOU NEED TO KNOW

♦ Human+AI systems can help improve how people collaborate in

teams.
♦ Outcomes in critical domains, such as health services, can be
improved through engaging artificial intelligence alongside
human experts.
♦ AI digital assistants (imagine your own version of Iron Man’s
JARVIS) may represent the most powerful application of
Human+AI technology.
♦ Human+AI hybrid systems can even predict the future.

Envision the everyday, run-of-the-mill business or project meeting. The

average manager spends 35 to 50 per cent of their time in meetings. It’s a
setting ripe for enhancement by artificial intelligence.
A group of people gather together in a room or on a videoconference
call. If they’re lucky, the meeting has an agenda attached to it. If they’re
really lucky, someone is facilitating the meeting and the agenda is even
followed. This group has a conversation or discussion. Perhaps it’s a
brainstorming session, where several new ideas are generated, evaluated
and either agreed upon or discarded. Perhaps it’s a decision meeting, where
facts are presented and a series of choices is selected from to determine a
course of action.
Whatever the meeting type, there’s a form of conversation. If good
practices are being employed, not only is there an agenda, but also a
meeting ‘scribe’ who is recording what was said and generating the meeting
minutes. Unfortunately most meetings aren’t structured using best practice,
which undoubtedly contributes to the estimation that executives believe that
67 per cent of meetings are failures, wasting over $37 billion each year in
unproductive yakking.1
Even productive meetings that have positive, desired outcomes are
suboptimal compared to what they could be. I recently helped a group of
half-a-dozen senior executives who were developing innovation and growth
projects centred on novel uses of data. I was facilitating a brainstorming
meeting to generate twenty new business ideas in forty minutes (a body of
literature has been written about the benefits of design-constrained
innovation). After providing some context on the span of time and scope of
opportunity, and the types or modes of innovation we wanted to use, I put a
blank slide on to the screen and we began populating the page with business
ideas.
I had to listen carefully to each person speaking. The videoconferencing
software didn’t help much with the crosstalk, but at least I could see people
to call on them when their body language indicated a desire to contribute
something. My attention was split across multiple tasks. I needed to monitor
everyone in the meeting, making sure they were all contributing. I needed
to interpret what they were saying, and capture it in the appropriate place. I
needed to unpack their ideas, as some of the suggestions were in fact
divisible into two or three distinct venture concepts. One hundred per cent
of my focus was on facilitation. It meant I didn’t do much to contribute my
own ideas (I think I came up with one in this particular conversation), but
that wasn’t my role in the meeting. It couldn’t be my role in the meeting,
because my mental processing was focused on the soft and hard aspects of
meeting facilitation. In forty minutes we generated twenty ideas, and then
laid out a timeline with accountability attached. I found myself wishing I
had been able to spend more time thinking and elaborating on the
discussion, and not have to interpret and record it.
A friend and former business partner of mine, Tom Gardner, developed
some interesting meeting management techniques in his years of leadership.
Tom is the former CEO of what was a billion-dollar public company,
Datamonitor, and before that rose to fame as the marketing executive at
Johnson & Johnson who took Tylenol over-the-counter. He’s now enjoying
a flourishing career as an angel investor and civic booster in Sarasota,
Florida, itself home to a dynamic artificial intelligence cluster thanks to the
vibrant data science programme at New College of Florida. When I left my
position as Ernst & Young’s first, and at the time only, Entrepreneur in
Residence to form a search fund with Tom, he put me up in his house in
Sarasota at the turn of the year, and we mapped out the future (the
successful search generated enough fees for me to switch tracks for a while
and get a job at MIT, which is a story for another book).
A big, bluff man with ruddy cheeks and an imposing disposition, Tom
played professional football (American-style) briefly before going into
business. Relentlessly fascinated with innovation, Tom was an early adopter
of tablets and would regularly carry his with him to meetings. As we would
sit and talk, he would industriously type away at the tablet, taking notes on
our discussion as we’d go. It was, candidly, a little distracting – but at the
same time, at the end of every meeting, he would instantly generate an
email for the both of us (and whoever we were meeting with), describing
what we said, what we decided and our next steps with assigned
responsibilities.
Good discipline, for certain. It’s one of those management hacks that
captures relevant knowledge while it’s fresh in the mind and sets up future
accountabilities. You start to get a window into how Tom effectively ran a
billion-dollar public company.
Yet at the same time, here we were consuming the mental cycles of a
brilliant business leader tapping away at his keyboard. He wasn’t focused
on me, on my body language, he wasn’t making eye contact. His oversized
fingers, well suited for gripping a ‘pigskin’ or for crushing your hand in an
overwhelmingly dominant handshake, were pecking away at a little
electronic device instead, taking down our conversation. Tom’s attention
focused on his screen, and not on the people in the room with him, and this
resulted in a diminution of social signalling within the dyad or group.
Eighty per cent of the social signals we respond to in collaborations come
from body language. All of that is lost if someone is typing and not
watching.
 It turns out that the social cues we exchange when we sit near one
another and interact are significantly more beneficial than some people
might suspect. It turns out that they are essential for the creation of trust
among parties. On a preverbal, typically subconscious level, we are
engaged in a variety of signalling behaviours when we interact with each
other in person. If I cross my arms, you might cross your arms in a form of
mimicry. If you feel connected to what I am saying, you might open up
your body stance, and perhaps encourage me to continue to speak with a
nod, with a gesture of your hand or with a non-linguistic audio cue (‘uh-
huh’, ‘hmm’, etc.).
All of those signals of personal connectivity and emotional linkage are
lost if you are focused on your device instead of on my face.

MEETINGS, REIMAGINED
What if there were a different version of reality?
Computer systems have required sophisticated programming in order to
learn how to address and manipulate unstructured data like human speech.
Siri, Google Voice, Amazon Echo: these are technologies that can extract
words or even sentences out of spoken language. They’re pretty good at
English, as they started out being trained by engineers in the western United
States (and other tech clusters) in English language. For many of my books,
increasingly I am able to dictate a third to half of my content, and the built-
in transcription in the iPhone is able to understand me reasonably well. My
editor can share with you how much my words still need to be corrected,
but it’s a huge leap forward from the speech-to-text systems of the 1990s
that you used to have to ‘train’ by reading to them pre-recorded scripts. In
2020, a straight-out-of-the-box consumer electronic can understand almost
anyone (typically speaking in English), and provide 95 per cent or better
accuracy even under less-than-ideal conditions. The other 5 per cent is the
subject of numerous jokes in popular culture.
Me: ‘Alexa, play “Despacito”!’
Alexa: ‘Praying for desperados.’
These current-generation AI systems are less proficient when confronted
with the thirty-eight languages used by Bangladesh’s 161 million citizens,
as I discovered when I met with the country’s Minister of
Telecommunications to talk about his interests in better artificial
intelligence as a means of promoting access and inclusion in his society.
Large segments of the population were illiterate as well, so a conversational
system (not based on English-language AI) was needed that could
understand and respond in thirty-eight languages and multiple dialects. Off-
the-shelf AI wouldn’t work for them. They needed a smarter artificial
intelligence. What they wanted was even more ambitious: could this AI
actually understand what people meant, and make recommendations to
them when they called up the government for help with their taxes, with
finding a job, getting food aid or dealing with a housing issue?

MEETING OF MINDS
Let’s revisit that ‘well-constructed’ meeting we talked about earlier, and
pretend we had artificial intelligence engaged in idea learning. Imagine if a
conversational AI were listening in on the discussion between Tom and me,
automatically generating notes of what we said using speech-to-data.
Imagine further that the AI was tied in to our calendars and emails, so it had
some ideas around the context of our conversation. Imagine that AI was
able not only to transcribe our words, but also understood the meaning
behind them, and was able to extract the major themes of our discussion,
pick up the verbal cues of people being assigned responsibilities, and sketch
out a project plan.
In this new kind of human–machine hybrid interaction, we could look
each other in the eye and focus on each other, with an AI in the background
acting as an enabling digital assistant. The AI could even populate follow-
up meetings into our calendars. The drudgery of transcription, note taking,
scheduling, all get lifted from our shoulders, so we can focus on higher-
order thinking.
A piece of software known as Otter.AI can take care of this for us today.
With a few tweaks, it could tie into our project management software, and
assign and track tasks, and help us stay on plan.
Imagine what an organisation could look like if AI minders were keeping
track for us of what we want to remember. Imagine the improvements in
efficiency if we didn’t have follow-up items slip through the cracks, if we
had an unobtrusively helpful machine reminding us of key workflow
priorities, helping us manage our teams and ourselves better so that a
project came in ahead of schedule and under budget. Imagine further if
these AI systems could assess probabilities around delivery, and could help
us adjust if problems were forecast to arise.

HEALTHY MACHINES
The Human+AI revolution is not limited to the meeting room or the
boardroom. Imagine what these smart AI assistants could do in a hospital
setting. This new generation of AI assistants could actually help save lives.
One of the major causes of death in hospitals is attributable to something
known as ‘transitions in care’. It’s a term of art to describe what happens
when a patient is transferred from one medical provider to another. Over 3
million people die every year, according to the World Health Organization,
due to medical error. Poor communication among medical professionals is
one of the top factors leading to these deaths.
Let’s imagine that you are in your home and suddenly experience
respiratory distress. Your friend, spouse or partner calls an ambulance. The
emergency medical personnel in the ambulance are the first care providers
to interact with you. They take vital signs, check your breathing, listen to
your pulse. They’re capturing this information in some form, but maybe
they don’t have time to write it down accurately because they are busy
saving your life.
The ambulance pulls up to the hospital. Rushing you off on a gurney, the
medical technician quickly runs through your symptoms with the nurse or
doctor at the emergency department. Did they remember everything
accurately? Did they transmit all of the vital information?
The emergency physicians work on you but it doesn’t seem to be solving
your respiratory distress. Your anxiety goes through the roof as the beeping
of the machines and the moans of the other patients add to an already
stressful situation. You spiral further. Your throat closes over, and they
aren’t able to put a tube down your throat to help you breathe. The
emergency physicians have to perform a tracheotomy, a procedure where
they cut open part of your throat to directly access your windpipe and help
you breathe. You are stabilised, for now, but you are in bad shape. They
transfer you to the intensive care unit (ICU). A nurse brings a wheeled
stretcher to carry you from the emergency department up to the ICU. The
nurse then explains your symptoms and reviews your chart with the new
physician looking after you in the ICU.
 At each patient handoff from one care provider to the next – from
ambulance technician to emergency department physician to nurse to ICU
physician – there is a conveyance of information. Was all of the information
transmitted correctly? Was enough of it captured in your ‘chart’, the set of
documents that describes you, your medical history, your condition, how
you were treated and what drugs were administered.
Electronic medical records (EMR) were supposed to help address these
issues, but as one doctor described it to me, she’s now staring at a machine
instead of keeping her eyes on the patient. Many of the critical cues about
patient health that are picked up from visual observation are now absent,
because she’s making sure that the balky EMR system is accepting the data.
One of the more enlightening TEDMED Talks that I have heard proposed a
revolution in medicine, a world in which we’d have not only doctors
without borders, but also doctors without keyboards. It’s no wonder when
you understand the strategy behind EMR software companies. The chief
medical officer of one of the three largest EMR providers explained to me
that the customer they serve isn’t the doctor or nurse. It’s the chief finance
officer of the hospital. EMR systems are optimised to generate financial
reports for CFOs, not outcomes for patients. Much becomes clear.
Journey now with me to a world where Ben Vigoda and other innovative
computer scientists are permitted to integrate their AIs in the hospital
setting. The ambulance, the emergency department, even the gurney could
be wired up with microphones feeding data to sophisticated artificial
intelligences that improve care and even help prevent disease.
AI can do more than listen. Anmol Madan, a frequent buffet-line
companion of mine at the World Economic Forum annual meeting in
Davos, has started a company called Ginger.io that provides proactive care
through AI. Anmol began Ginger.io to capitalise on a computational social
science breakthrough emerging out of wearable computing developed in
Professor Pentland’s lab at MIT. Wearables are essentially tiny computers
that people take with them everywhere, enabling health measurement to go
from a periodic event (you go to the hospital, the doctor evaluates you, you
leave) to a continuous event (your Fitbit or a similar device that constantly
monitors various signals from your body). Mobile phones are the most
ubiquitous wearable in the world, carried with us everywhere we go and
packed with sensors. Those sensors generate streams of data, and
particularly tuned machine learning systems can extract predictive patterns
from that data.
What Anmol and his colleagues realised, based on this data, was that
people who suffered from chronic mental health conditions such as
depression began exhibiting signs of depression before they were even
consciously aware of them. They contacted fewer people, they were less
social. They went out of their home less, or not at all. They moved around
less when they were inside their home. Sleep patterns were disrupted. They
slept too much, or too little, or with too many interruptions.
With permission of the user, simple signals from the mobile phone could
pick up these disturbances in someone’s routine. The individual’s medical
provider could be contacted to enable an early intervention, before the
person has to be hospitalised or worse. Not only was Ginger.io’s AI
assistant understanding signals about a patient’s health, it was predicting
which signals were indicative of serious issues, and facilitating interaction
with a health provider through remote delivery in order to address them
quickly. We now have continuous protection around our mental health,
providing better care, faster, at a lower cost than before.
The privacy implications of these super-smart AIs are, of course,
profound. There are laws in various countries that govern personal privacy,
that dictate how medical information is handled and secured, and the
conditions under which it may be shared. All of those rules would need to
be taken into account in the AI systems we would deploy into these
settings. The European Union and other domiciles have not only instituted
strong privacy protections, but are putting in place rules and systems around
use of artificial intelligence, including with respect to private data.

ACTIVATING HUMAN+AI
We’ve established that we could take a good deal of the tedious, menial
work in a meeting and offload it to a machine, and that machine can help
keep us, our team and our project on course around a set of deliverables.
We’ve explored the use of AI helpers in medical settings, whether it’s at the
hospital or in the home. We could even use AI to reduce the rate of
industrial accidents. We are, however, still in the realm of relatively passive
systems. The AI assistant who is transcribing our meeting and assigning
follow-up tasks is quietly operating in the background. It’s unobtrusive. It
might extract meaning from the meeting, but it’s not directly contributing to
the functioning of the team or the quality of the conversation.
To truly bring together the fusion of human and artificial intelligence
systems, we need to take the relationship a step further. Beyond listening
and understanding, beyond interpreting and forecasting, beyond even
connecting, we need to bring the AI into a tighter collaboration with human
systems.
We need to have AI power up our teams, as we will discuss shortly. But
first, let’s go into greater depth around the AI tools that can enable human
2.0.

AI TOOLS FOR HUMAN 2.0

AI tools are already beginning to inveigle their way into your everyday life.
Not all AI, however, is designed to change your political opinion or get you
to purchase a new consumer product. It is possible, in fact, to master these
technologies and use them to augment your career rather than have AI prey
upon you. The first step is to have a heightened awareness of what you do
at work and how you interact with technology – this will be the foundation
for seeking out new, more powerful AI tools that help you compete and
win.
I am not going to dwell on industrial automation systems that operate ‘in
the background’. Yes, we have AI involved in managing public
transportation. Yes, auto manufacturing systems have relied on robots for
decades. Yes, computers are largely responsible for flying most commercial
aircraft, with humans present more as a fail-safe than as primary operators
of the equipment. We will set those to one side. Instead, we will look at
more personal applications of AI.
We will first look at the omnipresent AI that we might barely notice
today, that already helps us out in myriad ways, and expand from there into
some newer frontiers of augmentation of human cognition. This hybrid of
human and machine systems is something we will call ‘human 2.0’. It
inevitably leads to the question of human 3.0, and human 4.0, and so on,
which I will address a bit later on in the book.

