What Is Ai v2

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

McKinsey Explainers

What is AI (artificial
intelligence)?
Artificial intelligence is a machine’s ability to perform some
cognitive functions we usually associate with human minds.

April 2024
Humans and machines: a match made in individual could in 31,688,765,000 years. And beyond
productivity heaven. Our species wouldn’t have computation, which machines have long been faster
gotten very far without our mechanized work­horses. at than we have, computers and other devices are
From the wheel that revolutionized agriculture to now acquiring skills and perception that were once
the screw that held together increasingly complex unique to humans and a few other species.
construction projects to the robot-enabled
assembly lines of today, machines have made life AI is a machine’s ability to perform the cognitive
as we know it possible. And yet, despite their functions we associate with human minds, such as
seemingly endless utility, humans have long feared perceiving, reasoning, learning, interacting with the
machines—more specifically, the possibility that environment, problem-solving, and even exercising
machines might someday acquire human intelligence creativity. You’ve probably interacted with AI even if
and strike out on their own. you don’t realize it—voice assistants like Siri and
Alexa are founded on AI technology, as are some
But we tend to view the possibility of sentient customer service chatbots that pop up to help
machines with fascination as well as fear. This you navigate websites.
curiosity has helped turn science fiction into actual
science. Twentieth-century theoreticians, like Applied AI—simply, artificial intelligence applied to
computer scientist and mathematician Alan Turing, real-world problems—has serious implications for
envisioned a future where machines could the business world. By using artificial intelligence,
perform functions faster than humans. The work companies have the potential to make business
of Turing and others soon made this a reality. more efficient and profitable. But ultimately, the
Personal calculators became widely available in the value of AI isn’t in the systems themselves. Rather,
1970s, and by 2016, the US census showed that it’s in how companies use these systems to assist
89 percent of American households had a computer. humans—and their ability to explain to shareholders
Machines—smart machines at that—are now just and the public what these systems do—in a way
an ordinary part of our lives and culture. that builds trust and confidence.

Those smart machines are also getting faster and For more about AI, its history, its future, and how to
more complex. Some computers have now crossed apply it in business, read on.
the exascale threshold, meaning they can perform
as many calculations in a single second as an Learn more about QuantumBlack, AI by McKinsey.

Beyond computation, computers and


other devices are now acquiring skills
and perception that were once unique to
humans and a few other species.

What is AI (artificial intelligence)? 2


What is machine learning? What is deep learning?
Machine learning is a form of artificial intelligence Deep learning is a more advanced version of
that can adapt to a wide range of inputs, including machine learning that is particularly adept
large sets of historical data, synthesized data, or at processing a wider range of data resources (text
human inputs. (Some machine learning algorithms as well as unstructured data including images),
are specialized in training themselves to detect requires even less human intervention, and can
patterns; this is called deep learning. See Exhibit 1.) often produce more accurate results than
These algorithms can detect patterns and learn traditional machine learning. Deep learning uses
how to make predictions and recommendations by neural networks—based on the ways neurons
processing data, rather than by receiving explicit interact in the human brain—to ingest data and
programming instruction. Some algorithms can also process it through multiple neuron layers that
adapt in response to new data and experiences recognize increasingly complex features of the data.
to improve over time. For example, an early layer might recognize
something as being in a specific shape; building on
The volume and complexity of data that is now being this knowledge, a later layer might be able to
generated, too vast for humans to process and identify the shape as a stop sign. Similar to machine
apply efficiently, has increased the potential of learning, deep learning uses iteration to self-correct
machine learning, as well as the need for it. In the and improve its prediction capabilities. For example,
years since its widespread deployment, which once it “learns” what a stop sign looks like, it can
began in the 1970s, machine learning has had an recognize a stop sign in a new image.
impact on a number of industries, including
achievements in medical-imaging analysis and Learn more about QuantumBlack, AI by McKinsey.
high-resolution weather forecasting.

