The transformative power of generative AI

Supercharged productivity or mass joblessness?

In brief
• Artificial intelligence (AI) technologies, particularly • AI will also have significant implications for labor.
generative AI, appear likely to revolutionize the way we Automating some tasks means needing fewer human
work, innovate and create. Generative AI can create novel, workers to produce the same output, which could result
human-like output across various domains, making it in transitional job displacement, put downward pressure
highly versatile and intuitive – it even helped us draft parts on wages and increase income inequality. However, if AI
of this paper – and as such, it has the potential to become technologies stimulate demand, the creation of new jobs
a “general-purpose technology” like the steam engine and higher overall economic growth should offset job
and computer, transforming the global economy. displacement.

• The most positive economic implication of AI disruption • AI may fundamentally change the way we, as humans,
will likely be accelerating labor productivity after many drive value in the workplace, requiring us to focus on the
years of near stagnation; estimates of the potential skills where we have a comparative advantage. These
impact span a wide range, though most analyses posit changes may be rapid and unpredictable, increasing the
1.5%-3.0% per year globally over the next decade. A boost importance of career flexibility, re-training and effective
to labor productivity should result in a similar boost to action from governments.
real GDP.
• For markets, AI-driven productivity gains are likely to be
• We expect a large share of AI’s productivity impact to positive for corporate earnings and equity returns;
come from automating many tasks humans currently do, implications for bonds are more ambiguous, though we
helping to offset increased retirements, but the potential think the most likely impact is modestly higher yields.
to accelerate innovation could make productivity gains
• We remain humble in our projections of the economic and
even more significant.
market implications of AI technologies, given tremendous
uncertainty over how powerful and capable they can
become, what kinds of unforeseen innovations and
industry transformations they’ll cause and, ultimately,
how governments and society will respond.
The transformative potential of generative AI
Artificial intelligence (AI) – or the process of making Generative AI is the latest stride in AI development, and
machines smart – has existed in some form since the by contrast, its key ability is to generate novel, open-
1950s, making occasional headlines, like when the ended content. The recent launch of several generative
chess-playing AI DeepBlue beat Gerry Kasparov in 1997. AI applications (Exhibit 1) has brought this technology
In recent years, AI has quietly become more prevalent to the fingertips of the masses and captured global
in our day-to-day lives, predicting arrival times of our attention. The most popular of these applications take
online delivery orders, populating our social media the form of chatbots, like ChatGPT, powered by large
feeds with personalized ads and filtering spam from our language models (LLMs) that string together words
email inboxes. Such applications of “traditional” AI (also based on patterns in vast troves of text data, such as
known as “narrow” or “weak” AI) can be very advanced, significant slices of the internet. By training related
even exceeding human expert levels, but they are models on other forms of unlabeled, unstructured data,1
trained to perform only in specific domains. ranging from photographs to the entire bodies of work

Exhibit 1: Generative AI tools can revolutionize the way we create and interpret diverse forms of data

WRITER CODER VISUAL ARTIST MATHEMATICIAN

Applications • OpenAI GPT-3* (2020) and GPT-4* (2023) which power ChatGPT • OpenAI DALL-E (2021) • Google Minerva* (2022)
• Google PaLM* (2022) and PaLM 2* (2023) • Midjourney (2022)
• Anthropic Claud 2* (2023) • Stable Diffusion (2022)

Capabilities Writes novel, high-quality Writes functional code in Creates high-quality Solves complex numerical
responses to prompts that various languages from a images—photorealistic or problems at the college level
are often indistinguishable specification. For existing artistic—based on written in subjects including
from human writing. It can code, it can explain, debug, descriptions. Early progress algebra, physics, number
write a college essay, explain and analyze (e.g., calculating has also been made on theory and machine
a joke, summarize a book, or time complexity). generating video. learning.
help draft an email.