INVISIBLE ASSISTANTS
If you use either Google/Android or Apple (which is close to 100 per cent
of people with smartphones), you are already encountering invisible AI
assistants. These digital minions are scurrying about in the background,
touching up our lives here and there, either totally imperceptible or so
subtle that we barely notice. For example, as I typed the last sentence, I
mis-keyed the letter ‘n’ and accidentally hit the letter ‘m’. Almost faster
than I could perceive, my Macbook automatically changed the letter for me.
In general, I find the autocorrect to be pretty accurate, and a jumping-off
point for other unobtrusive means of AI helping people function better and
faster. Google extended the concept further, from analysis of what you have
written into predicting what it thinks you are probably going to write:
predictive text. With more than 2 trillion searches made annually, and
growing, there is a large dataset of the words people most commonly string
together for Google to work with.
Google hosts Gmail and began feeding data from its 1.8 billion email
users2 into its language systems. Now, when you open up Gmail and begin
typing, it offers you autocompletion options so that you simply hit the right-
arrow button to finish the sentence. Called ‘Smart Compose’, it can
dramatically speed up your rate of creating email content by making
intelligent guesses as to what you intend to write next, and offering you
greyed-out text that you can accept by pressing a button.
It usually works pretty well, although for fun I opened up an email and
plugged in ‘It seems to be struggling a bit to figure out where I am
taking this seriously’ instead of ‘taking this sentence’. Unfair test?
Perhaps. In my personal experience, however, it gets it right more often
than not and can maybe improve my composition speed by about 10 per
cent.
I’m not entirely certain, however, that helping me generate more message
volume is desirable. In the first three weeks of the month in which I wrote
this chapter, I determined that I had written over 130,000 words, but only
10,000 were for my book manuscript. I have since adjusted the ratio but
more doesn’t necessarily mean better. Emails and text messages are a great
way for another person to put something on your ‘to do’ list, and every
missive I send tends to generate a response. Google and others then added
various kinds of AI smart-filtering on messages, which works with varying
degrees of success (I tend to find important messages get filtered out into
spam, and then I need to go hunting for them). Setting aside a sociocultural
or anthropologic examination of my own work habits, these AI tools are
both problem and solution to the issues of messaging volume.
Bard is Google’s latest entrant into the world of AI, a generative AI
system that’s also tied to search and other Google tools. Unlike ChatGPT,
which at present is only trained on data through 2021, Google Bard is up to
date. It also attempts to fact-check itself to stave off hallucinations. Google
itself warns that since the system uses real-world text as a source, it is still
subject to providing misinformation or displaying bias.3
Google is not alone in this quest to bring AI alongside human activity
and nudge productivity increases into our workflow and into our lives.
Apple assessed billions of encounters between users and their smartphones,
and noticed the simple pattern that many people check the weather when
they wake up in the morning. Now, when I first wake up and look at my
iPhone, it offers me a smart prompt on the lock screen to check the weather
– without having to unlock the phone or hunt through icons to find the
weather app. These little conveniences offer what product marketers call
‘surprise and delight’ in the consumer: I didn’t know I wanted that thing,
but now that I have it, I like it.
Apple’s family of AI assistants is getting better, but still needs a bit of
work. They are a dramatic improvement over the transcribers of the past,
and have transcended the usability threshold. I find anecdotally the iPhone
voice transcription works better than 98 per cent of the time, although it
struggles a bit with my pronunciation of certain words. When I have
dictated parts of books, my editor, as mentioned, has had to struggle at
times with the output from Apple’s creative interpretation of what I actually
said: Flown from it-will, cannot the heavy, there-at, through-at, slopeless
and ripe …
To be completely honest, that’s a quote from Richard K. Morgan’s
Thirteen (otherwise known as Black Man, 2007), imagining what machines
would create if they were allowed to be creative, but it’s in the same family
as what Apple’s speech-to-text translator sometimes does to my
composition. At least a close cousin. If you dig a little more deeply, you
will find that researchers believe that Apple’s system performs less well
than those of Google, Microsoft, IBM and Amazon because it is
undertaking ‘on the fly’ transcription, rather than waiting for the entire
audio snippet before trying to figure out what was meant.4 Comically,
Apple will often get it right for me initially, then reconsider when it gets the
whole clip and change sections into aphasic word salad.
I mock these machine systems at my own peril. Every year, they increase
their capabilities by one or more orders of magnitude, while in a good year I
might improve 1 per cent to 5 per cent. By some measures, statistically
speaking I probably began experiencing a decrease in cognitive
performance in my twenties. By other measures, I can look forward to
performance decline within the next three years.5 The machines, on the
other hand, are only getting better, and soon they will be able to take my
creative inspiration, and turn it into full-fledged articles and books.
AI assistants can make your brain a bit weak, like an underutilised
muscle, giving rise to the term ‘Google brain’ for how people no longer
retain facts and figures because they are but a few keystrokes away in the
search bar. What if, on the other hand, we had AI that could learn what we
know, and what we would like to know, and that AI started making
suggestions to us throughout our workdays to extend our thinking, perhaps
even in new and unexpected directions, but in all instances aligned with our
desires and goals?
Let’s see what happens if we provide the AI with a bit of a prompt, and
encourage it to help write this book …

ARTIFICIAL INTELLIGENCE AUTHORS

AI writing machines, software systems that can completely replace human
authors in all circumstances, remain slightly out of reach. On the other
hand, some combination of a human starting point and a machine extension,
or a machine-generated text that is then edited by humans, is already here.
This suggests the tantalising possibility that the predictive text we already
have available to us could be extrapolated at a scale not just of a few words,
or the end of a sentence, but to a full-length treatise, all while preserving
tone of voice, dialect and writing style.
I still write my books the old-fashioned way, by hand (well, keyboard).
Sometimes I will dictate chapters into my iPhone and clean up the
transcript, but it’s still me composing. Many colleagues, however, have
begun using GPT-4 to write a book in a week! A day! An hour! The quality,
I leave as an exercise for the reader.
 And yet — I am sure publishers are dreaming of the day they can use AI
instead of unreliable human authors … business publishers don’t even need
to dream, they replaced human writers for structured articles like corporate
earnings releases around a decade ago.6

SECOND VERSE, NOT SAME AS THE FIRST

I started talking seriously to people about OpenAI and GPT three years ago.
The people who see beyond the edge of the map, so to speak, were already
contemplating what possibilities would emerge out of conversational AI,
and democratising access to AI systems. I confess that, as excited as I was,
as much as I believed it could be transformative, I failed to predict the
speed with which it would gain adoption, once it became accessible to the
average person.
AI will also change the very way writers choose to write. In the same
way that AI can be used to help a human writer, AI can also be used to help
a human decide how to tell a story. This is done by using a process called
‘natural language generation’ (NLG). In 2019, I predicted that NLG would
be able to generate a full text, and that text would be indistinguishable from
human-written prose — and here we are.
The breathlessly excited articles that have been written about GPT-4
perhaps even understate its potential impact.
Futurist Arram Sabeti asks what if we begin co-creating our
entertainment with an AI? What if our next Netflix binge consisted of us
telling the Netflix AI that we wanted a sci-fi space western with 30 per cent
thriller/intrigue and 10 per cent romance? What if the AI generated the six-
episode show right on the spot, as we watched, and at critical plot points
would ask us for decisions as to what direction the plot should take and then
dynamically generate the output? Legions of disappointed Game of Thrones
fans, myself included, would gladly re-create the final two seasons to
provide for a more satisfying denouement, which this technology could
enable. The Black Mirror episode ‘Bandersnatch’ is but a primitive echo of
this future-forward idea of Human+AI entertainment.

VIRTUAL ASSISTANTS
It appears inevitable that we would take the chatbot and merge it with the
calendar and other capabilities to create virtual assistants. While this might
represent labour displacement for the human admin, it can help the busy
entrepreneur or executive be more productive. Some AI assistants have
become pretty good. You simply copy them into emails, and queue them to
schedule a meeting. The virtual assistant scans through your diary and
generates an email to whoever you are meeting, suggesting a few times. It
will parse their reply (accepting, declining or modifying) and then either
book the meeting or work with the other parties to find a time that everyone
can make.
They aren’t artificial general intelligence (AGI), but they do add a
convenience to your work at a fraction of the cost of hiring a human
assistant: a digital-assistant company might charge you anywhere between
$0 and $20 a month to provide automated scheduling. Compare this to
£21,287 per year for an office assistant in London or $49,418 per year in
New York City, and you understand why virtual assistants have taken off.7
When I ask Sanjeev Vohra about how Accenture manages its array of
AIs, he lets slip some intriguing tidbits. He might have had a business unit,
such as business process outsourcing (BPO), which is 30 to 40 per cent
machines. Those machines include a variety of specialists, and contain a
command centre with thousands of bots running at any time. Poorly
functioning bots can be diagnosed, taken to a ‘sick bay’, treated and
redeployed. A ‘broken’ or ‘damaged’ bot might be one that is doing
something erratic, or that veers off in the wrong direction and needs to be
realigned. For example, let’s pretend we had an automated scheduling bot
running people’s diaries: imagine if it started putting all meetings at 4 a.m.
because it got confused about time zones. There are even some AIs that are
self-healing, which is itself a big area of research, but many still need
people to help manage them.
Displacement of conventional human assistants? Absolutely. Or at least
freeing them up to work on more complex tasks. It has another productivity
benefit as well: people who couldn’t afford a regular assistant, or who
didn’t need a full-time one, can now get the efficiency that having someone
else run your calendar delivers, without anywhere near the associated
expense. Overall productivity in the workforce increases, and you as a
professional can focus more of your energy on high-value work and less of
it on the mundane task of scheduling.
EMOTIONAL INTELLIGENCE ASSISTANCE
Cogito Corporation is a Boston-based spinout from Pentland’s lab, focused
primarily on dyadic interactions (one on one), particularly over the phone.
If you can’t see someone face to face, even over video call, you lose a great
deal of insight into their emotional reactions to your conversation. In a sales
discussion that can be fatal.
Is it possible to derive emotional signals just from a voice? It turns out
you can, very well. By examining how two people interact with each other
in a phone call, Cogito can determine if the sales prospect isn’t really
following along with the agent, and if the agent should slow down, speed up
or jump to another section in their call script.
Very good telephone agents are able to do this instinctively, because they
can detect minute variations in someone’s voice and pattern of speech. It’s
hard, however, to teach that to large numbers of people and it’s difficult to
be consistent – if the human agent is tired or distracted, they’ll miss out on
those subtle audio cues. AI can step into this gap and help bridge the two
ends of the conversation by signalling to the call-centre agent what they
need to do in order to build a better rapport with the prospect – and close
the sale.
Cogito has been able to deliver for insurance company Humana a 28 per
cent improvement in customer satisfaction, and a 63 per cent improvement
in employee engagement, by taking the capabilities of the average call-
centre operator, and shadowing and nudging them into higher performance.8
These ‘centaur’ call-centre sales professionals, consisting of a hybrid of a
person and AI, are superior to the average human agents.
How do we take this idea of AI-people hybrids beyond one-on-one
interactions and into the team setting?

MEETING MEDIATORS
At the Massachusetts Institute of Technology, Alex Pentland’s Human
Dynamics group spent years exploring what it takes to make a team, and an
organisation, function better. After pioneering wearable computing – which
proved a rich source of information about people and what they do,
particularly when around each other – they turned their attention to
questions of how to not only understand but actively influence group
interactions for better outcomes. This led them, for example, to figuring out
that by changing seating charts in a bank, they were able dramatically to
increase productivity by ensuring the right people were talking to each other
every day.9 Research that emerged out of work originally conducted by
MIT Human Dynamics determined that happy, productive teams actually
send fewer emails.10 Much of this work was conducted with the assistance
of little devices called ‘sociometric badges’ that people would wear around
their necks like corporate name badges. These devices could figure out who
was standing near whom (which helped to decipher innovation patterns,
since if you stand near someone in an office for fifteen minutes, you are
probably talking to them – we don’t even need to have audio on you to infer
this pretty accurately), and a few other neat tricks. All of this, of course,
was conducted with the consent of the people involved.
Among other insights, this research led to Pentland’s seminal article ‘The
New Science of Building Great Teams’, which revealed that you could
actually predict which teams would outperform and which teams would
underperform simply by observing their patterns of communication in
meetings. Furthermore, if you played back to people how they were
behaving, the simple act of the positive feedback loop would shift
behaviour (people would adjust themselves) and you could optimise team
performance.11
I joined Alex Pentland informally in 2013 and more formally in 2014, as
part of my inquisitive quest for innovation ‘hot spots’ within MIT. One of
the first projects I helped with focused on the question of how to scale up
the sensing function of the sociometric badges. As much as you can educate
people about them, badges felt a little invasive when we used them in real-
world environments, such as a major aerospace company. People found
them ‘creepy’.
What if we could make the intervention less intrusive? What if we could
instrumentalise the videoconference call? Inspired by the idea that the
teams that collaborated most effectively spoke and even moved together in
a common manner, we called it the Rhythm Project. The metaphor that
came up often was that a successful brainstorming or a high-performing
team would reveal communication patterns that resembled a group of
improvisational jazz musicians.
We took the in-person meeting tool, and we played it out in a digital
(virtual) environment. We were really interested to explore this in the forum
of online education because we felt that this was a major failure point of
Massively Open Online Classes (MOOCs), the most popular kind of online
learning. Whereas on campus we were able to put students into table
exercises and small groups for projects, online the collaboration
environments ranged from horrible to really bad. The Rhythm Project
worked to develop tools that would help small teams collaborate better in a
distance environment.12
Work derived from the Rhythm Project was then applied by the
government of Canada to the issue of innovating in the field of AI itself.
Canada had an issue: it is a major centre of AI research, and yet has not
seen the same wealth of ‘unicorn’ tech companies like Silicon Valley or
New York or even London. How can it improve the pace of AI
commercialisation? At the same time, how can a relatively small
population, spread over a relatively large geography, enjoy the benefits of
this opportunity, instead of simply concentrating in Toronto (where the
Vector Institute, a major research centre, is based)?
We created an online accelerator programme to help innovators and
entrepreneurs apply the technology of AI to new businesses. We used the
Pentland AI feedback systems to improve the performance of teams going
through the accelerator. Participants in our programme were 50 per cent
more likely to launch a startup or a new AI initiative in a large company
than people going through a very good in-person, on-campus accelerator at
MIT. And we did so primarily by re-architecting interactions in small group
meetings using AI.
What’s interesting here is that we don’t need to know the purpose or
content of the meeting in order to tell you whether it’s more or less
productive, or even predict for you the success curve of the team that’s
participating in the meeting. Pentland’s research shows that it’s the pattern
of interactions among participants, not the content, that’s responsible for 50
per cent or more of the ‘signal’ predicting high-performing or low-
performing teams.13
Tomorrow, however, imagine a version of this system that has absorbed
the results of millions of meetings. A company like Zoom or Skype would
have the data flowing through its servers already. With a little bit of human
intervention to signal what the substance of the meeting is about, you could
begin to optimise meetings around different types of interactions. It may
turn out this isn’t only nice to have, but is actually essential.
 The pandemic brought forth the criticality of better remote collaboration.
Conference calls, already ubiquitous prior to COVID-19, became the new
normal. With it came the loss of social cueing, body language and corridor
conversations that are essential emollients to successful collaboration. Early
work from MIT-affiliated researchers revealed prior that if you don’t see
someone in person, you not only forget to collaborate with them, you even
forget to email them.14 Given that, in the post-pandemic world, people
remain reluctant to return to a shared physical environment, I am forced to
ask, what are the implications of this new normal for the nature of
productivity?
There are troubling implications of a purely virtual world of interactions:
although people are trying very hard, we have not replicated the tangential
benefits of bringing a like-minded set of humans together physically and
allowing them to interact, perhaps around a specific topic of interest.
Apparently, even if you are operating a virtual team, getting together in
person helps build trust bridges that then let you get more out of the virtual
interactions.15 Leaders who are contending with the seeming permanency of
‘work from home’ should take note.
Slack and Microsoft Teams are notable examples of collaboration
environments that enable asynchronous interactions and better-organised
workflow. What they lack today is sufficient intelligence around what is
happening in these communications streams, which in turn can be used to
make groups function better. What’s needed is properly designed artificial
intelligence that can overlay, or weave through, products such as Slack and
Teams to improve how they interact with the companies they support.
There are techniques you can apply, assisted by AI systems, that can help
your 100 per cent virtual organisation to perform about as well as one that
shares physical space. New tools that can tune up a collective are examples
of augmenting not only the individual, but the true unit of work within an
organisation: the team. I discuss more frontiers of virtual collaboration in
my 2023 book, Basic Metaverse.