Web 2024
GenAI2
Exhibit
Exhibit 2b 1of 2

Artificial intelligence is a machine’s ability to perform some cognitive


functions we usually associate with human minds.

The evolution of artificial intelligence

Artificial intelligence
The science and
engineering of Machine learning
making intelligent A major
machines breakthrough Deep learning
in achieving AI An advanced
branch of machine Generative AI
learning An advanced branch
of deep learning

AI is the broad field of Machine learning Deep learning uses neural Generative AI is a branch of
developing machines algorithms detect networks, inspired by the deep learning that uses
that can replicate patterns in large ways neurons interact in the exceptionally large neural
human behavior, data sets and learn human brain, to ingest data networks called large
including tasks related to make predictions and process it through language models (with
to perceiving, by processing data, multiple iterations that learn hundreds of billions of neurons)
reasoning, learning, rather than by increasingly complex that can learn especially
and problem-solving. receiving explicit features of the data and abstract patterns. Language
programming make increasingly models applied to interpret and
instructions. sophisticated predictions. create text, video, images, and
data are known as generative AI.

McKinsey & Company

What is AI (artificial intelligence)? 3


The volume and complexity of data that
is now being generated, too vast for
humans to process and apply efficiently,
has increased the potential of machine
learning, as well as the need for it.

What is generative AI? MIT physicist Rodney Brooks shared details on the
Generative AI (gen AI) is an AI model that generates four previous stages of AI:
content in response to a prompt. It’s clear that
generative AI tools like ChatGPT and DALL-E (a tool — Symbolic AI (1956). Symbolic AI is also known as
for AI-generated art) have the potential to change classical AI, or even GOFAI (good old-fashioned
how a range of jobs are performed. Much is still AI). The key concept here is the use of symbols
unknown about gen AI’s potential, but there are and logical reasoning to solve problems. For
some questions we can answer—like how gen AI example, we know a German shepherd is a dog,
models are built, what kinds of problems they are which is a mammal; all mammals are warm-
best suited to solve, and how they fit into the broader blooded; therefore, a German shepherd should
category of AI and machine learning. be warm-blooded.

For more on generative AI and how it stands to The main problem with symbolic AI is that humans
affect business and society, check out our Explainer still need to manually encode their knowledge
“What is generative AI?” of the world into the symbolic AI system, rather
than allowing it to observe and encode
relationships on its own. As a result, symbolic AI
What is the history of AI? systems struggle with situations involving
real-world complexity. They also lack the ability
The term “artificial intelligence” was coined in 1956
to learn from large amounts of data.
by computer scientist John McCarthy for a workshop
at Dartmouth. But he wasn’t the first to write about
Symbolic AI was the dominant paradigm of AI
the concepts we now describe as AI. Alan Turing
research until the late 1980s.
introduced the concept of the “imitation game” in a
1950 paper. That’s the test of a machine’s ability
— Neural networks (1954, 1969, 1986, 2012).
to exhibit intelligent behavior, now known as the
Neural networks are the technology behind
“Turing test.” He believed researchers should focus
the recent explosive growth of gen AI. Loosely
on areas that don’t require too much sensing and
modeling the ways neurons interact in the
action, things like games and language translation.
human brain, neural networks ingest data and
Research communities dedicated to concepts
process it through multiple iterations that learn
like computer vision, natural language understanding,
increasingly complex features of the data. The
and neural networks are, in many cases, several
neural network can then make determinations
decades old.