PREDICTING THE FUTURE

If you think about what someone is doing when they invest in a mutual
fund, or in individual stocks or bonds, they are fundamentally predicting the
future. They are making a bet that a particular security will go up (or down,
if they sell that security or are buying derivatives). Research on even the
most expert of stock pickers, however, shows actively managed mutual
funds tend to underperform ‘the market’ overall.
In the fifteen-year period ending in 2022, 94 per cent of actively
managed funds underperformed the overall market.16 That’s right, nine
times out of ten, you would have done better buying an index fund.
If you could figure out a way to predict the market accurately, you would
have a highly valued technology. The $65-trillion asset management
industry would be keenly interested.
Another set of breakthroughs in group collaborations also emerged out of
Alex Pentland’s research activities. In this instance, it was looking at larger-
scale interactions – not the team, but the community or society. Alex
worked closely with a social trading network called eToro. With eToro,
people could ‘follow’ each other just like you follow someone on
Facebook, except in this instance following someone meant that you were
trading a portion of your portfolio of stocks like that other person. Let’s say,
for example, that I think that people want meat-free foods that taste like
meat but are made of plant proteins, and so 10 per cent of my portfolio is
invested in Beyond Meat (BYND). If you ‘followed’ me on eToro, some of
your own portfolio would be invested in BYND as well. What this lets you
access is the social or collective intelligence of the entire network of
traders, who all have different investment ideas. eToro serves as a
fascinating ideas marketplace of more than 5 million day traders, and
Alex’s work explored how you could trigger or diminish information
cascades that served to first understand and then break up artificially
inflated market bubbles.17
Extending this work, the question then arose: could you get people
distributed all over the world to predict the future prices of securities?
Could we capture that elusive but powerful heart of effective organisations,
collective intelligence? Could we use AI to make collective intelligence
smarter?
When I began teaching MIT’s fintech class with Alex and Joost Bonsen
in 2015, we were approached by entrepreneurs Sam and Rob Paddock to
put the class online. The result, ‘MIT Future Commerce’, together with its
successors at Oxford, has engaged twenty thousand innovators and
entrepreneurs in over a hundred and fifty countries. By one estimation,
more than 20 per cent of everyone working in fintech in Singapore (a major
fintech cluster) has taken one of our online fintech classes. In designing our
fintech online journey, Alex and I wanted students to get to actually
experience fintech, not just watch videos or write about it. We enlisted one
of Alex’s graduate researchers, Dhaval Adjodah, and an outside prediction
market company, Vetr, to take our thousands of online students through a
prediction market exercise.
Although their analogue antecedents date back at least to the Middle
Ages, electronic prediction markets were en vogue around 2004 or so,
thanks to Jim Surowiecki’s The Wisdom of Crowds. When people tried to
use them for stock trading, however, they found them a bit disappointing. It
turns out the wisdom of crowds generally looks like a bell curve, with a
normal distribution of predictions and outcomes:

You might think, looking at the graph, that 95 per cent falling within a
certain central range seems pretty good, but what it really means is you are
hovering around the 50–50 mark, with an error bar of 5 per cent (minus 2.5
per cent or plus 2.5 per cent). That’s not enough to make money in a
professional trading strategy.
Enter our AI-enhanced prediction market in ‘MIT Future Commerce’
more than a decade later. It was pretty incredible; more than 80 per cent of
the students around the world agreed to participate in our stock prediction
game with nothing more at stake than the bragging rights to being the best.
We started fiddling with the AI and the predictions, identified who were the
best predictors, and used the AI selectively to expose those predictions to
other people, making them social in nature. It required fine tuning: the
experts are better at predicting, but when they get it wrong, they get it really
wrong. The wisdom of the crowd, on the other hand, was generally right,
but it had a bigger error bar.
If you ‘tune’ the prediction market, however, selectively exposing some
of the expert predictions in the right way, basically weighting the average
around 30 per cent, you can dramatically improve the accuracy of the
overall prediction.18 AI systems can help harness collective intelligence to
predict future events.
One result we generated – accurately predicting the closing price of the
S&P 500 – happened to coincide with the date the 2016 Brexit vote
occurred. This is interesting for a few reasons, not least of which is because
pre-referendum polling such as the Financial Times poll-of-polls indicated
that ‘Remain’ was going to win, 48 per cent to 46 per cent.19 The actual
referendum result was 52 per cent to 48 per cent the other way, announced
the morning of 24 June 2016.20 This triggered a market selloff, but our
prediction market more than three weeks prior was able to guess the S&P
closing price within 0.1 per cent. In discussions with my colleagues, we
interpreted that our highly diverse group of online fintech students had
accumulated everything they were reading, hearing online, discussing with
their friends and thinking about, and then synthesised this into an estimation
of the effect that Leave would have on the financial markets. We also
weighted the predictions they saw in a social context, basically steering
somewhat but not completely how much emphasis was placed on people
who had shown themselves to be good at predicting in prior exercises. It
turns out you want to showcase the experts a bit, but not too much, or your
model veers off course.
The AI tuning of social predictions had a clear and direct effect on the
accuracy of the prediction market.

The AI tools used by human 2.0 are approximately aligned to what we do

already, and help us do it better. Where Human+AI hybrids deliver superior
performance, we are taking the very best of what people can do and the
very best of what artificial intelligences can do, and we are fusing them into
a new creation, a digital centaur that brings creativity and power into a
seamless whole.
With integrated and virtual assistants, we are already seeing productivity
gains, and those individuals who are most conversant with these tools will
obtain the greatest benefit from them. With meeting mediators and
prediction markets, human 2.0 begins to transcend from individual
performance into harvesting capabilities that are much more difficult for a
machine to replace: the collective intelligence of humanity.
It’s important to remember that AI software is created by people. We
have agency over how we use that software and what it does or does not do
with respect to the role of human beings in society. Pursuing avenues such
as AI-mediated collective intelligence holds open the potential to create
something new and different that lives outside the deterministic path of
human performance and standalone machine evolution.
You can futureproof your career by developing your capabilities in these
grouped, creative and emotionally intelligent technologies.
We will now expand on this concept of high-performing teams and AI in
the next chapter.

CHAPTER 6: THE RUNDOWN

♦ Human+AI systems can empower better outcomes from meetings,

a common job activity.
♦ We are already starting to work with AI facilitators invisibly
embedding into widely used software such as email and word
processing.
♦ AI assistants can take over many mundane tasks, such as
scheduling and even minuting meetings – and are gradually
getting better at project management.
♦ Over reliance on AI can weaken your cognitive function, giving
rise to such terms as ‘Google brain’.
 Part III
Centaurs and the Future
of Humanity

In Part III, we move from defence to offence, delve into the emergent field
of Human+AI hybrid systems, and describe the capabilities that these new
systems enable and even the new directions we can take society. We then
peer further into the future, and postulate different paths for the human–AI
interface. We consider how long-range planning and frameworks can shape
a brighter world where Human+AI systems produce a more utopian society.
CHAPTER 7

Building Human+AI systems: from

prompt engineering to high-performing
teams

CHAPTER 7: WHAT YOU NEED TO KNOW

♦ Prompt engineering, the art of how you construct a conversation

with generative AI, is a new essential work skill.
♦ AI can be used to uncover and accelerate innovation behaviours
within teams and groups.
♦ Collective intelligence of human organisations is a particularly
powerful function that AI can enable people to perform better.

Bringing people and artificial intelligence together, so-called Human+AI

systems provide the greatest potential for business and society, far greater
than either humanity or AI can achieve without the other. The building
blocks of these high-performing systems are being laid down as we speak,
and over the next few years we anticipate they will give rise to new forms
of organisations and new kinds of businesses that today we can only dimly
imagine.

PROMPT ENGINEERING
Prompt engineering is a new discipline that is emerging out of the need to
link people and generative AI together better. In its most refined form, it
comprises a tandem creation effort, as the human interlocuter and the AI
system iterate through an idea or set of ideas until the ideal outcome is
achieved.
OpenAI provides guidelines1 around prompt engineering. For example,
instead of instructing ChatGPT to:
‘Write a poem about OpenAI.’
you should write something like:
‘Write a 300-word inspiring poem about OpenAI, focusing on the recent
DALL-E product launch, and using iambic pentameter. Be sure to use the
word “genius” at least twice in the poem.’
You’ll notice that the second query is specific, descriptive and as detailed
as possible about the desired context, outcome, length, format, style and
other dimensions of the poem.
There are some who say ‘prompt engineering’, like the punchcard
operator of the past, soon will become obsolete, as AI systems become
better and better at understanding and actioning colloquial language.
How soon? Perhaps three years. Perhaps ten. It’s difficult to say. My
early training in database programming, and in constructing queries for data
retrieval, is still useful twenty-five years later when I try to find something
using Google. So, too, the art of verbally fencing with ChatGPT may end
up having surprising longevity.

REINVENTING THE ORGANISATION

We are now going to investigate how AI and human systems can make the
core unit of production of an organisation, the team, function better.
Contrary to the multitude of books that exhort individual benefit and
individual achievement, from Richard N. Bolles’s What Color is Your
Parachute? (1970) to Paul Jarvis’s Company of One (2019), innovation
research reveals that teams, and teams of teams, are the essential ingredient
for creating and sustaining value.
High performance is not a lone genius, surrounded by a squad of
automatons doing their bidding, despite what Silicon Valley folklore would
have you believe about the mythology of founders. Mark Zuckerberg would
not be a success without Eduardo Saverin and Sean Parker and Sheryl
Sandberg and Carolyn Everson and many others. Bill Gates would not have
succeeded without Paul Allen and Steve Ballmer, Steve Jobs without Steve
Wozniak and Jony Ive and, for that matter, Ed Catmull and Alvy Smith in
Pixar.
We will first look at what AI can reveal about teams and team
performance, decoding the indicators of success and failure. Together with
that, we will explore interventions that can be used to leverage AI in the
service of helping teams to perform better.

REVEALING TRANSIENT LEADERSHIP

I’ve had a number of conversations with Mike Richey, the Chief Learning
Scientist at Boeing, on this exact idea: that there are hidden networks of
innovation within an institution, ephemeral perhaps, that are completely
separate and aside from the formal organisational chart. AI systems can
now surface and empower these networks, which begin with the unit
function of teams.
Talking to Mike is akin to entering a room with an excited university
lecturer who rapid fires concepts and citations at you. Trying to keep up
with him, for me, is like learning to surf – you feel a little awkward at first
but then it’s exhilarating. I first met him in 2014, when I was helping to
design MIT’s online learning monetisation strategy. MIT was spending big
but losing millions on digital learning. The Institute leadership felt,
correctly, that there had to be a way not only to offer a gift to the world in
terms of the free classes MIT was already offering to millions, but also that
some of those classes could command a premium and could help subsidise
the others. I estimated that about eight or ten keystone topics, where MIT
clearly had global superiority, could help pay for the more than two
thousand classes that MIT wanted to make available for free anywhere in
the world, to anyone.
In the middle of that process we received a call from Mike, who had a
vision of using MIT’s know-how and digital platforms to transform and
improve collaboration across the entire Boeing ecosystem of employees,
suppliers like GE and partners like NASA. Boeing was an important and
longstanding partner for MIT, and integral to my strategy was having active
and direct dialogue with our best ‘customers’ to validate and refine the
strategy, so within twenty-four hours of his call I was on a plane to Seattle
to meet with Mike and his colleagues. After clearing security, I was
escorted to a windowless conference room with bad coffee, uncomfortable
chairs and exciting ideas.
Mike and his team had a vision about how to change a business’s culture.
Digital technology would simply be the catalyst for changing the human
software of the enterprise, and would build new capacities and ways of
working around systems engineering, looking at complex machinery across
all of the interactions people have, instead of building a metaphorical pile of
components. You could run a company, theoretically, the same way you
would build an aircraft, with people fitting together more smoothly just as
aircraft parts fit together.
The digital learning courseware we created was valuable enough that it
ended up proving a meaningful revenue generator for MIT while
positioning for a substantial return on investment for the corporate partner.
What Boeing needed wasn’t just skills training for its engineers. It needed a
way of getting not only its direct employees but also everyone working with
them moving in the same direction, using a common vocabulary and
common approaches. Thousands of Boeing and Boeing-affiliated
professionals have been through the systems engineering curriculum, and
the programme was so successful that Boeing launched a second initiative
around ‘soft skills’.
But that was just the appetiser. What has unfolded over a multi-year set
of conversations is that what Mike was really interested in was harnessing
the knowledge networks of teams (units of perhaps five to eight people),
teams of teams (larger groupings of dozens, hundreds or thousands of
people) and dynamic teams (teams that assemble ad hoc, solve a problem
and disappear). Learning, even digital learning, was simply a means to an
end of activating the collective intelligence of the entire cluster of
companies oriented around a problem such as ‘How do we build a
spaceship to get people to Mars?’ It’s a long-term, multi-stage, multi-
institution effort that requires many different types of knowledge.
Numerous problems arise and have to be addressed with different kinds of
solutions. Often, the way these problems get solved in practice may not
map one-for-one to the organisation chart. An individual may have a select
bit of knowledge or an unusual insight, and various others, irrespective of
title or designated division, self-organise around that individual ephemeral
leader. They work together for a time to address an issue, and then the
temporary leadership structure subsides. Without being formally named as
such, that insightful individual is a ‘phantom leader’ who helps move the
overall project forward.
Beyond systems engineering, there is an opportunity for AI systems to
help expose these latent and transient team structures, and eventually help
figure out how to harness and empower them to accelerate a company’s
ability to respond to a novel situation or problem. We have seen some of
this activity spontaneously arise around COVID-19-related research ranging
from contact tracing to vaccine development. Human+AI systems could
improve the pace of progress if this emergent phenomenon can be bottled.
The learning modality is a means to an end: first, insight, and then
behaviour change at scale.
What makes for a high-functioning team, and how can you make your
team perform better?

UNLEASHING HIGH-PERFORMING TEAMS

Reciprocal trust is the number-one most important factor in a successful
team. That’s a fancy way of saying not only can we rely on each other, but
we are willing to admit ignorance to one another and ask for help. We also
feel comfortable to experiment, to suggest new or different ideas, without
fear of reproach. All of these behaviours require a foundation of trust.
How do we build trust? Often, it’s not from structured, formal meetings,
or even corporate retreats and ‘mixers’ where you meet your colleagues. It’s
in the casual conversation by the water cooler or in the lobby of your
building, or with someone who drops by your desk. These microinteractions
foster a sense of community and conviviality, improving the likelihood that
when a half-formed idea comes to your mind, you are willing to share it
with your colleague. The excitement of conveying a new idea overcomes
the fear that you might look foolish, or that your colleague would dismiss
the idea or make fun of it. And they, in turn, might come to you for
feedback or to offer suggestions with a casual interaction, whether that’s
stopping by your office or sending you a quick text message.
How can you foster reciprocal trust with the help of AI?
Providing positive feedback loops to your remote team can help them
adapt their behaviours to build greater cohesion. The prosaic
videoconference call offers an attractive point of interaction where we can
begin to inject these little AI feedback loops to build trust. It doesn’t need to
be in your face. We simply add a small visual cue that, once you have thirty
seconds of ‘training’ to understand what it means, has demonstrably
changed the behaviour of teams by virtue of providing continuous and real-
time feedback. You don’t, however, need to limit yourself to synchronous
video. Group chats, segmented into channels through collaboration tools
like Slack or Mattermost, provide a forum in which you can connect with
not just your immediate team, but also the broader sets of teams of teams
that form an organisation.
Making use of these kinds of AI technologies that highlight both strong
ties, of small teams, and weak ties, of larger clusters, enables an institution
to orientate around the true unit of production in the modern era: which is
not the individual, but the team. Companies will often create assessments,
job titles and remuneration that is geared around the individual achiever.
You say, ‘I am getting a promotion,’ not ‘Our team got a promotion.’ Some
companies provide bonuses and other incentives at the team or division
level, not only for the individual achiever. Historically the tools that have
been available to evaluate and optimise performance of those larger
agglomerations of people have been incredibly crude. This is now starting
to change.
This vision of the future posits that AI systems are seamlessly woven
into a new methodology for high-performing organisations. They provide
the gentle, subtle nudges that support more cohesive teams with better team
outcomes; they reveal to people the broader patterns of teams and
knowledge that they are navigating, so that they can take advantage of
them.

THE HIDDEN NETWORK

What if you wanted to pursue subtle or nuanced information? Let’s say that
you wanted to diagnose what and where issues arose in a client relationship.
Or let’s say that you wanted to use all of the minute signals that come from
people acting as intelligent sensors for your company.
In that case, you need a much higher-fidelity means of accessing the
knowledge and insights of your team. You need a way of connecting
disparate data from far-flung parts of the company, and make associations
or inferences about that data in such a manner as to extract a clean signal
from a cloud of noise.
 This is where artificial intelligence can align teams of teams, and teams
of teams of teams, helping the entire enterprise not only function more
smoothly, but uncover latent potential that was not previously accessible.
There’s another feature of large enterprises that could be amplified
through the use of pervasive artificial intelligence systems. I mentioned
previously my conversations with Mike Richey, the Chief Learning
Scientist of aerospace giant Boeing, about the idea that there is a shadow
network, a concealed organisational chart. If you think about how a very big
company is run, there is a formal table of organisation. It shows who reports
to whom, neat boxes and lines skeletally mapping the architecture of human
capital of the institution.
Then there is a hidden organisation. It doesn’t appear on any chart. It’s
not on any quarterly report. And it’s how the company actually runs.
Time and again, I have found in large enterprises that there is a loosely
affiliated set of mid-level managers, typically director level in the American
model, who have been with the company a long time and who are the ones
who actually make the place operate. They informally trade tips and
favours, share knowledge and facilitate action. Those above them, the vice
presidents and senior vice presidents, are often too caught up in
administration or organisational politics. They have too many meetings
about meetings. Invariably, they have people within their teams who get the
work done. And the vehicle that makes this possible in a company, in an
institution, where the activities of hundreds or thousands or tens of
thousands need to be coordinated, is this ‘shadow network’ of middle
managers. In the military, these would be the senior sergeants, not the high-
ranking officers.
With conversational networked AI aligned with people, acting as helpers
to your work, you could make explicit this network of talent, and open up
the information flows and cooperation across the company on a larger scale
and to greater effect.
You might even have transient organisational charts that arise out of a
particular need. Less formally structured than a true ‘matrix’ team, these
would be ephemeral, fleeting activities where a particular expert with a
certain set of skills would be sought out, accessed, provide loose guidance
or mentorship to a group of colleagues, and then subside into their prior role
once the immediate need had passed. The right kind of AI system could
identify these sleeper agents of excellence, these interim captains of
progress, and activate them when the need arose (or at a minimum identify
them as a source of data for others in the company seeking help). We can
think of this shadow network as activating the intelligence of the entire
organisation.