What is AI (artificial intelligence)? 4


about the data, learn whether a determination is In 2012, Hinton and two of his students
correct, and use what it has learned to make highlighted the power of deep learning. They
determinations about new data. For example, applied Hinton’s algorithm to neural networks
once it “learns” what an object looks like, it can with many more layers than was typical,
recognize the object in a new image. sparking a new focus on deep neural networks.
These have been the main AI approaches of
Neural networks were first proposed in 1943 recent years.
in an academic paper by neurophysiologist
Warren McCulloch and logician Walter Pitts. — Traditional robotics (1968). During the first few
Decades later, in 1969, two MIT researchers decades of AI, researchers built robots to advance
mathematically demonstrated that neural research. Some robots were mobile, moving
networks could perform only very basic tasks. around on wheels, while others were fixed, with
In 1986, there was another reversal, when articulated arms. Robots used the earliest
computer scientist and cognitive psychologist attempts at computer vision to identify and
Geoffrey Hinton and colleagues solved the navigate through their environments or to
neural network problem presented by the MIT understand the geometry of objects and
researchers. In the 1990s, computer scientist maneuver them. This could include moving
Yann LeCun made major advancements in around blocks of various shapes and colors.
neural networks’ use in computer vision, while Most of these robots, just like the ones that have
Jürgen Schmidhuber advanced the application been used in factories for decades, rely on
of recurrent neural networks as used in highly controlled environments with thoroughly
language processing. scripted behaviors that they perform repeatedly.

Case study: Vistra and the Martin Lake Power Plant

Vistra is a large power producer in the Vistra and a McKinsey team, including data Overall, the AI-powered HRO helped Vistra
United States, operating plants in 12 states scientists and machine learning engineers, achieve the following:
with a capacity to power nearly 20 million built a multilayered neural network model.
— approximately 1.6 million metric tons of
homes. Vistra has committed to achieving The model combed through two years’
carbon abated annually
net-zero emissions by 2050. In support worth of data at the plant and learned
of this goal, as well as to improve overall which combination of factors would attain — 67 power generators optimized
efficiency, QuantumBlack, AI by McKinsey the most efficient heat rate at any point
— $60 million saved in about a year
worked with Vistra to build and deploy in time. When the models were accurate to
an AI-powered heat rate optimizer (HRO) 99 percent or higher and run through a
Read more about the Vistra story here.
at one of its plants. rigorous set of real-world tests, the team
converted them into an AI-powered engine
“Heat rate” is a measure of the thermal that generates recommendations every
efficiency of the plant; in other words, it’s 30 minutes for operators to improve the
the amount of fuel required to produce plant’s heat rate efficiency. One seasoned
each unit of electricity. To reach the optimal operations manager at the company’s
heat rate, plant operators continuously plant in Odessa, Texas, said, “There are
monitor and tune hundreds of variables, things that took me 20 years to learn
such as steam temperatures, pressures, about these power plants. This model
oxygen levels, and fan speeds. learned them in an afternoon.”

What is AI (artificial intelligence)? 5


They have not contributed significantly to the our lives. Executives should begin working to
advancement of AI itself. understand the path to machines achieving human-
level intelligence now and making the transition to
But traditional robotics did have significant a more automated world.
impact in one area, through a process called
“simultaneous localization and mapping” (SLAM). For more on AGI, including the four previous
SLAM algorithms helped contribute to self- attempts at AGI, read our Explainer.
driving cars and are used in consumer products
like vacuum cleaning robots and quadcopter
drones. Today, this work has evolved into What is narrow AI?
behavior-based robotics, also referred to as Narrow AI is the application of AI techniques to a
haptic technology because it responds to specific and well-defined problem, such as chatbots
human touch. like ChatGPT, algorithms that spot fraud in credit
card transactions, and natural-language-processing
— Behavior-based robotics (1985). In the real engines that quickly process thousands of legal
world, there aren’t always clear instructions for documents. Most current AI applications fall into
navigation, decision making, or problem-solving. the category of narrow AI. AGI is, by contrast,
Insects, researchers observed, navigate very AI that’s intelligent enough to perform a broad
well (and are evolutionarily very successful) with range of tasks.
few neurons. Behavior-based robotics
researchers took inspiration from this, looking Learn more about QuantumBlack, AI by McKinsey.
for ways robots could solve problems with
partial knowledge and conflicting instructions.
These behavior-based robots are embedded
How is the use of AI expanding?
with neural networks.
AI is a big story for all kinds of businesses, but some
companies are clearly moving ahead of the pack.
Learn more about QuantumBlack, AI by McKinsey.
Our state of AI in 2022 survey showed that adoption
of AI models has more than doubled since 2017—
and investment has increased apace. What’s more,
What is artificial general intelligence?
the specific areas in which companies see value
The term “artificial general intelligence” (AGI) was from AI have evolved, from manufacturing and risk
coined to describe AI systems that possess to the following:
capabilities comparable to those of a human. In
theory, AGI could someday replicate human-like — marketing and sales
cognitive abilities including reasoning, problem-
solving, perception, learning, and language — product and service development
comprehension. But let’s not get ahead of ourselves:
the key word here is “someday.” Most researchers — strategy and corporate finance
and academics believe we are decades away from
realizing AGI; some even predict we won’t see One group of companies is pulling ahead of its
AGI this century, or ever. Rodney Brooks, an MIT competitors. Leaders of these organizations
roboticist and cofounder of iRobot, doesn’t believe consistently make larger investments in AI, level up
AGI will arrive until the year 2300. their practices to scale faster, and hire and upskill
the best AI talent. More specifically, they link AI
The timing of AGI’s emergence may be uncertain. strategy to business outcomes and “industrialize” AI
But when it does emerge—and it likely will— operations by designing modular data architecture
it’s going to be a very big deal, in every aspect of that can quickly accommodate new applications.

What is AI (artificial intelligence)? 6


What are the limitations of AI It’s also important to keep a human in the loop (that
models? How can these potentially is, to make sure a real human checks the output
be overcome? of a gen AI model before it is published or used) and
avoid using gen AI models for critical decisions,
We have yet to see the longtail effect of gen AI
such as those involving significant resources or
models. This means there are some inherent risks
human welfare.
involved in using them—both known and unknown.

It can’t be emphasized enough that this is a new field.


The outputs gen AI models produce may often
The landscape of risks and opportunities is likely
sound extremely convincing. This is by design. But
to continue to change rapidly in the coming years.
sometimes the information they generate is just
As gen AI becomes increasingly incorporated
plain wrong. Worse, sometimes it’s biased (because
into business, society, and our personal lives, we
it’s built on the gender, racial, and other biases of
can also expect a new regulatory climate to take
the internet and society more generally).
shape. As organizations experiment—and create
value—with these tools, leaders will do well to keep
It can also be manipulated to enable unethical or
a finger on the pulse of regulation and risk.
criminal activity. Since gen AI models burst onto the
scene, organizations have become aware of users
Learn more about QuantumBlack, AI by McKinsey.
trying to “jailbreak” the models—that means trying
to get them to break their own rules and deliver
biased, harmful, misleading, or even illegal content.
Gen AI organizations are responding to this threat What is the AI Bill of Rights?
in two ways: for one thing, they’re collecting The Blueprint for an AI Bill of Rights, prepared by
feedback from users on inappropriate content. the US government in 2022, provides a framework
They’re also combing through their databases, for how government, technology companies, and
identifying prompts that led to inappropriate content, citizens can collectively ensure more accountable
and training the model against these types AI. As AI has become more ubiquitous, concerns
of generations. have surfaced about a potential lack of transparency
surrounding the functioning of gen AI systems, the
But awareness and even action don’t guarantee data used to train them, issues of bias and fairness,
that harmful content won’t slip the dragnet. potential intellectual property infringements,
Organizations that rely on gen AI models should privacy violations, and more. The Blueprint comprises
be aware of the reputational and legal risks five principles that the White House says should
involved in unintentionally publishing biased, “guide the design, use, and deployment of automated
offensive, or copyrighted content. systems to protect [users] in the age of artificial
intelligence.” They are as follows:
These risks can be mitigated, however, in a few ways.
“Whenever you use a model,” says McKinsey partner — The right to safe and effective systems. Systems
Marie El Hoyek, “you need to be able to counter should undergo predeployment testing, risk
biases and instruct it not to use inappropriate or identification and mitigation, and ongoing
flawed sources, or things you don’t trust.” How? For monitoring to demonstrate that they are adhering
one thing, it’s crucial to carefully select the initial to their intended use.
data used to train these models to avoid including
toxic or biased content. Next, rather than employing — Protections against discrimination by algorithms.
an off-the-shelf gen AI model, organizations Algorithmic discrimination is when automated
could consider using smaller, specialized models. systems contribute to unjustified different
Organizations with more resources could also treatment of people based on their race, color,
customize a general model based on their own data ethnicity, sex, religion, age, and more.
to fit their needs and minimize biases.