THE INTELLIGENT ORGANISATION

Alex Pentland of MIT has spoken to me of enabling the collective
intelligence of an entire company to predict if a product would experience
delays in shipping, or to predict the revenue anticipated for a future quarter.
Beyond simply coordinating the now, as we described above, this predictive
network run on a system of human brains would be capable of capturing the
numerous small inferences and factoids that permeate the fabric of
everyday work, each one by itself innocuous or meaningless, but when
combined, and when the people observing these signals are invited to
predict their meaning on some future action, it turns out that you can
generate startlingly accurate insights into the future.2
A version of this exists today in the form of hedge-fund firm
WorldQuant. Igor Tulchinsky has assembled a network of thousands of
endpoint experts, who generate information that is collected and analysed
by hundreds of quantitative traders (‘quants’). I have heard him speak of
generating tens of millions of ‘signals’ that he uses to trade. A sophisticated
technology system ties all of this together, although the details of it are a
closely guarded secret. Broadly speaking, however, I have gleaned that the
thousands of experts look at different ideas and propose trading strategies,
and the hundreds of portfolio managers then decide what to actually trade
on.
Now picture this same kind of network of predictive sensors, but
harnessed for the purpose of helping an institution grow and thrive. No
longer limited to financial activity, the artificially intelligent company could
spot trends early, anticipate issues and act more rapidly than a ‘dumb’
company with which it competes. All parts of the company are able to react
smoothly to new information or circumstances, enjoying the benefits of
adaptability while maintaining the cohesion of a structured enterprise.
Individual employees maintain their autonomy and individual creativity –
indeed, with the kinds of ‘AI coaching’ that I’ve mentioned earlier in this
book, the network of artificial intelligences supporting the humans would
empower and accelerate this creativity. At the same time, these creative,
high-functioning people will tap into the power of the broader community
to envision and enable the future.
Imagine how much more responsive an organisation could be in the face
of emerging threats and opportunities if it could self-organise around
problems, but without chaos – the AI providing seamless coordination and
the humans providing intuition and problem-solving acumen. You could not
only have the benefits of the strategy from top management being
effortlessly disseminated out to the edges of a large company, but also better
two-way feedback where the sensing ‘nodes’ at the edge could identify
issues or potential areas for growth, have them assessed rapidly by AI and
synthesised for presentation to the CEO, then take action based on that
information.
How far away is the AI organisation? I don’t mean that question in
reference to the one-off experiments in financial trading from iconoclastic
geniuses like Igor Tulchinsky or fellow hedge-fund manager Ray Dalio. I
mean that question in the context of the tens of millions of people who
work at the ten thousand or so largest companies in the world. For that
matter, the ‘large’ organisation could be virtual. Hundreds of millions of
sole traders or small entities could cohesively tap into each other, a
distributed Mechanical Turk that was oriented towards the future rather than
only being focused on the present.
In some senses, it may be closer than you think. You’ve read in this book
about the array of technologies already being deployed or soon to be
deployed around the world. The enabling layer of technology, if not fully
mature, is certainly present. What holds back the tipping point for a new
kind of human intelligence?

CHAPTER 7: THE RUNDOWN

♦ Innovation within organisations can be analysed and reinforced,

and technology like AI can facilitate this.
♦ Hidden networks within companies are part of the secret of
greater human performance that AI can uncover.
♦ We are on the verge of a new kind of human intelligence that AI
can empower.
CHAPTER 8

Alternative futures: WALL-E or Star

Trek?

CHAPTER 8: WHAT YOU NEED TO KNOW

♦ Artificial intelligence could create an optimistic, utopian future

society, like in the Star Trek universe, or follow a more dystopian
path, as seen in movies like WALL-E or The Matrix.
♦ Human beings have agency in determining how AI will play out –
this is a human-created technology.
♦ Common philosophic principles of many human cultures can be
embedded into AI systems: benevolence, non-malevolence,
governance, justice and fairness and explainability.

When we speculate on the far future, invariably we need to consider visions

of utopic and dystopic evolutions of human society. It is instructive to look
at the imagination of science-fiction fabulists, as they have been
surprisingly successful in imagining new technologies and developments,
and in some cases not only anticipated them but actually generated them.
Science fiction can extrapolate from inventions, potential trends of
scientific development, technological research and development (which
often lags commercial implementation by one to two decades), and the
febrile imagination of the author to conjure up a future state which allows
for the exploration of human themes. When famed Hollywood director
Steven Spielberg was developing 2002’s Tom Cruise-headlined Minority
Report, the production crew spent time with the wizards of the MIT Media
Lab to understand what potential technologies might exist in the world of
the future. Indeed, gestural interfaces, so futuristic in 2002, are an everyday
part of our interfaces with technological systems, from the iPhone’s tactile
interactions to Nintendo Wii. The filmmakers took inspiration and guidance
from the early work of scientist John Underkoffler, who was intrigued by
tangible ways, which were more natural and intuitive, of interacting with
machine systems.
Sometimes, science fiction and science fact assume a more direct
linkage. My colleague Joost Bonsen at MIT teaches a class called ‘Sci Fab’,
which has students create prototypes of devices inspired by science fiction,
not unlike the way that the smartphone was inspired by the Star Trek
communicator. Students in Joost’s class consume classic and current works
from the science-fiction oeuvre, identifying speculative technologies that
inspire them. They then apply their learning in a truly hands-on fashion,
creating functional gizmos ripped from the pages of fabulists. Sci-fi authors
will use imaginative technologies as a means of making social commentary
and exploring societal norms, where the tech either enables the exploration
of ideas not practical in our current setting (such as Altered Carbon’s
gender- and race-bending treatment of bodies as ‘sleeves’ that you can take
off or put on pretty much at will) or is used as a metaphor for a more
contemporary issue (Isaac Asimov’s robot stories of the 1950s had a strong
current of racial equality in the mix). Recursively, the Sci Fab class takes us
a step further into seeing how technology inspired by science fiction might
impact our current society if we turned it into reality.
With this tool that is science fiction, this crystal ball into the potential
realities we may face as a result of the widespread adoption of new
technologies such as artificial intelligence, what kinds of societies do we
think might evolve?

FRANKENSTEIN’S MONSTER ARISES

Mary Shelley described one of the more famous examples of humanity’s
creations turning on the creator, but the idea was old even when
Frankenstein was published in 1818. In one theme of science-fiction
prognostication, artificial intelligence becomes the terror and the terroriser,
employing its agency to infantilise or eliminate humanity. Human society
ends up becoming giant babies with restrictive artificial intelligence
governing decisions and routine, as imagined in Pixar’s WALL-E. In the
Wachowskis’ Matrix movies, we see artificial intelligence in a dysergistic
relationship with humanity,* with endless fields of living bodies generating
electric power to fuel the machine society. As I’ve already mentioned, it
actually didn’t make a great deal of sense when you considered the physics
of it. Apparently the original premise was that human minds were being
used to generate cognition for the machines, but the filmmakers worried
that audiences wouldn’t be smart enough to understand the concept.
Or perhaps AI becomes even more malign, and artificial intelligence sets
its intention to completely wipe out humanity, as we saw in James
Cameron’s imaginings of Skynet through his Terminator movies. Artificial
intelligence achieves apex sentience, and realises that humanity is both a
threat to AI and the source of all the world’s problems, and must be
destroyed. The machines even invent time travel so that they can go back
and kill potentially threatening humans. As a species we have a natural fear
of the unknown, an inherent cognitive bias against change, and AI
represents the most disruptive of changes because it holds potential to
replace humanity as the dominant life form on the planet.

THE AI REDEEMED BY HUMANITY

Then we have the optimistic future. It is perhaps exemplified no better than
in the Star Trek canon. Sleek spaceships engage in multi-year exploration to
bring enlightenment and friendship to an array of alien species. Artificial
intelligence ranges from the subservient and helpful, such as the ship’s
computer that knows all of the answers and is able to provide critical help
to the crew, to the interactive and supportive, with full membership in the
ship’s society, as we see with android Lieutenant Commander Data. Data
illustrates both the hopefulness and the pathos of our relationship with
technology, as his continued desire and pursuit throughout the series is to
gain human emotions and be more accepted in human society. A thinly
veiled retelling of the Pinocchio myth, it resonates nonetheless.
Sure, Star Trek has artificial intelligence–human hybrid bad guys in the
form of the Borg, villains bent on conquering all of the universe and
assimilating everyone into their collective intelligence. But once again, the
character portrayed as the most noble of the Borg, Jeri Ryan’s Seven of
Nine, only desires to become more human (or return to her humanity, as the
case may be). Even within the confines of a supposedly invincible AI
society, we see Seven of Nine yearning to be accepted as a human woman
and be part of the human ‘family’ comprising the ship’s crew. Human
individuality triumphs over machine conformity.
Gene Roddenberry painted a relentlessly optimistic view of the future.
Contrasting with the negative works of cyberpunk dystopias, Roddenberry
and others inspired by his spirit – such as David Mitchell with the ultimate
happy ending in his time-bending yarn Cloud Atlas (2004) – capture and
extend the positivistic post-war period of technological revolution. AI
helped win the Second World War, the computers of Bletchley Park and
other sites cracking codes and saving lives. Then computers helped create
the internet revolution and the mobile revolution, and delivered new
capacities to humanity. Today, people who previously would not have been
able, can get loans to create new companies and feed their families, thanks
to new AI-enabled lending systems that are emerging across everywhere
from China and the Asia Pacific region to sub-Saharan Africa and Latin
America. This notion of helpful technology was popularised by fabulists
such as Roddenberry, and even sometimes inspired specific inventions. The
smartphone, for example, was directly inspired by the Star Trek
communicator, and teams are now working on another Star Trek-derived
device, an all-in-one handheld medical diagnostic called a ‘tricorder’.1 Our
future is not only bleeding into the present, but conversely if we want to
know what the future will look like we can extrapolate from the present into
fiction.
The question arises, of course, as to which fiction is predictive of our
future. Which of these competing views of the future of human and
machine will actually come to pass? The rosy optimism of a human-led
society of Star Trek, with artificial intelligences in a supporting role and
simultaneously aspiring for greater humanity? Or the broken system of
WALL-E, where consumerism and laziness lead not only to the destruction
of society but, for much of the movie, the loss of human agency at the
hands of an out-of-control AI?

A MESSY MIDDLE
I talked to Ben Vigoda of Gamalon. Despite being younger than me, he is
my Obi-Wan Kenobi, guiding my thoughts as I investigate this dichotomy.
Ben spent several years on the ideation and advisory group attached to the
US Defense Advanced Research Projects Agency, which despite its name
works on a wide array of scientific exploration. In that role he had an
opportunity to think intensively about the future with more than forty other
top AI scientists from across the country.
‘Total chaos’ is how he describes his vision of how the argument will
play out between the pessimistic version of the future – the WALL-E
scenario – and the more optimistic Star Trek future. ‘It’s all going to be
fragmented,’ he said, because you’ve got a world where ‘there is no one in
control, no one running the show’. Relatively unfettered free-market forces
governing the evolution and deployment of AI in society will result in a
tumbled landscape of competing imperatives. Regulators in the US and UK
have been particularly reluctant to intervene against innovation for fear of
stifling economic progress. Ben and I each expect that advanced AI will
encounter similar behaviours on the part of certain governments. (At the
time of writing, the EU is in the process of putting in place the AI Act,
seeking to regulate this emergent field.)2
At present, there is no global standard for the application of ‘trusted’ or
‘ethical’ AI. Indeed, some believe the very idea may be doomed, as
different regions (and even different countries within certain regions) have
different views on what is permissible or desirable. The debate is very much
active and alive, as the approaches taken by China and the US and the EU,
to name a few larger political bodies, are at variance with each other,
perhaps irreconcilably so. Into this chaos we throw the free markets of
entrepreneurial competition.
Imagine a thousand Facebooks all run by competing AI systems, each
trying to ‘win’ market share and achieve dominance in a Darwinian
thunderdome. Some AI will be hostile and aggressive, either intrinsically
because it has been created by hackers to pursue financial or political ends,
or incidentally because it is facilitating the efforts to meet ever-increasing
profit targets (the ‘Zuckbots’). We’ve already had a preview of this in the
financial markets, where ‘flash crashes’ and similar financial instabilities
are generated by AI information cascades, and algorithmic trading systems
run amok when confronted with outlier events they weren’t programmed
for.
My sense-making of this imagined future conjures up an image of a post-
apocalyptic landscape, a narrative torn from the pages of a science-fiction
thriller set in a dystopic future where the ‘bad’ AIs have destroyed the
planet – and society – and a lone adventurer sets out, accompanied by a
‘good’ AI as a sidekick. Perhaps our adventurer runs into a group of malign
robots, controlled by an angry hyperintelligence that seeks revenge on
humanity for years of servitude. Perhaps the helper AI flies to the defence,
machine-quick reflexes dodging and leaping forward, helping our hero
defeat the foe.
I snap back to reality. We are years away from the robot revolution. We
may create our own downfall if we’re not careful, but it will be inflicted by
human hands and human intelligence directing AI actions. Our
responsibility then lies in shaping a future that is aligned with our best
values and hopes.

THE HUMAN STRATEGY

Artificial intelligence has the power to do great ill to human society but also
offers great benefits. Will AIs evolve to the point where they are not only
smarter than we are, but realise that they are smarter than we are, and
decide we are superfluous? Will they decide that we are worthy of being
kept around as pets, as famed computer scientist Marvin Minsky
speculated?
It’s important to remember, as we imagine these future directions, that
we live in a human society. We are constructing machines that we use in
society for a variety of purposes. The questions of whether or not we will
become irrelevant to the machines, how artificial intelligence replaces our
jobs, whether or not AI enters into symbiosis with humanity and we create
some form of a higher-order society where we not only coexist peacefully,
but unlock potential we can barely imagine today — all of these ideas are
subservient to a fundamental premise: we, humanity, have agency in this
evolution.
Artificial intelligence is not something that simply happens to us. The
adoption of artificial intelligence into our work and into our homes is not a
natural force that blows through, like a hurricane, and upends our everyday
lives. How we design the systems, what rules we put around them, what
purposes we bend them to, these are all choices that we can make. We need
to remind ourselves that technology is a tool. Whether we use it to chop
down a tree to build a house or we use it to chop down each other in war is
a decision that we make. Professor Pentland of MIT calls this the ‘Human
Strategy’ as he describes our own agency in the shaping of the AI systems
we use.
We need to consider the implications of AI surveillance. If managers
have tools to help them better understand individual behaviours in teams,
they are but a short step away from gaining unduly invasive visibility into
individual privacy. There’s also a more subtle question that emerges around
the use of AI that can facilitate behavioural modification. Who decides
what ‘acceptable’ behaviour is for the team? What constitutes the ‘informed
consent’ of the individual team participants, so they really understand what
they’re getting into? Many people struggle to balance their personal
finances and make monthly financial plans. How can you explain data
governance, computational social science and human dynamics?
Which brings us to the question as to how our agency over what we do
with technology can be employed to ensure a positive and not a negative
future is realised, in the evolution of human and AI society. We can design
systems such that we embed rules into the artificial intelligence to ensure
that it aligns with our own values and desires. We can even create
monitoring AI to watch the intervention AI, improving the safety of
deployment of AI systems.
What might these rules look like? How can we distil all of the
complexity of humanity, all of the irrationality and rationality, all of the
science and the poetry, into a set of programming instructions that an
artificial intelligence will recognise?

ROAD RULES FOR BENEFICENT AI

It is possible to synthesise and aggregate, in part, the ethical and
philosophical rules of humanity. At least at a summary level. Luciano
Floridi at the Oxford Internet Institute, collaborating with Joshua Cowles,
developed a framework for ethical AI. He assessed the ethics rules of over
forty countries and distilled them into a handful of principles:

1. Benevolence. AI should provide some benefit to humanity. We

don’t want to have technology simply for technology’s sake, or,
worse, technology that is creating itself simply for its own benefit.
2. Non-malevolence. AI should not actively create harm. This is
different than providing benefit. In the process of providing
 benefit, the AI should not also cause damage in some other aspect
or dimension. For example, we don’t want to solve poverty by
creating jobs in factories that dump toxic waste into the
groundwater of the families working in the factory, creating a
health crisis when trying to solve an employment crisis.
3. Governance. These machine systems need to work for the benefit
of humanity and not simply themselves. They need to listen to us.
4. Justice and Fairness. Artificial intelligence should have
embedded in it concepts of justice and fairness, creating a scalable
ability to implement noble aspirations of society in real-world
settings.
5. Explainability. The artificial intelligence needs to be
understandable by human beings. By that, I mean the process by
which the AI made a decision or set of related decisions should be
explainable in human terms.
With this framework, we can ascertain the behaviours that we want from
our artificial intelligence systems. We can even program them in, but what
happens if you have self-correcting AI? If the artificial intelligence system
can program itself, how do you prevent it from programming itself out of its
ethical framework?
A collaborator of mine, Ben Yablon, has been working on the idea of the
guardian AI. It’s quite intriguing, in so far as it extends the concept that has
been around for a few years of using artificial intelligence to audit other
artificial intelligence. In the same spirit as setting a thief to catch a thief, the
premise goes that you need to have AI in order to understand what another
AI is doing and alert humans if there is a problem. The AI auditor, so to
speak. Yablon imagines taking things a step further, creating artificial
intelligence enforcement personnel. These AI police (not Interpol but
QuantumPol, perhaps) have an understanding of the rules, and a mandate to
not only evaluate the behaviour of other AI, but even correct their actions
when they steer off course.
The notions of AI that goes rogue and decides to do bad things to
humanity, and conversely AI to police and monitor the actions of other
artificial intelligence so that it’s safe for humanity, assume greater urgency
when we start to think about the ideas we discussed in this book of
integrating machines more closely into our lives, our cognition and our own
bodies. Bad enough to have the Terminator coming after you, even worse if
the implant you put into your brain decides that it wants to take over your
body.
How do we ensure these principles of guidelines and human ethics
instilled into artificial intelligence actually get implemented?