What is AI (artificial intelligence)? 7


— Protections against abusive data practices, via Even though AI regulations are still being developed,
built-in safeguards. Users should also have organizations should act now to avoid legal,
agency over how their data is used. reputational, organizational, and financial risks. In
an environment of public concern, a misstep
— The right to know that an automated system is could be costly. Here are four no-regrets, preemptive
being used, and a clear explanation of how actions organizations can implement today:
and why it contributes to outcomes that affect
the user. — Transparency. Create an inventory of models,
classifying them in accordance with
— The right to opt out, and access to a human who regulation, and record all usage across the
can quickly consider and fix problems. organization that is clear to those inside
and outside the organization.
At present, more than 60 countries or blocs have
national strategies governing the responsible — Governance. Implement a governance structure
use of AI (Exhibit 2). These include Brazil, China, the for AI and gen AI that ensures sufficient
European Union, Singapore, South Korea, and oversight, authority, and accountability both
the United States. The approaches taken vary from within the organization and with third
guidelines-based approaches, such as the parties and regulators.
Blueprint for an AI Bill of Rights in the United States,
to comprehensive AI regulations that align with — Data, model, and technology management.
existing data protection and cybersecurity
regulations, such as the EU’s AI Act, due in 2024. • Data management. Proper data
management includes awareness of data
There are also collaborative efforts between sources, data classification, data
countries to set out standards for AI use. The US– quality and lineage, intellectual property,
EU Trade and Technology Council is working and privacy management.
toward greater alignment between Europe and the
United States. The Global Partnership on Artificial • Model management. Organizations should
Intelligence, formed in 2020, has 29 members establish principles and guardrails for AI
including Brazil, Canada, Japan, the United States, development and use them to ensure all AI
and several European countries. models uphold fairness and bias controls.

Web 2024
GenAI2
Exhibit 1 of 2
Exhibit 2

Regulations related to AI governance vary around the world.

As of November 2023, nonexhaustive


Type of policy: General AI legislation proposed Example countries without general
Nonbinding principles (eg, OECD) or being finalized AI legislation
Japan Brazil Australia
Singapore Canada India
United Arab Emirates China New Zealand
United Kingdom South Korea Saudi Arabia
United States European Union

Source: OECD; McKinsey analysis

McKinsey & Company

What is AI (artificial intelligence)? 8


• Cybersecurity and technology management. Learn more about QuantumBlack, AI by McKinsey,
Establish strong cybersecurity and and check out AI-related job opportunities if you’re
technology to ensure a secure environment interested in working at McKinsey.
where unauthorized access or misuse
is prevented. Articles referenced:

— Individual rights. Make users aware when they — “As gen AI advances, regulators—and risk
are interacting with an AI system, and provide functions—rush to keep pace,” December 21,
clear instructions for use. 2023, Andreas Kremer, Angela Luget, Daniel
Mikkelsen, Henning Soller, Malin Strandell-
Jansson, and Sheila Zingg
How can organizations scale up
their AI efforts from ad hoc projects — “What is generative AI?,” January 19, 2023
to full integration?
— “Tech highlights from 2022—in eight charts,”
Most organizations are dipping a toe into the AI
December 22, 2022
pool—not cannonballing. Slow progress toward
widespread adoption is likely due to cultural
— “Generative AI is here: How tools like ChatGPT
and organizational barriers. But leaders who
could change your business,” December 20, 2022,
effectively break down these barriers will be best
Michael Chui, Roger Roberts, and Lareina Yee
placed to capture the opportunities of the AI era.
And—crucially—companies that can’t take full
— “The state of AI in 2022—and a half decade in
advantage of AI are already being sidelined by those
review,” December 6, 2022, Michael Chui,
that can, in industries like auto manufacturing
Bryce Hall, Helen Mayhew, Alex Singla, and
and financial services.
Alex Sukharevsky
To scale up AI, organizations can make three
— “Why businesses need explainable AI—and how
major shifts:
to deliver it,” September 29, 2022, Liz Grennan,
Andreas Kremer, Alex Singla, and Peter Zipparo
1. Move from siloed work to interdisciplinary
collaboration. AI projects shouldn’t be limited
— “Why digital trust truly matters,” September 12,
to discrete pockets of organizations. Rather,
2022, Jim Boehm, Liz Grennan, Alex Singla, and
AI has the biggest impact when it’s employed
Kate Smaje
by cross-functional teams with a mix of skills
and perspectives, enabling AI to address broad
— “McKinsey Technology Trends Outlook 2023,”
business priorities.
July 20, 2023, Michael Chui, Mena Issler, Roger
Roberts, and Lareina Yee
2. Empower frontline data-based decision making.
AI has the potential to enable faster, better
— “An AI power play: Fueling the next wave of
decisions at all levels of an organization. But for
innovation in the energy sector,” May 12, 2022,
this to work, people at all levels need to trust the
Barry Boswell, Sean Buckley, Ben Elliott, Matias
algorithms’ suggestions and feel empowered to
Melero, and Micah Smith
make decisions. (Equally, people should be able
to override the algorithm or make suggestions
— “Scaling AI like a tech native: The CEO’s role,”
for improvement when necessary.)
October 13, 2021, Jacomo Corbo, David
Harvey, Nicolas Hohn, Kia Javanmardian, and
3. Adopt and bolster an agile mindset. The agile
Nayur Khan
test-and-learn mindset will help reframe
mistakes as sources of discovery, allaying the
fear of failure and speeding up development.

What is AI (artificial intelligence)? 9


— “What the draft European Union AI regulations — “Executive’s guide to developing AI at scale,”
Find more content like this on the
mean for business,” August 10, 2021, Misha October 28, 2020, Nayur Khan, Brian McCarthy,
McKinsey Insights App
Benjamin, Kevin Buehler, Rachel Dooley, and and Adi Pradhan
Peter Zipparo
— “An executive primer on artificial general
— “Winning with AI is a state of mind,” April 30, 2021, intelligence,” April 29, 2020, Federico Berruti,
Thomas Meakin, Jeremy Palmer, Valentina Sartori, Pieter Nel, and Rob Whiteman
and Jamie Vickers
— “The analytics academy: Bridging the gap
— “Breaking through data-architecture between human and artificial intelligence,”
Scan • Download • Personalize
gridlock to scale AI,” January 26, 2021, Sven McKinsey Quarterly, September 25, 2019, Solly
Blumberg, Jorge Machado, Henning Soller, Brown, Darshit Gandhi, Louise Herring, and
and Asin Tavakoli Ankur Puri

— “An executive’s guide to AI,” November 17, 2020, This article was updated in April 2024; it was
Michael Chui, Brian McCarthy, and originally published in April 2023.
Vishnu Kamalnath

Get to know and directly engage with senior McKinsey experts on AI


Sven Blumberg is a senior partner in McKinsey’s Düsseldorf office; Michael Chui is a partner at
the McKinsey Global Institute and is based in the Bay Area office, where Lareina Yee is a senior partner;
Kia Javanmardian is a senior partner in the Chicago office, where Alex Singla, the global leader of
QuantumBlack, AI by McKinsey, is also a senior partner; Kate Smaje and Alex Sukharevsky are senior
partners in the London office.

Designed by McKinsey Global Publishing


Copyright © 2024 McKinsey & Company. All rights reserved.

What is AI (artificial intelligence)? 10

You might also like