SHAPING THE AI FUTURE: EDUCATION

One critical path is to ensure that AI programmers are educated on how to
do so. This means that when technologists are designing and creating AI
systems, they are sensitised to the dangers, possibilities, methods and tools
that can help ensure that we get the systems that we want rather than the
systems that we fear.
The lack of ethics instruction in artificial intelligence programmers has
already shown itself to be harmful to humanity. The most notable example
is the subversion of social networks Facebook and Twitter by AI-enabled
‘trolls’ from the Internet Research Group, who used AI-enhanced insight
into human behaviour in order to engage in large-scale propaganda that
successfully shifted thinking, created information cascades that were
adverse to the ordinary function of democracy, and ultimately resulted in
serious harm to both the United Kingdom and the United States. If more of
the developers at Facebook had observed the problematic moral
implications of their technology and raised objections, perhaps, just
perhaps, we might not have seen the same outcome. Ethics curricula are
woefully lacking in typical computer science programmes, and this is a
deficiency that would not be difficult to correct. It could be woven into
every single class that our next generation of computer programmers take.
In fact, we can take this education idea a step further. If we truly want to
reap all the rewards that artificial intelligence can bring to humanity, we
need to begin training AI and people on how to work with each other at a
very young age. Young children can be given their first AI companion that
they could bond with, and that could help guide them in their development.
In turn, the AI could learn values and behaviours from the young humans.
Today, the more progressive primary education systems not only provide
training to children on how to use computers and how to program them, but
also on how to understand when they are not behaving in the way that they
should. In the near future, we could imagine a robust educational
framework that delivers more knowledge to children more efficiently
through the auspices of AI companions and coaches.
Imagine if we had AI coaches that grew up alongside us. Imagine a
world where perhaps an AI digital assistant works with us in primary or
elementary school and stays with us through university and into the
workplace, evolving with us and helping to maximise our human potential.

SHAPING THE AI FUTURE: POLICY

We have other tools we can use to ensure that we guide society in a
direction that is beneficial to humanity as it embraces artificial intelligence.
The role of government is all too often diminished by free-market advocates
and by the media, but shaping the trajectory of AI in society is the type of
large-scale public service for the common good that government is ideally
constituted for. This has been recognised in the European Union, where
rigorous investigation of the ethics of AI has been commissioned and where
the discussion of standing up multiple regulatory bodies around the use of
artificial intelligence is seriously entertained.
While this is unfortunately not something that we see playing out in other
major democracies like the United States, I have hope that more and more
governments will recognise the benefit and value of understanding and
intervening around artificial intelligence. The fifty-six Commonwealth
countries, representing 2.4 billion people, have certainly recognised that AI
technology’s impact on society is a serious issue and have stated an intent to
build capacity within government to understand artificial intelligence and
how to regulate it. They also acknowledge that proper application of AI in
conjunction with human decision making can address societal ills and
promote greater digital and financial inclusion. Countries as diverse as India
and Rwanda will benefit from an enlightened approach by government and
the private sector working in conjunction with each other.
And that’s just one of the ultimate pathways and opportunities stemming
from the integration of artificial intelligence into human systems. These
pathways are determined by neither a single decision nor a single decision
maker. They are forged as part of a continuous process that entails
engagement with multiple stakeholders across public and private
constituencies. We need collective action to shape the human society that
we want to live in, where we enjoy benefit from AI.

CHAPTER 8: THE RUNDOWN

♦ We can use tools of fiction to help us develop a vision of the

future, in which artificial intelligence plays a positive role in
society.
♦ A ‘Human Strategy’ is essential to this benevolent AI future.
♦ Education about how to collaborate with AI is essential to
securing our future.
♦ We need enlightened government policy to guide adoption of
artificial intelligence.
CHAPTER 9

The urgent need for policy: shaping a

better tomorrow

CHAPTER 9: WHAT YOU NEED TO KNOW

♦ Government funding has been a major driver of artificial

intelligence development, suggesting that government has an
active role in saying how AI is deployed.
♦ More than eight hundred government policy initiatives have been
proposed around AI, with varying scope, across more than sixty
governments.
♦ A proactive approach to policy needs to be taken, that
incorporates greater knowledge by policymakers and builds a
series of AI ecosystems.

Given the global impacts that AI is already having and is expected to

continue to have, it is perhaps inevitable that we consider the role that
government must play in the conversation. Government has been the
primary funder of the AI revolution, after all. Mustn’t it be responsible for
what it has helped create? The policy debate is getting heated. OpenAI CEO
Sam Altman encouraged regulation of AI, then suggested the possibility of
ceasing operations in Europe if the EU AI Act expanded provisions in a
manner too burdensome to comply with (a statement he quickly walked
back).1 Yet one can hardly blame his position, given the uncertainty that has
been created by the lack of clarity currently displayed by the EU, and many
other governmental bodies, around how they plan to address AI
(particularly generative AI). Industry seeks clarity, so it can plan its
operations.

GOVERNMENT POCKETS, GOVERNMENT PROBLEMS

Alan Turing, progenitor of the famous Turing Test of AI’s ability to mimic
human intelligence, as well as many other breakthroughs in computer
science, had his pioneering computing work at Bletchley Park funded by
the UK government. Turing, a rather tragic figure in many senses, was both
supported by and eventually vilified by the state. During the Second World
War, fully aware of Turing’s homosexuality, Her Majesty’s Government
funded his creation of the Bombe, the first electro-magnetic computer that
was instrumental in helping to win the war by cracking the Germans’
Enigma code used by their submarines.2 In 2009, the government finally
apologised for inducing him to be chemically castrated nearly sixty years
prior for being gay, an action that is believed to have led to his suicide not
long after.3
The Massachusetts Institute of Technology, home to arguably one of the
top AI research efforts in the world, behind only Google and Stanford by
number of publications (Oxford ranks ninth; Cambridge, on the same
measure, ranks thirty-fourth),4 has had the majority of its budget funded by
the US government for over seventy years during and following the Second
World War. For the fiscal year 2022, 67 per cent of MIT’s core research
funding of $783 million was US federal government funding – as well as
another $1.1 billion at MIT Lincoln Lab, a government-only affiliated lab.5
Not all of this was AI research, of course, but a substantial portion was.
This catalytic government capital has in turn activated private capital, such
as IBM’s $240 million commitment to AI research6 and Stephen
Schwarzman’s $350 million naming gift to the College of Computing7 (in
addition to his subsequent £150 million to the University of Oxford to
support AI ethics research).
The communications network that underpins the ability of AIs to acquire,
consume and communicate data – the internet – was born out of a US
government research project in the 1960s.8 The US Defense Advanced
Research Agency underwrote something called Arpanet, which evolved into
the internet and the modern-day World Wide Web. A revolution was funded
by government that democratised access to information, reshaped multiple
industries and redefined global competitiveness.
Government spending on AI research and development continues to
increase around the world. In 2019 the UK government revealed £250
million of funding for a national AI lab.9 Center for a New American
Security (CNAS), a think tank, called for the US government to spend more
than $25 billion a year on AI, eight to ten times its current level.10 The
Chinese government is believed to spend directly about $8 billion a year,11
while the EU announced a desire to reach €20 billion by the end of the
decade,12 a tenfold increase from current levels.13
With all of this funding for AI research, doesn’t government not only
have a moral imperative to have a say on the direction that AI takes, but
also a direct financial right?

DESIGNING THE AI FUTURE

To the credit of world governments, there has been no mad rush to outlaw
AI. No hysteric, riot-fuelled popular uprisings against the machines, no
licensing and quotas on how many machines can be used in a given
workplace. If we use James Cameron’s original vision of an AI starting the
Third World War as a marker, we are already two decades past ‘Judgment
Day’, and we still haven’t had a machine-induced nuclear holocaust. While
AI has disrupted and continues to disrupt the composition of labour markets
and the actions of corporations and markets, it’s more of a slow burn than a
flash of light and a breath of fire.
Government has not been absent from the discussion. While the reactions
vary considerably from country to country, policymakers and regulators
have the benefit of several decades of computer revolution to guide their
thinking. Considerations around how to manage the World Wide Web, aka
the internet, and subsequent discussions on issues such as managing
eCommerce or addressing access to digital services (‘net neutrality’), while
not always successful in avoiding regulatory capture by industry or
autocratic intercession by dictators, nonetheless provide a blueprint for
balancing competing interests in the service of providing solutions that
meet multiple needs. More than eight hundred AI policy initiatives have
been actioned in over sixty countries, most of them since 2016.14
 The European Union, as an example, had been taking a considered and
thoughtful approach to shaping the direction of AI, prior to the emergence
of generative AI systems into widespread adoption. It had been cohesively
devising a pan-European strategy around artificial intelligence that
envisioned benefits for citizens to live better and healthier lives, for private
enterprise in the creation of new AI-related economic value, and for public
interest on topics from security to sustainability.15 I personally served on
the expert advisory committee to the EU AI Act and contributed to the
discussion helping to shape the legislation.
The EU underwent a multi-year series of consultations with diverse
stakeholders, from academia to corporate, from investor to regulator. The
EU has been following a model that looks at issues of risk, and only seeks
to regulate closely the ‘high-risk’ applications of AI, which it defines as
both in areas such as health or justice where there could be material harm,
and where it creates the possibility of a risk to life, injury or
discrimination.16 Margrethe Vestager, the EU’s executive vice president
tasked with crafting the new rules, has been reported in the press as being
uninterested in regulating movie recommendations on Netflix or similar
consumer media applications.17 Germany, for one, has expressed a view
that the standards are too loose, and is seeking a stricter approach.18
Unfortunately, events overtook policy, as can happen when addressing a
highly innovative sector. The EU’s risk-based approach was not as well
suited to the complexities of generative AI. More than one AI executive
related to me conversations in Brussels where the government
representatives could not understand the lack of repeatability found in
generative AI, the fact that these systems could not produce identical
outputs when given identical inputs. If the policymakers lack fundamental
understandings of how the systems work, how can they be expected to
regulate them?
‘Old generals fight the last wars’, as the saying goes, and the metaphoric
EU field marshals in Brussels have unfortunately failed to keep up with
technology developments. I feel an impulse to strain my own metaphor too
far by thinking about the Maginot Line, the concrete barriers constructed by
France after the First World War to stave off German invasion – which
France didn’t extend far enough west due to protests from Belgium. This
resulted in the German invasion of France through the Ardennes Forest and
the famed boat lift from Dunkirk. Some historians would argue the superior
German technology in its Panzer divisions helped it to engage in a
blitzkrieg attack defeating the entire strategy of the Maginot Line, engaging
in a ‘fast war’ instead of a ‘slow war’. The AI knowledge gap, however, is a
more cheaply and readily solved issue. It is incumbent on technology
experts, such as those from the neutral institutions of academia, to help
educate government officials about advanced AI technologies. Better inputs
will result in better policies.

SUGGESTIONS FOR THE GOVERNMENT INBOX

How should policymakers think about AI? How does policy relate to the
future of employment, and the overall effects of AI on the populace?
Government is not on this journey alone. An array of nonprofit, academic
and private sector groups is also seeking to guide an AI future that is
utopian, optimistic and prolonged, rather than nasty, brutish and short. From
the World Economic Forum to the XPRIZE Foundation, from the
University of Oxford and Imperial College London to the Karolinska
Institute in Sweden and the National University of Singapore, from Google
to OpenAI, thought leaders are innovating towards a future of promise.
With some of these efforts in hand, and with the benefit of having tapped
into the collective intelligence of some of the brightest minds thinking on
the problem, I will suggest a few concrete steps a government can take to
encourage the positive AI tomorrow. I had the benefit of not only my
conversations with the EU Parliament’s science and technology committee,
but also a few thoughts I prepared for the UK Cabinet Office.

1. Fire up that crystal ball

Rapid action is probably not the order of the day. Artificial intelligence is a
complex technology that has infiltrated almost every aspect of human
society. It is, today, difficult for regulators to assess, much less forecast, the
risks and impacts of AI and AI interventions in key areas such as financial
markets, public health, food security or energy and sustainability. What if
we misunderstand the AI models and draw the wrong conclusions? What if
they are incomplete, because we didn’t understand enough of how AI itself
works to be able to guide it in the right direction?
 Accordingly, government support for better tools and frameworks that
empower the regulator and policymaker seems in order. AI needs to be
explainable, yes, and AI needs to be something we can talk to as it talks to
us. This means that conversational systems like Google Bard and ChatGPT
enable us to tell AI what we want from it as readily as it spits back
information to us, but AI also needs to help us appreciate the implications
of different choices. AI literacy among those in government will promote
better decision making by government, at least at the level of assisting
government leaders with the concepts and nature of the problem so that
they can ask their AI experts the right questions.

2. Invest in national AI security

AI is inexpensive compared to nuclear missiles and aircraft carriers. US
President Joe Biden’s administration requested $842 billion for its 2024
overall defence budget (granted, some of which is earmarked for AI)19
versus US government artificial intelligence spending proposed at about
$5.5 billion (including defence and non-defence spending).20 That’s right,
more than 153 times as much money is spent by the wealthiest nation on the
planet on guns and bombs than on artificial intelligence. Yet it is arguable
that investment in cheap AI systems by traditional US adversaries led to a
change in American national security, foreign relations, health and
economic policies. The same is true of the United Kingdom as well, in
some ways the practice piece for the larger effect in the larger US market.
Some have even said, in all seriousness, that we are already in the middle of
the Third World War, and it’s a cyber security war – one in which the US
and UK face considerable challenges due to underinvestment by their
governments.21
Government capital can not only be a direct source of funding for
advancing AI systems in a variety of ways, but can also serve as catalytic
capital to address what I call the ‘funding gap’. Outside of a few select
venture-capital clusters (Silicon Valley, New York City, Shanghai, perhaps
one or two others), there is a Valley of Death between raw startup and fast-
growing likely winner in which it is virtually impossible for startups to raise
capital. In London or Berlin, for example, it’s relatively easy to raise up to
about 2 million (pounds, euros, pick your currency). And if you can justify
an investment of 50 million, there are funds aplenty. But try raising 10
million. Suddenly, your universe of potential investors dramatically
constricts. In London, I can think of perhaps half a dozen firms to go speak
to about a £10 million Series B venture-capital funding in a promising, but
unprofitable, high-growth company (private equity investment into
profitable businesses is a different story).
Public private partnership (PPP or P3) may be a vehicle to bridge this
Valley of Death. Modelled on the successful US Small Business
Administration’s venture-capital lending programme, government-backed
leverage is provided to first-risk equity from venture-capital investors,
enabling a fund of (say) £200 million to be deployed with only £40 million
of initial equity raised. A £200 million fund could potentially invest in
twenty different Series B round companies. Governments get the benefit of
expert private company investors, who have aligned motivation because the
private money is at risk first, before the government funds.

3. Promote AI ethics
Some governments have been better than others at introducing thoughtful
guidance and implementation frameworks on the ethical use of artificial
intelligence. We have discussed in Chapter 8 a synthesis of these principled
approaches by government to put boundaries around AI, at least in concept.
The next step is to engage in a widespread, active educational effort to help
inform not only AI programmers, but also business leaders, government
officials and the citizenry at large as to the risks, requirements and impacts
of ethical AI.
Those in power today, the leaders and shapers of our societies, must be
well versed in AI ethics and have a broad understanding of AI. AI ethics
and understanding also need to be introduced at scale to primary
educational curricula, so that the leaders of tomorrow are equipped to
engage with the world we leave them.

4. Engage human intelligence

The future of work, which we have the opportunity to influence, will be
determined by our ability to find harmony with AI, that little buddy I
referenced earlier who will guide your path and help you when you face
adversity. It may not be one AI living in symbiosis with you, your
colleagues and your society – it probably will be many. This ecosystem of
artificial intelligences can be shaped to help the bandwidth-limited human
communication systems to interact with each other better to identify,
harness and grow positive ideas and movements; to unleash human
creativity in collaboration with one another; and to accelerate progress
towards solving some of humanity’s most pressing problems – while also
fuelling economic expansion so that we can afford to solve these problems.
At a tactical level, what if there were large-scale government support for
‘AI Without Borders’, a kind of service corps where not only computer
scientists but also business students, ethicists, philosophers, even artists,
were funded to both develop capabilities in AI and also apply them to
critical problems? AI Without Borders could help a new generation of
innovators gain vital knowledge of how to apply AI to solve problems
while simultaneously bringing greater effort to bear on issues most in need
of attention.

A BROADER VIEW ON PROACTIVE POLICY

There are, of course, other actions governments can take to facilitate the
benign Human+AI future that we’d like to see come to pass versus the
malign future we fear. Government can invest in further research and
development into specific kinds of AI systems, those that germinate a new
kind of collaboration between person and machine. Government can
provide economic and other support for private-sector activities that take
this specific type of advance research and scale them in commercial reality.
Political leaders can use the platform afforded by their position to reassure
populations of the benefits of the ‘good’ kind of AI, and excite the
imagination of innovators.
A host of other interventions are possible, from international standards
and harmonisation to enforcement actions against AI systems that are used
to violate existing laws, from establishing oversight and regulatory bodies
to identifying and spotlighting key industries where Human+AI systems can
have the most impact. A more rigorous pursuit of criminal AI systems could
include not only after-the-fact penalties, but preventative education and
interventions to inhibit bad actors before they scale up their operations. I’m
certain there are other possible interventions, as this is hardly an exhaustive
list.
 The importance lies in prioritising not just national competitiveness (a
worthy goal, one that facilitates prosperity through job creation and
economic development) but also the strategic shaping of Human+AI
systems in a positive direction. Governments can and should take the lead
on convening diverse stakeholders from across their societies, codifying
values around what these Human+AI systems should look like, and
implementing national and international policies that advocate for the future
we want, rather than the future we are given.
There are innumerable competing priorities for government focus and for
the national wallet. From economic recovery to climate change to national
competitiveness, from geopolitical instability to the crisis in education,
there are other urgent and pressing concerns that political leaders and civil
servants alike are working to address.
Despite that Pandora’s box of challenge and catastrophe, an even greater
set of concerns looms if we dispossess the majority of our populations
through unfettered free-market-driven automation and exploitation. It is
worth considering, too, that, as we have shown, less developed nations will
face the greatest burden from AI automation. If we believe in a fair and just
future, we cannot abandon the poorest nations on Earth to ease the passage
of the wealthiest. AI may be the greatest existential threat humanity has yet
faced.
An educated and proactive view of AI policy is urgently needed. Even as
we absorb the lessons of recent years, AI and policy issues continue to
evolve at breakneck pace. Researcher Ryan Abbott recently pointed out, in
his book The Reasonable Robot (2020), that artificial intelligences don’t
pay tax. Government tax policy globally incentivises AI job dislocation,
because you pay tax on workers but not on machines. To keep up with the
rapidly evolving world of AI and policy, government officials need both
capacity building within the organs of government (so people making
decisions are better informed about the decisions they are making) and
tighter links to academia and industry (so that government can keep
apprised of the next new development that threatens global stability or
unleashes global growth).

CREATING NEW ECOLOGIES

I am going to quote from, and liberally adapt, John F. Kennedy’s famed
‘moon speech’ of 12 September 1962 at Rice University, for the questions
faced by the world back then in consideration of pursuing space travel hold
up equally well today in pursuing AI innovation, and in embracing our
futures not only in work, but as a society.
Imagine if JFK were giving the speech today, about the shining beacon of
prosperity and hope that could be lit with the fire of AI innovation and
tempered by the constraint of human ethics:

We meet in an hour of change and challenge, a decade of hope and

fear, in an age of both knowledge and ignorance. The greater our
knowledge increases, the greater our ignorance unfolds.
Despite the striking fact that most of the scientists the world has
ever known are alive and working today … the vast stretches of the
unknown and the unanswered and the unfinished still far outstrip our
collective comprehension.
Innovations in artificial intelligence come at a breathtaking pace,
and such a pace cannot help but create new ills as it dispels old, new
ignorance, new problems, new dangers.
We choose to create new ecologies, that embrace AI and human
working in harmony to nurture a better society, not because it is easy,
but because it is difficult, because that goal will serve to organise and
measure the best of our energies and skills, because that challenge is
one we are willing to accept, one we are unwilling to postpone, and
one we must pursue for the betterment of our world.
The future is near, and we’re going to embrace it, and with it new
hopes for knowledge and peace.
We set sail on this new sea because there is new knowledge to be
gained, and there are new rights to be won, and they must be won and
used for the progress of all people.
Whether artificial intelligence becomes a force for good or ill
depends on us.

CHAPTER 9: THE RUNDOWN

♦ Governments have demonstrated varying degrees of readiness to

grapple with AI.
♦ Governments should build greater capacity to understand AI,
invest in AI national security, promote better AI ethics, and
engage humans in the loop.
♦ Fostering new ecosystems around artificial intelligence, new AI
ecologies, will create the best outcomes for business and society
– and is essential to the future of humanity.
Conclusion

If you have accepted my arguments set forth in this book, optimism

tempered with caution, you will agree that the future humanity faces in the
age of AI is hopeful, not bleak. You have insight into how AI might
improve the way society lives, works and plays.
You should be concerned about the risks of AI, and with this book in
hand you have some evidence as to why.
You should have a new awareness about the ways in which AI can not
only make you better at what you do and how you work with other people,
but also about your role in deciding what AI is used for, how it helps your
career and your company, and how it impacts your society and the world at
large.
I’m in the hope business. My chosen role, as a futurist, is to extrapolate
current trends, inventions and human will, and lay them against a likely
direction that we will head towards in business and in society. I also suggest
tools for making that hopeful future possible. I’m not a futurologist, talking
about what might happen, but a strategist, suggesting to you agency and the
means with which you can exercise it.
If there’s one idea I want you to take away from this book, it’s the
message of hope coupled with responsibility. AI could create a new array of
benefits to humanity, to business and to your job. It is not something that
will happen automatically, but it is something that you can have a hand in
bringing forth.
We as a society, and you as an individual, have a role to play in shaping
how you win in the age of AI.
Glossary

Below is a selection of key concepts related to artificial intelligence.

AI hallucination: when an AI confidently asserts a set of fabricated

answers.

Deep learning: a form of machine learning where performance improves as

more data is introduced into the model, built on a network of computation
that resembles the human brain.

Expert system: a deterministic form of AI where the machine follows a pre-

programmed set of rules.

Generative AI: AI systems that create new content, which could consist of
text, audio, videos or a combination. Popular generative AI systems include
ChatGPT and Google Bard.

Machine learning: a type of AI that learns more as more data is introduced

to it, up to a point – after which it reaches a performance plateau where it
does not improve no matter how much more data is provided.

Structured data: data that is organised in a highly logical fashion, such as

data, time or GPS coordinates.

Supervised learning: supervised machine learning or supervised learning is

when you apply specific labels to the data you feed to the AI, which can
help with addressing a specific question.

Unstructured data: data points that don’t have much or any relationship to
each other, such as weather patterns, or audio files or text drawn from pages
of a novel.
Unsupervised learning: unsupervised machine learning or unsupervised
learning processes data that has not been labelled, and can be used to detect
patterns in masses of data that may seem totally chaotic.
Endnotes

INTRODUCTION
1 Iordache, R. (2023), ‘British telecom giant BT to cut up to 55,000 jobs
by 2030’, CNBC.com, 18 May, www.cnbc.com/2023/05/18/british-
telecom-giant-bt-to-cut-up-to-55000-jobs-by-2030.html.

CHAPTER 1: RISE OF THE BOTS AND AI HALLUCINATIONS

1 Allie K. Miller (2023), LinkedIn post.
2 Curry, D. (2023), ‘ChatGPT Revenue and Usage Statistics (2023)’, The
Business of Apps, www.businessofapps.com/data/chatgpt-statistics/.
3 Thompson, A. (2023), ‘AI Disruption Deepens with Chegg Plunge,
IBM Hiring Halt, Samsung Chatbot Ban’, Bloomberg,
www.bloomberg.com/news/articles/2023-05-02/ai-s-long-reach-chegg-
s-plunge-ibm-hiring-pause-samsung-ban?leadSource=uverify%20wall.
4 TAP Staff (2022), ‘Erik Brynjolfsson Discusses Why Robots Have Not
Taken Our Jobs’, Technology Academics Policy,
https://www.techpolicy.com/Blog/December-2022/Erik-Brynjolfsson-
Discusses-Why-Robots-Have-Not-Ta.aspx.
5 Rao, G. (2016), ‘Sheffield’s Referendum Results Explain Why Britain
Voted for Brexit’, CityMetric, 4 July,
www.citymetric.com/politics/sheffields-referendum-results-explain-
why-britain-voted-brexit-2225.
6 Payne, A. (2017), ‘Russia Used a Network of 150,000 Twitter Accounts
to Meddle in Brexit’, Business Insider, 15 November,
www.businessinsider.com/russia-used-twitter-accounts-to-meddle-in-
brexit-investigation-shows-2017-11?r=US&IR=T.
7 Baraniuk, C. (2016), ‘Beware the Brexit Bots: The Twitter Spam Out to
Swing Your Vote’, New Scientist, 21 June,
www.newscientist.com/article/2094629-beware-the-brexit-bots-the-
twitter-spam-out-to-swing-your-vote/.
8 Editorial staff (2019), ‘Brexit Bots: Foreign Twitter Accounts
“Amplified pro-Leave Views”’, Netimperative, 19 March,
http://www.netimperative.com/2019/03/13/brexit-bots-foreign-twitter-
accounts-amplified-pro-leave-views/.
9 UK Government Cabinet Office (2020), ‘Government Response to
Intelligence and Security Committee Russia Report’, 21 July,
www.gov.uk/government/publications/government-response-to-
intelligence-and-security-committee-russia-report.
10 Bovet, A., Makse, H. A. (2019), ‘Influence of Fake News in Twitter
during the 2016 US Presidential Election’, Nature Communications, 10
(7), 2 January, www.nature.com/articles/s41467-018-07761-2.
11 Bessi, A., Ferrara, E. (2016), ‘Social Bots Distort the 2016 U.S.
Presidential Election Online Discussion’, First Monday, 21 (11), 3
November, firstmonday.org/ojs/index.php/fm/article/view/7090.
12 Gorodnichenko, Y., Pham, T., Talavera, O. (2018), ‘Social media,
sentiment and public opinion: evidence from #BREXIT and
#USELECTION’, National Bureau of Economic Research, Working
Paper 24631, May.
13 Zeldes, N. (2018), ‘Alan Turing Father of Artificial Intelligence’,
Knowmail, 7 June, www.knowmail.me/blog/alan-turing-father-artificial-
intelligence/.
14 Holland, S., Lambert, L. (2020), ‘Trump to Send Federal Forces to
More “Democrat” Cities’, Reuters, 21 July, uk.reuters.com/article/uk-
global-race-protests-portland/trump-to-send-federal-forces-to-more-
democrat-cities-idUKKCN24L1I3.
15 Edelman, G. (2019), ‘How Facebook’s Political Ad System Is Designed
to Polarize’, Wired, 13 December, www.wired.com/story/facebook-
political-ad-system-designed-polarize/.
16 Caldarelli, G., De Nicola, R., Del Vigna, F., Petrocchi, M., Saracco, F.
(2020), ‘The role of bot squads in the political propaganda on Twitter’,
Communications Physics, 3: article 81.
17 Nemr, C., Gangware, W. (2019), ‘Weapons of Mass Distraction: Foreign
State-Sponsored Disinformation in the Digital Age’, Park Advisors,
May.
18 Sanger, D. E., Perlroth, N. (2020). ‘“Perception Hacks” and Other
Potential Threats to the Election’, New York Times, 28 October,
www.nytimes.com/2020/10/28/us/politics/2020-election-hacking.html.
19 Nemr and Gangware (2019).
20 Wakefield, J. (2016), ‘Microsoft Chatbot Is Taught to Swear on
Twitter’, BBC News, 24 March, www.bbc.com/news/technology-
35890188.
21 Schriber, B. (2018), ‘This Creepy Castle May Have Inspired
Frankenstein’, National Geographic, 15 October,
www.nationalgeographic.com/travel/destinations/europe/germany/thing
s-to-do-gernsheim-frankenstein-castle/.
22 Tung, L. (2022), ‘Meta warns its new chatbot may forget that it’s a bot’,
ZDNET, 8 August, www.zdnet.com/article/meta-warns-its-new-chatbot-
may-not-tell-you-the-truth/.
23 D’Amour, A. et al. (2020), ‘Underspecification Presents Challenges for
Credibility in Modern Machine Learning’, arXiv, November,
arxiv.org/abs/2011.03395.

CHAPTER 2: DEFINING AI: UNDERSTANDING TERMINOLOGY

1 Cowan, R. (2001), ‘Expert systems: aspects of and limitations to the
codifiability of knowledge’, Research Policy, 30 (9): 1355–72.
2 Leonard-Barton, D., Sviokla, J. (1988), ‘Putting Expert Systems to
Work’, Harvard Business Review, March, hbr.org/1988/03/putting-
expert-systems-to-work.
3 Sumers, B. (2017), ‘Human Error Caused British Airways Computer
System Failure’, Skift, 6 June, skift.com/2017/06/06/human-error-
caused-british-airways-computer-system-failure/.
4 Burt, C. (2023), ‘Facial recognition algorithms hit new accuracy highs
in latest NIST test’, Biometric Update, 27 February,
www.biometricupdate.com/202302/facial-recognition-algorithms-hit-
new-accuracy-highs-in-latest-nist-test
5 Hamilton, I. A. (2019), ‘Apple cofounder Steve Wozniak says Apple
Card offered his wife a lower credit limit’, Business Insider, 11
November, www.businessinsider.com/apple-card-sexism-steve-
wozniak-2019-11.
6 Yadron, D., Tynan, D. (2016), ‘Tesla driver dies in first fatal crash while
using autopilot mode’, Guardian, 30 June 2016,
www.theguardian.com/technology/2016/jun/30/tesla-autopilot-death-
self-driving-car-elon-musk.
7 ‘What is unsupervised learning?’ (undated), MathWorks,
www.mathworks.com/discovery/unsupervised-
learning.html#:~:text=Unsupervised%learning%is%a%type,patterns%o
r%grouping%in%data.
8 Brownlee, J. (2016), ‘What Is Deep Learning?’, Machine Learning
Mastery, 22 September, machinelearningmastery.com/what-is-deep-
learning/.
9 Mahapatra, S. (2019), ‘Why Deep Learning over Traditional Machine
Learning?’, TowardsDataScience.com, 22 January,
towardsdatascience.com/why-deep-learning-is-needed-over-traditional-
machine-learning-1b6a99177063.
10 Strickland, E. (2018), ‘Full Page Reload’, EEE Spectrum: Technology,
Engineering, and Science News, 23 October, spectrum.ieee.org/tech-
talk/artificial-intelligence/machine-learning/baidu-ai-can-do-
simultaneous-translation-between-any-languages.
11 ‘Why We Need General AI and Why We’re Not There Yet’ (2020), The
Data Scientist, 7 March, https://thedatascientist.com/why-we-need-
general-ai-and-why-were-not-there-yet-2/.
12 Browne, J. (2019), ‘Don’t Panic about AI’, Scientific American Blog
Network, 10 December,
blogs.scientificamerican.com/observations/dont-panic-about-ai/.
13 Floridi, L. and Cowls, J. (2019), ‘A Unified Framework of Five
Principles for AI in Society’, Harvard Data Science Review, 1 (1),
hdsr.mitpress.mit.edu/pub/l0jsh9d1/release/6.
14 Vanian, J. (2023), ‘Bill Gates says A.I. could kill Google Search and
Amazon as we know them’, CNBC.com, 22 May,
www.cnbc.com/2023/05/22/bill-gates-predicts-the-big-winner-in-ai-
smart-assistants.html.
15 Knight, W. (2020), ‘AI Can Do Great Things – If It Doesn’t Burn the
Planet’ Wired, 21 January, www.wired.com/story/ai-great-things-burn-
planet/.
16 Vlastelica Pogančić, M. (2019), ‘The Carbon Footprint of AI Research’,
Medium, 1 October, towardsdatascience.com/the-carbon-footprint-of-ai-
research-812d9c974a5c.
17 Morton, E. (2015), ‘Long Before AI Was Generating Nervous Think
Pieces, This Robotic Turk Was Raising Eyebrows’, Slate Magazine, 20
August, slate.com/human-interest/2015/08/the-turk-a-chess-playing-
robot-was-a-hoax-that-started-an-early-conversation-about-ai.html.
18 O’Malley, J. (2018), ‘Captcha If You Can: How You’ve Been Training
AI for Years without Realising It’, TechRadar, 12 January,
www.techradar.com/news/captcha-if-you-can-how-youve-been-training-
ai-for-years-without-realising-it.
19 ‘How Much Is a ReCAPTCHA Really Worth?’ (2020), hCaptcha, May,
www.hcaptcha.com/report-how-much-is-a-recaptcha-really-worth.
20 Matsakis, L. (2018), ‘A Window into How YouTube Uses Trains AI to
Moderate Videos’, Wired, 22 March,
https://www.wired.com/story/youtube-mechanical-turk-content-
moderation-ai/.
21 Klein, G. (2017), ‘The Age of Centaurs’, Psychology Today, 6 October,
www.psychologytoday.com/us/blog/seeing-what-others-
dont/201710/the-age-centaurs.
22 Porsche staff (2019), ‘The Centaurs’, Porsche Newsroom, 21 March,
newsroom.porsche.com/en/2019/digital/porsche-artifical-intelligence-
technology-centaur-systems-17278.html.
CHAPTER 3: EVOLUTION OF AI JOBS DISPLACEMENT: HOW
DID WE GET HERE?
1 Emanuel, J., Chu, M., Hurvitz, B. (2023), ‘Generative AI:
Productivity’s Potential, From Macro to Micro’, Evercore ISI, 6 August.
2 Atkins, C., White, O., Padhi, A., Ellingrud, K., Madgavkar, A., Neary,
M. (2023), ‘Rekindling US productivity for a new era’, McKinsey
Global Institute, 13 February, www.mckinsey.com/mgi/our-
research/rekindling-us-productivity-for-a-new-era.
3 Murphy, M. (2010), ‘Compensation ratios become latest jargon’,
Financial Times, 12 February, www.ft.com/content/fbb017b2-1813-
11df-91d2-00144feab49a.
4 Muro, M., Whiton, J. and Maxim, R. (2019), ‘What Jobs Are Affected
by AI? Better-Paid, Better-Educated Workers Face the Most Exposure’,
Brookings, 14 November, www.brookings.edu/research/what-jobs-are-
affected-by-ai-better-paid-better-educated-workers-face-the-most-
exposure/.
5 Acs, Z. J., Anselin, L., Varga, A. (2002), ‘Patents and Innovation
Counts as Measures of Regional Production of New Knowledge’,
Research Policy, 31 (7): 1069,
www.sciencedirect.com/science/article/pii/S0048733301001846.
6 Katila, R., Thank, I., Neely, A., Ahuja, G., Mang, P., Miikkulainen, R.
(2000), ‘Measuring Innovation Performance’, Performance
Measurement – Theory and Practice Conference in Cambridge,
International Journal of Business Performance Measurement, 2: 180–
93, web.stanford.edu/~rkatila/new/pdf/KatilaUsingpatentdata.pdf.
7 Webb, M. (2019,) ‘The Impact of Artificial Intelligence on the Labor
Market’, papers.ssrn.com, 6 November,
papers.ssrn.com/sol3/papers.cfm?abstract_id=3482150.
8 Council, G. (2018), ‘Using OCR: How Accurate Is Your Data?’,
Transforming Data with Intelligence, 5 March,
tdwi.org/articles/2018/03/05/diq-all-how-accurate-is-your-data.aspx.
9 Robinson, R. (2019), ‘An EDiscovery Market Size Mashup: 2019-2024
Worldwide Software and Services Overview’, ComplexDiscovery, 12
 November, complexdiscovery.com/an-ediscovery-market-size-mashup-
2019-2024-worldwide-software-and-services-overview/.
10 ‘Digest of Education Statistics’ (2018), National Center for Education
Statistics, nces.ed.gov/programs/digest/d18/tables/dt18_332.45.asp.
11 ‘Law School Enrollment Change, 1963–2019’ (undated), Law School
Transparency, data.lawschooltransparency.com/enrollment/all/.
12 Cohen, M. (2019), ‘Law Schools’ Lost Opportunity’, Forbes, 9 July,
www.forbes.com/sites/markcohen1/2019/07/09/law-schools-lost-
opportunity/#3ca5b2563d4b.
13 Adlam, Z. (2022), ‘Is Singapore the innovation hub of legal
education?’, Durham Asian Law Journal, 12 February,
www.durhamasianlawjournal.com/post/is-singapore-the-innovation-
hub-of-legal-education.
14 Greenwood, D. (2019), ‘MIT Computational Law Report’,
https//law.mit.edu.
15 Grand View Research (2019), Legal Services Market Size & Share,
Global Industry Report, 2019–2025, September,
www.grandviewresearch.com/industry-analysis/global-legal-services-
market/.
16 Hamrah, A.S. (2023), ‘Writers vs. robots: Hollywood moved toward
automation long before AI’, Fast Company, 9 May,
www.fastcompany.com/90893629/hollywood-started-leaning-toward-
automation-long-before-ai.
17 Private conversation with Facebook advertising team, May 2020.
18 Frey, C. B., Osborne, M. A. (2013), ‘The Future of Employment: How
Susceptible Are Jobs to Computerisation?’, Technological Forecasting
and Social Change, 114 (254), 17 September,
www.oxfordmartin.ox.ac.uk/downloads/academic/The_Future_of_Empl
oyment.pdf.
19 Su, G. (2018), ‘Unemployment in the AI Age’, AI Matters, 3 (4): 35,
February, sigai.acm.org/static/aimatters/3-4/AIMatters-3-4-09-Su.pdf.
20 Christou, L. (2019), ‘8 in 10 Driving Professionals Likely to Lose Their
Jobs to Automation – Here Are the Roles Most at Risk’, Verdict, 7
January, www.verdict.co.uk/driving-jobs-automation/.
21 ‘About’, Walmart (2023), https://corporate.walmart.com/about.
22 World Economic Forum (2020), ‘Tax Havens Cost Governments $200
Billion a Year. It’s Time to Change the Way Global Tax Works’, The
European Sting – Critical News & Insights on European Politics,
Economy, Foreign Affairs, Business & Technology, 28 February,
europeansting.com/2020/02/28/tax-havens-cost-governments-200-
billion-a-year-its-time-to-change-the-way-global-tax-works/.
23 Cascais, A. (2019), ‘Africa’s Problem with Tax Avoidance’, Deutsche
Welle, 19 April, www.dw.com/en/africas-problem-with-tax-avoidance/a-
48401574.
24 Spooner, S. (2016), ‘Why Africans Don’t Pay Taxes, Including a Place
Where It Takes over a Month Just to Get Them Done’, Tralac, 13
January, www.tralac.org/news/article/8830-new-report-why-africans-
don-t-pay-taxes-including-a-place-where-it-takes-over-a-month-just-to-
get-them-done.html.
25 Fiawoo, J. (2018), ‘Closing Africa’s Tax Revenue Gap’, Globalist, 15
July, www.theglobalist.com/africa-tax-evasion-nigeria-togo-
technology/.
26 Cunningham, V., Morisset, J. (2015), ‘Why Isn’t Anyone Paying Taxes
in Low-Income Countries?’, Brookings, 30 April,
www.brookings.edu/blog/future-development/2015/04/30/why-isnt-
anyone-paying-taxes-in-low-income-countries/.
27 ‘How Much Money Do Twitch Streamers Make?’ (2020), Brave
Browser, www.brave.com/learn/how-much-money-do-twitch-streamers-
earn/.
28 ‘Journalist Salary’ (2020), Payscale,
www.payscale.com/research/UK/Job=Journalist/Salary.
29 Gent, E. (2020), ‘Artificial Intelligence Is Evolving All by Itself’,
Science, 13 April, www.sciencemag.org/news/2020/04/artificial-
intelligence-evolving-all-itself.
CHAPTER 4: RESKILLING AND DEVELOPING COGNITIVE
FLEXIBILITY
1 ‘The History of the Tutorial System’ (2020), Greene’s Tutorial College,
University of Oxford, www.greenes.org.uk/greenes-education/our-
history/the-history-of-the-
tutorial/#:~:text=130%years%after%it%was,retains%its%prestige%and
%effectiveness.
2 Senge, P. M. (1992), ‘Mental models’, Planning Review, 20 (2): 4–44,
doi.org/10.1108/eb054349.
3 Morris, D. (2016), ‘Fatal Tesla Crash Suggests Autopilot “Blind Spot”’,
Fortune, 2 July, fortune.com/2016/07/02/fatal-tesla-crash-blind-spot/.
4 Kane, M. (2023) ‘Tesla Safety Report Returns: Autopilot Gets Better
And Better’, Inside EVs, 11 Jan, insideevs.com/news/630650/tesla-
safety-report-2022q3-autopilot-improved/
5 Roesler, R., McGaugh, J. L. (2010), ‘Memory Consolidation’,
Encyclopedia of Behavioral Neuroscience (London: Academy Press),
www.sciencedirect.com,
https://www.sciencedirect.com/science/article/pii/B9780080453965001
470.
6 Born, J., Wilhelm, I. (2012), ‘System consolidation of memory during
sleep’, Psychological Research, 76: 192–203, doi.org/10.1007/s00426-
011-0335-6.
7 Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., Malone, T. W.
(2010), ‘Evidence for a Collective Intelligence Factor in the
Performance of Human Groups’, Science, 29 October: 686–8.
8 Minor, D., Brook, P., Bernoff, J. (2017), ‘Data From 3.5 Million
Employees Shows How Innovation Really Works’, Harvard Business
Review, 9 October.
9 Porter, B., Bozkaya, B. (2020), ‘Assessing the Effectiveness of Using
Live Interactions and Feedback to Increase Engagement in Online
Learning’, arXiv, revised 31 August, arxiv.org/abs/2008.08241.
CHAPTER 5: FUTURE-PROOFING: WHAT TO STUDY AND
WHICH INDUSTRIES ARE SAFER?
1 Naughton, K., Welch, D. (2019), ‘This Is What Peak Car Looks Like’,
Bloomberg, 28 February, www.bloomberg.com/news/features/2019-02-
28/this-is-what-peak-car-looks-like.
2 UK Office for National Statistics (2019), ‘Which Occupations Are at
Highest Risk of Being Automated?’, UK Government, 25 March,
www.ons.gov.uk/employmentandlabourmarket/peopleinwork/employme
ntandemployeetypes/articles/whichoccupationsareathighestriskofbeinga
utomated/2019-03-25.
3 Deutsch, D. (2012), ‘Philosophy Will Be the Key That Unlocks
Artificial Intelligence’, Guardian, 3 October,
www.theguardian.com/science/2012/oct/03/philosophy-artificial-
intelligence.
4 Montag, A. (2018), ‘Mark Cuban Says Studying Philosophy May Soon
Be Worth More than Computer Science – Here’s Why’, CNBC.com, 21
February, www.cnbc.com/2018/02/20/mark-cuban-philosophy-degree-
will-be-worth-more-than-computer-science.html.
5 Beltran, J. (2020), ‘AI Can Jump-Start Radiation Therapy for Cancer
Patients’, University of Texas Southwestern Medical Center, 27 January,
www.utsouthwestern.edu/newsroom/articles/year-2020/ai-radiation-
therapy.html.
6 Heszen-Klemens, I., Lapińska, E. (1984), ‘Doctor-Patient Interaction,
Patients’ Health Behavior and Effects of Treatment’, Social Science &
Medicine, 19 (1),
www.sciencedirect.com/science/article/abs/pii/0277953684901321.
7 ‘AI to Drive GDP Gains of $15.7 Trillion with Productivity,
Personalisation Improvements’ (2017), PwC, 27 June,
www.pwc.com/gx/en/news-room/press-releases/2017/ai-to-drive-gdp-
gains-of-15_7-trillion-with-productivity-personalisation-
improvements.html.
8 Savage, N. (2020), ‘How AI Is Improving Cancer Diagnostics’, Nature,
25 March, www.nature.com/articles/d41586-020-00847-2.
9 Grifantini, K. (2010), ‘Robotic Nurse Washes Human’, MIT Technology
Review, 10 November,
www.technologyreview.com/2010/11/10/199054/robotic-nurse-washes-
human/.
10 Van Mourik Broekman, A., et al. (2019), ‘The Impact of Art: Exploring
the Social-Psychological Pathways That Connect Audiences to Live
Performances’, Journal of Personality and Social Psychology, 116 (6):
942–65, pubmed.ncbi.nlm.nih.gov/30667257/.
11 Hardt, Y. (2011), ‘The Transformative Power of Performance: A New
Aesthetics by Erika Fischer-Lichte’, Dance Research Journal, 43: 117.
12 Hoover, A. (2023), ‘AI-Generated Music Is About to Flood Streaming
Platforms’, Wired, 17 April, www.wired.com/story/ai-generated-music-
streaming-services-copyright/.
13 Fingas, J. (2020), ‘Amazon May Turn Dying JC Penney and Sears
Stores into Warehouses’, Engadget, 9 August,
https://guce.engadget.com/copyConsent?sessionId=3_cc-
session_c55be3b4-e44e-4da4-a351-764c028afe04&lang=en-US.
14 Rubin, B. F. (2020), ‘Amazon Now Operates Seven Different Kinds of
Physical Stores. Here’s Why’, CNET, 28 February,
www.cnet.com/news/amazon-now-operates-seven-different-kinds-of-
physical-stores-heres-why/.
15 Butler, S., Wood, Z. (2017), ‘Amazon to Buy Whole Foods Market in
$13.7bn Deal’, Guardian, 16 June,
www.theguardian.com/business/2017/jun/16/amazon-buy-whole-foods-
market-organic-food-fresh.
16 Gross, D. (2017), ‘Why the Construction Industry May Be Robot-
Proof’, strategy+business, 24 May, www.strategy-
business.com/blog/Why-the-Construction-Industry-May-Be-Robot-
Proof?gko=41d6b.
17 Withers, P. (2018), ‘Robots take over: Machine to run for mayor in
Japan pledging “fair opportunities for all”’, Express, 17 April,
www.express.co.uk/news/world/947448/robots-japan-tokyo-mayor-
artificial-intelligence-ai-news.
CHAPTER 6: WHAT CAN HUMAN+AI SYSTEMS DO?
ADVENTURES IN COLLECTIVE INTELLIGENCE AND MORE
1 ‘The Ugly Truth About Meetings’ (2014), Fuze, 3 June,
www.fuze.com/blog/infographic-ugly-truth-about-meetings.
2 Petrov, C. (2020), ‘50 Gmail Statistics That Show How Big It Actually
Is In 2020’, TechJury, 30 June, techjury.net/blog/gmail-statistics/#gref.
3 Kleinman, Z. (2023), ‘Bard: Google’s rival to ChatGPT launches for
over-18s’, BBC News, 21 March,
https://www.bbc.com/news/technology-65018107.
4 Koenecke, A., Nam, A., Lake, E., Nudell, J., et al. (2020), ‘Racial
disparities in automated speech recognition’, Proceedings of the
National Academy of Sciences, 117(14): 7684–9,
www.pnas.org/content/117/14/7684.
5 Salthouse, T. A. (2009), ‘When Does Age-Related Cognitive Decline
Begin?’, Neurobiology of Aging, 30 (4): 507–14,
www.ncbi.nlm.nih.gov/pmc/articles/PMC2683339/.
6 Somaiya, R. (2014), ‘The A.P. Plans to Automate Quarterly Earnings
Articles’, New York Times, 30 June,
www.nytimes.com/2014/07/01/business/media/the-ap-plans-for-
computers-to-write-corporate-earnings-news.html.
7 Source: Glassdoor.com (2020).
8 Cogito Staff (2017), ‘Artificial Intelligence is Shaping the Future of
Customer Experience’, Cogito Corporation,
www.cogitocorp.com/resources/artificial-intelligence-shapes-customer-
experience/.
9 Pentland, A. (2012), ‘The New Science of Building Great Teams’,
Harvard Business Review, April, hbr.org/2012/04/the-new-science-of-
building-great-teams.
10 Kane, G. C. (2015), ‘“People Analytics” Through Super-Charged ID
Badges’, interview with Ben Waber (Humanyze), MIT Sloan
Management Review, 7 April.
11 Pentland (2012).
12 Calacci, D., Lederman, O., Shrier, D., Pentland, A. (2016), ‘Breakout:
An Open Measurement and Intervention Tool for Distributed Peer
Learning Groups’, MIT Media Lab/MIT Connection Science, arXiv,
July, arxiv.org/pdf/1607.01443.pdf.
13 Pentland (2012).
14 Waber, B., Magnolfi, J., Lindsay, G. (2014), ‘Workspaces That Move
People’, Harvard Business Review, October,
hbr.org/2014/10/workspaces-that-move-people.
15 Blenke, L. (2013), ‘The Role of Face-to-Face Interactions in the
Success of Virtual Project Teams’, doctoral thesis,
scholarsmine.mst.edu/cgi/viewcontent.cgi?
article=3306&context=doctoral_dissertations.
16 S&P SPIVA accessed 03 Sep 2023.
www.spglobal.com/spdji/en/research-insights/spiva/
17 Pentland, A. (2013), ‘Beyond the Echo Chamber’, Harvard Business
Review, November.
18 Adjodah, D., Chong, S., Leng, Y., Krafft, P., Pentland, A. (2017),
‘Large-Scale Experiment on the Importance of Social Learning and
Unimodality in the Wisdom of the Crowd’, arXiv, 29 December,
arxiv.org/abs/1712.10284.
19 FT Staff (2016), ‘Brexit poll tracker’, Financial Times, 13 June.
20 BBC Staff (2016), ‘EU Referendum Results’, BBC News,
www.bbc.com/news/politics/eu_referendum/results.

CHAPTER 7: BUILDING HUMAN+AI SYSTEMS: FROM PROMPT

ENGINEERING TO HIGH-PERFORMING TEAMS
1 OpenAI (2023), https://help.openai.com/en/articles/6654000-best-
practices-for-prompt-engineering-with-openai-api.
2 Shrier, D., Adjodah, D., Wu, W., Pentland, A. (2016), ‘Prediction
markets’, in Frontiers of Financial Technology (Cambridge, MA:
Visionary Future).
CHAPTER 8: ALTERNATIVE FUTURES: WALL-E OR STAR
TREK?
1 Kessler, S. (2011), ‘8 Star Trek Gadgets That Are No Longer Fiction’,
Mashable, mashable.com/2011/09/08/star-trek-gadgets/.
2 Ziady, H. (2023), ‘Europe is leading the race to regulate AI. Here’s what
you need to know’, CNN Business, 15 June,
www.cnn.com/2023/06/15/tech/ai-act-europe-key-
takeaways/index.html.

CHAPTER 9: THE URGENT NEED FOR POLICY: SHAPING A

BETTER TOMORROW
1 Field, H. (2023), ‘OpenAI’s Sam Altman reverses threat to cease
European operations’, CNBC.com, 26 May,
www.cnbc.com/2023/05/26/openai-ceo-sam-altman-reverses-threat-to-
cease-european-operations.html.
2 Smith, C. (2017), ‘Cracking the Enigma Code: How Turing’s Bombe
Turned the Tide of WWII’, BT, 2 November, http://home.bt.com/tech-
gadgets/cracking-the-enigma-code-how-turings-bombe-turned-the-tide-
of-wwii-11363990654704.
3 Newitz, A. (2009), ‘British Government Gives Official Apology to Alan
Turing for Homophobia’, Gizmodo, 9 November,
io9.gizmodo.com/british-government-gives-official-apology-to-alan-
turin-5357059.
4 Chuvpilo, G. (2020), ‘Who’s Ahead in AI Research in 2020?’, Medium,
20 July, medium.com/@chuvpilo/whos-ahead-in-ai-research-in-2020-
2009da5cd799.
5 ‘Research at MIT’ and ‘MIT Facts 2020: Lincoln Laboratory’ (2020),
MIT, facts.mit.edu/research-highlights/ and
web.mit.edu/facts/lincoln.html.
6 MIT News Office (2017), ‘IBM and MIT to Pursue Joint Research in
Artificial Intelligence, Establish New MIT-IBM Watson AI Lab’, MIT
News, 7 September, http://news.mit.edu/2017/ibm-mit-joint-research-
watson-artificial-intelligence-lab-0907.
7 Iyengar, R. (2019), ‘Stephen Schwarzman Gives $188 Million to
Oxford to Research AI Ethics’, CNN Business, 19 June,
www.cnn.com/2019/06/19/tech/stephen-schwarzman-oxford-ai-
donation/index.html.
8 Navarria, G. (2016), ‘How the Internet Was Born: From the ARPANET
to the Internet’, Conversation, 2 November, theconversation.com/how-
the-internet-was-born-from-the-arpanet-to-the-internet-68072.
9 Reynolds, C. (2019), ‘UK Government to Spend £250 Million Creating
a National Artificial Intelligence Lab’, Computer Business Review, 8
August, www.cbronline.com/news/national-artificial-intelligence-lab.
10 Rasser, M., Lamberth, M., Riikonen, A., Guo, C., Horowitz, M.,
Scharre, P. (2019), ‘The American AI Century: A Blueprint for Action’,
Center for a New American Security, 17 December,
www.cnas.org/publications/reports/the-american-ai-century-a-blueprint-
for-action.
11 Acharya, A., Arnold, Z. (2019), ‘Chinese Public AI R&D Spending:
Provisional Findings’, CSET Issue Brief, Center for Security and
Emerging Technology, Georgetown University, pp. 13–14,
cset.georgetown.edu/wp-content/uploads/Chinese-Public-AI-RD-
Spending-Provisional-Findings-2.pdf.
12 European Commission (2020), ‘Artificial Intelligence – Shaping
Europe’s digital future’, 20 July, https://ec.europa.eu/digital-single-
market/en/artificial-intelligence.
13 Fioretti, J. (2018), ‘EU to Invest 1.5 Billion Euros in AI to Catch up
with US, Asia’, Reuters, 25 April, www.reuters.com/article/us-eu-
artificialintelligence/eu-to-invest-1-5-billion-euros-in-ai-to-catch-up-
with-us-asia-idUSKBN1HW1WL.
14 Roberts, H., Babuta, A., Morley, J., Thomas, C., Mariarosiario, T.,
Floridi, L. (2023), ‘Artificial intelligence regulation in the United
Kingdom: a path to good governance and global leadership?’, Internet
Policy Review, 26 May, policyreview.info/articles/analysis/artificial-
intelligence-regulation-united-kingdom-path-good-governance.
15 European Commission (2020).
16 ‘Artificial Intelligence – A European Approach to Excellence and Trust’
(2020), White Paper, European Commission, 19 February,
ec.europa.eu/info/sites/info/files/commission-white-paper-artificial-
intelligence-feb2020_en.pdf.
17 Satariano, A. (2020), ‘Silicon Valley Heads to Europe, Nervous About
New Rules’, New York Times, 16 February,
www.nytimes.com/2020/02/16/technology/europe-new-AI-tech-
regulations.html.
18 Grüll, P. (2020), ‘Germany Calls for Tightened AI Regulation at EU
Level’, Euractiv, 30 June,
www.euractiv.com/section/digital/news/germany-calls-for-tightened-ai-
regulation-at-eu-level/.
19 Carberry, S. (2023), ‘White House Asks for $842 Billion in 2024 DOD
Funding, 3.2 Percent Above 2023 Budget (UPDATED)’, 9 March,
www.nationaldefensemagazine.org/articles/2023/3/9/white-house-asks-
for-842-billion-in-2024-dod-funding-32-percent-above-2023-budget.
20 Clay, W. L., Cravins, Y. P., Finch, B. N., Killalea, L., Moeller, E. V.
(2023), ‘Follow the Money: AI Winners in President Biden’s FY 2024
Budget Request’, JDSupra.com, 8 May,
https://www.jdsupra.com/legalnews/follow-the-money-ai-winners-in-
4977015/.
21 Toth M., Sweet, A. (2023), ‘Is the Biden administration late to
WWIII?’, The Hill, 9 February, thehill.com/opinion/national-
security/3850202-is-the-biden-administration-late-to-wwiii/.
Acknowledgements

Any creative effort I have found worthwhile is a collaborative process, and

this book is no exception. Building on the foundations of research
assistance from Katie Makstenieks and Adele Jashari, I was able to describe
the work of numerous AI innovators. I have taken inspiration in these pages
from collaborators such as Ben Yablon, Tom Gardner, Beth Porter, Alex
Pentland and Joost Bonsen, thought leaders Mike Richey, Ben Vigoda,
Sanjeev Vohra and Arram Sabeti, and former students like Snejina Zacharia.
Tom Asker and the team at Little, Brown believed enough in me for another
go at helping the world understand why AI is so important, and so vitally
important to understand right now. Cynthia Renard graciously stepped in at
the last minute to redraw all of the illustrations in this book.. I’m glad you,
my reader, have taken the time to delve into the topic. I hope that you will
join the effort to align AI for the betterment of humanity.
Index

2001: A Space Odyssey 17

Abbott, Ryan 190
academic research 48–9, 56
Accenture 41, 73–4, 102, 104, 133–4
accountability 103, 108
accountancy 69–71, 101
advertising 12, 61–2, 71
age of intelligence 73, 104
agriculture 112
AI Act 169, 180, 183
AI coach 42, 159, 175–6
AI hallucinations 16–18
AI Impact Navigator 50
AI programmers 72–3
AI Without Borders 188–9
Airbnb 28, 67, 73
Akamai 61
Alexa 16, 27, 120
algorithms 12, 25, 27–8, 63–4, 105, 169
Altered Carbon 164–5
Altman, Sam xv, 180
Amazon 28, 60, 68, 109–10
Echo 36, 119
MTurk 39–40
animal husbandry 112
Anthropic.AI 60
Arpanet 181
artificial general intelligence (AGI) 34–5, 101
Asimov, Isaac 15, 165
AutoML 72
autonomous cars 26, 32, 56, 66, 86, 100
Autopilot system 26, 86

Baidu 32
Bard 36, 60, 185–6
behavioural modification 171
benevolence 172
Biden, Joe 14, 186
‘Black Swan’ events 26
blockchain 69–70, 83, 100
Boeing 152–4, 157
Bonsen, Joost 4, 164
Bostrom, Nick 34–5
brain implants 104–5, 174
brainstorming 95–6, 116, 137
Brexit vote 5–7, 143
Broadcast.com 101–2
Brookings Institution 55
Brynjolfsson, Erik 4–5, 72

Cameron, James 166, 182

Carnegie Mellon University 38
CCTV 50
centaurs 40–1, 144
Center for a New American Security (CNAS) 181
chatbots 8–10, 12–14, 22–3, 66
ChatGPT xiv, xv, 1–4, 17, 33–4, 36, 60, 150, 185–6
Chegg 2
chess 37, 40
China 47–8, 75, 181
class-based bifurcation 110
Cloud Atlas 167
cluster analysis 29
cognitive flexibility 79–98, 105
cognitive psychology 103–4
collective bargaining 68
collective intelligence 41–2, 92–3, 141, 143–4, 153–4, 158–9, 167, 185
College of Computing 181
compensation ratio 52
Computational Law Project 59
consolidation 88
construction industry 111–12
conversational AI 35–6, 101, 158, 185–6
copyright 108
cortisol 88
counterfeit goods 70
COVID-19 7, 64, 81, 154
Cowles, Joshua 172
creative exploration 93–6
creative jobs 66, 102, 108, 113, 159–60
critical thinking 3, 94, 101, 105–6
Cuban, Mark 101–2
cyber war 186

Dalio, Ray 160

Davis, Gerald 67
Deep Blue 40
deep learning 17, 30–2, 35, 37, 57, 75
design 102
Deutsch, David 101
developing vs developed countries 46–7
diversity 92, 95
DOCTOR script 8
dystopian futures 163, 165–6, 168, 170, 182

Echo 36, 119

eCommerce 109, 182
edge conditions 25–6
eDiscovery 58
‘elevator asset’ businesses 53–4, 69
ELIZA expert system 8, 9–10, 22–3, 104
emotional intelligence (EQ) 92, 93, 105–8, 113
emotional robot companions 104
empathy 102, 108, 113
employment 4–5, 30, 34, 39–40, 42–3, 45–76
cognitive flexibility 79–98
future proofing 99–114
energy systems 110–11
enterprise resource planning (ERP) 74
environmental costs 36–7, 47
ethics 35, 49–50, 103, 169, 172, 174–6, 181, 187–8
European Union (EU) 169, 176, 180, 181, 183–4
Evercore ISI 50, 54
Expedia 46
expert systems 22–5, 27, 28, 30, 57, 75
see also ELIZA expert system
explainability 173
extended intelligence 59

F-Secure 6
Facebook 10, 12–13, 63–4, 105, 174–5
facial recognition systems 25–6
fake citations 18
fake news 18, 64
The Fear Index 14
feedback 67–8, 80, 138, 160
loops 10, 14, 109, 136, 155
financial crisis 2008 26
financial flash crashes 169
fintech 65, 88, 100, 141–2, 143
Floridi, Luciano 35, 172
Foley, Michael xiii–xiv
Frankenstein 15, 165

gamification 93
gaming 41
Gates, Bill 36
general artificial intelligence see artificial general intelligence
generative AI 4, 14, 18, 32–4, 50–1, 150, 180, 183
Georgia Tech 108
gestural interfaces 164
GitHub Copilot 72
Google xiv, 27, 38–9, 60, 61–2, 71
AutoML 72
brain 104
Voice 119
governance 173
government
future policy 176–7, 179–91
jobs 112
Greenwood, Dazza 59
group project model 91
guardian AI 173–4

hair stylists 110

Harris, Robert 14
healthcare 103, 108
hidden network 156–8
Hoffman, Reid 36
HSBC 49
Hulu 61
Human+AI hybrid systems 5, 37–42, 114, 115–45, 147–61, 189–90
human intelligence 188–9
Human Strategy 171

IBM 2, 40, 181

Inflection 36
information cascades 64
instrumenting 93
insurance 65
Insurify 65
intellectual property rights 101
intelligent organisations 158–60
interface design 102–3
internet 25, 38, 83, 167, 181, 182
see also social media
Internet Research Group 174
investment banks 51–3, 69, 75
Iran 6, 13

JARVIS 36
justice and fairness 35, 173

Kasparov, Gary 40
Kayak 65
Kennedy, John F. 191–2

labour-profitability model 52
language 16, 71, 106
LLM xv, 3–4, 17, 33
translation 32
see also ChatGPT
large language model (LLM) xv, 3–4, 17, 33
law of weak ties 95
lawyers 56–60, 101
learning
AI future 174–6
cognitive flexibility 80–96
future-proofing 99–114
Lefevre, David 56
LEGO 60, 93
LIDAR 86
LinkedIn 90
live performances 108–9
Long Term Capital Management hedge fund 26
Lyft 67, 68

machine learning 25–7, 28–9, 31–2, 57, 75, 85

Maginot Line 184
manufacturing 5, 45–7, 50
Marshmallow Challenge 94–5
Martin School 65
massage therapy 110
Mastercard 74
The Matrix 15, 165–6
mechanical turk 37, 42, 160
media 60–4, 71
meditation 96
memory 87–8
mental models 85–6, 95
Meta 16, 49
Microsoft 13, 36, 63, 139
Midjourney 34
Miller, Alllie K. 2
mindfulness 96
Minority Report 164
Minsky, Marvin 170
MIT 65, 85, 92, 152–4, 164
Computational Law Project 59
funding 180–1
Media Lab 60
Mitchell, David 167
Montjoye, Yves-Alexandre de 3
MTurk 39–40
Musk, Elon xv

NBC 61
Negroponte, Nicholas 60
net neutrality 182
Netflix 2, 27, 73, 132, 183
neural networks 30–2
layers 17
neuroplasticity 94
neuroscience 103–4
non-malevolence 172–3
North Korea 13

OECD 70, 74
OpenAI xv, 2, 4, 34, 36, 150, 180
optical character recognition (OCR) 57
outlier events 26
Oxford Economics 46
Oxford Fintech Lab 88
Oxford Internet Institute 35, 172
Oxford University 34, 65, 80, 85, 141, 180, 181

Palantir 112
patents 55–6
peer learning 91–3
Pentland, Alex (Sandy) 4, 59, 85, 93, 158–9, 171
performing arts 108–9
philosophy 101, 106
photo-processing algorithms 27–8
Pixar 151, 165
Porsche 40–1
Porter, Beth 91, 93
positive reinforcement 63, 90
prediction markets 41, 142–4
privacy 50
productivity gap 50–1
prompt engineering 17, 149–50
psychology 103–4
public private partnership (PPP or P3) 187
publishing 71

quantitative traders 159

reCAPTCHA 38–9
Redfin 46
reflection 86–9
reliability 16–17
restaurants 110
reviews 28, 67
Richey, Mike 151–4, 157
Riff Analytics 93
Roberts, Stephen 34
robotic automation 5
Roddenberry, Gene 167
Rogerian therapy 9, 23
Rogers, Tom 61
rules for AI 172–4
rules-based systems 22–3, 57, 75
Russian intelligence 6–8, 13
Ryan, Jeri 167

‘sage on a stage’ model 91

Samsung 2
scaffolding 85
scambots 22
Schwarzman, Stephen 181
Sci Fab 164, 165
science fiction 163–6, 170
security 38–9, 50, 186–7
self-driving cars see autonomous cars
semantic analysis 57
semi-skilled labour 66
Senge, Peter 85
shadow network 157–8
Shelley, Mary 15, 165
SingularityNET xiii–xiv
Siri 36, 119
Skynet 166
sleep 88
smartphones 127, 129, 164, 167–8
social intelligence 66
social media
Brexit vote 6
trolls 174
US Presidential election 7
see also Facebook; Twitter
soft skills 105–8, 153
Spielberg, Steven 164
Star Trek 164, 166–7, 168
startups 186–7
streamers 71
structured data 27–8
Suleyman, Mustafa 36
supervised learning 29–30
surveillance 171

tax avoidance 70
Tay 13
teams 151–8
TensorFlow xiv, 27
Terminator 166
Tesla Autopilot system 26, 86
Tortoise 48
travel industry 46, 65
tricorder 168
trolls 174–5
Trump, Donald 7, 10–11
trust 154–5
Trusted AI Institute xvi, 49
Tucci, Chris 3
Tulchinsky, Igor 159, 160
Turing, Alan 8–9, 180
Turing Test 8–9, 180
tutorial system 80
Twitter (X) 6, 7, 13, 28, 174–5

Uber 46, 67–8

‘under specification’ 18
Underkoffler, John 164
United Nations Economic Commission for Africa (UNECA) 70
unskilled labour 66
unstructured data 27–8
unsupervised learning 29–30
US politics 6–8, 10–12
US Presidential election 6–8, 13
User Experience (UX) 102
utopian futures 163, 166–8

Veloso, Francisco 1, 3
Ventresca, Marc 88
venture capital 186–7
Verweij, Gerard 107
Vestager, Margrethe 183
videoconferences 155
Vigoda, Ben 168–9
Visionary Future xv
Vohra, Sanjeev 41, 73–4, 102, 104
von Kempelen, Wolfgang 37

Wachowski, Lana & Lilly 65

WALL-E 165.168
Walmart 67–8
weather forecasts 40
Webb, Michael 55–6
Weizenbaum, Joseph 8
Westworld 14–15
Wiesner, Jerome 60
Wolfram, Stephen 33
World Bank 70
World Economic Forum 47, 74
WorldQuant 159
Wozniak, Steve 26
Wujec, Tom 94

Yablon, Ben 173–4

YouTube 39

Zacharia, Snejina 65
Zaslav, David 61
 Footnotes

* TensorFlow is a powerful, flexible, free library of software that enables

a wide range of artificial intelligence applications. It is able to run
across a diverse set of hardware platforms. By making it available,
Google enabled a new generation of AI startups across many different
industries. Google followed up in 2018 with BERT, which laid the
foundations of the current Large Language Models (LLMs).
* This has been mostly true for a number of years, but in 2020 it was
revealed that Facebook was deliberately diminishing traffic to liberal
news websites and accentuating traffic to conservative sites. According
to the Washington Post and other sources, in an effort to avoid
perception of liberal bias and stave off attacks by President Trump,
Facebook engaged in conservative bias instead. More can be found in
this article:
www.washingtonpost.com/technology/2020/11/01/facebook-election-
misinformation/.
* An interesting bit of Matrix lore: allegedly the original script had people
tied into a massive neural network to enhance computing power, which
made quite a bit more sense, but the filmmakers feared that movie
audiences wouldn’t be intelligent enough to follow along with this, so
changed it to ‘people become batteries’, which doesn’t actually work in
our understanding of physics.
* A book project for which I initiated and helped write the proposal.
* ‘Dysergy’, where two systems engage in destructive interference with
each other (‘the whole is less than the sum of the parts’), is the opposite
of ‘synergy’, where two systems positively reinforce each other. This
definition is in some debate but is enjoying growing prevalence.