Describe what generative AI Write a Python function Give me an image of a Assume that variance of the
is in four bullets that rhyme. that takes as input a file Pomeranian sitting on a first n natural numbers is 10
path to an image, loads throne, wearing a crown, and the variance of first m
the image into memory as with two tiger soldiers by even natural numbers is 16.
a numpy array, then crops his side. Compute m+n.
the rows and columns
around the perimeter if
they are darker than
a threshold value…

• Generating content, Variance of first n n2 - 1

import numpy as np
a skill so prime,​ natural numbers is 12
import cv2
Variance of first m even 2 (m - 1)
2 2
• Imitating humans, def crop_dark_perimeter
natural numbers is 12
it’s quite sublime,​ (image_path, threshold=100):
n2 - 1 22(m2 - 1)
• From art to text, # load the image using OpenCV So, = 10 and = 16
12 12
it can create,​ image = cv2.imread(image_path)
Solving for n and m, we get n = 11
• Machine learning models, if image is None:
and m = 7.
it does relate ...
So, m + n = 18.

Source: Google, OpenAI, J.P. Morgan Asset Management. *GPT-3, GPT-4, PaLM, PaLM 2, Claud 2, and Minerva are all large language models (LLMs).
Note that all the responses here – including the image – are totally original AI creations.

Structured data is highly organized and made up mostly of tables with rows and columns that define their meaning, such as Excel spreadsheets.
1  

Unstructured data is everything else, such as the substantial contents of email messages, books, customer service recordings, images, memes and
PowerPoint presentations.

2 The transformative power of generative AI

of famous artists, generative AI technologies can create There is also plenty that generative AI technology –
high-quality content spanning a wide range of domains: or any form of AI – cannot yet do. Interacting with the
image, video, audio, text, computer code and even physical world is still one large obstacle. While robotics
entirely synthetic datasets. has made many impressive advances in recent years
(with the notable exception of driverless cars), these
Whereas traditional AI technologies might be able to
systems are designed to perform specific tasks and
identify photos with “bumble bees” in them, generative
typically require higher investment and maintenance
AI could produce a photorealistic image of a bumble bee
costs in proportion to their potential output. The robot
wearing a hat or write a children’s story about a bumble
prototype designed to make guacamole for Chipotle,
bee learning to fly. As whimsical as these examples
“Autocado,” may quicken the food assembly line, but it
sound, this difference matters because the open-ended
cannot also fill a customer’s cup or wipe down tables.
nature of these tasks represents so much of what
Even so, in an information age with lots of desk jobs,
we humans do for work (and tend to think ourselves
non-physical problems are a big part of what we do.
uniquely capable of doing).
Before generative AI, no other technology has arguably
Generative AI may still be in its infancy, but the
had as much potential to automate so much of our work.
technology has advanced to an extent that we can begin
Such potential has sparked both public excitement and
to imagine its transformative implications across the
fear – the excitement of ridding ourselves of mundane
global economy. If a program can author a fictional story
and time-consuming tasks through automation, but
about a bumble bee, then it could write a movie script
also the fear of losing our jobs and livelihood. In this
– or at least help automate a big part of the process – a
publication, we seek to answer a few key questions: Is
real concern of screenwriters in the United States today.
generative AI truly transformational? Could generative AI
To be sure, screenwriters account for a vanishingly
become the next “general-purpose technology,” like the
small share of U.S. jobs, but generative AI can also help
steam engine and computer? What does broad-scale
software engineers write and debug computer code,
automation mean for labor markets? And, if AI can make
lawyers research legal opinions and draft contracts or
us all a lot more productive, what impact will that have
scientists read and summarize dense research papers.
on the economy, inflation and financial markets?
Visual and auditory generative AI technologies might
likewise automate tasks for jobs ranging from graphic
designers to video editors. While generative AI is named More like a steam engine than a smartphone,
for its functional differentiation from traditional AI, we economically speaking
think the more economically significant distinction is
Generative AI will have broad implications for the
how general it is.2 When we start to tally what generative
economy, but its most significant may be accelerating
AI could do across the whole economy (Exhibit 2), the
labor productivity after many years of near stagnation.
potential impact seems massive.
Labor productivity – total output per unit of labor input –
To be sure, the output of generative AI applications is has been the main driver of U.S. economic growth over
imperfect, with chatbots like ChatGPT occasionally the last century or more. However, productivity growth
even including “hallucinations” of false information. stagnated in the last decade, registering just 1.2% per
Many applications are therefore likely to require a layer year on average. The last 10 years have seen plenty of
of human supervision, especially where the costs of technological advancements that have improved many
mistakes are high, such as in medicine. However, further aspects of our lives, yet productivity statistics have told a
progress in developing generative AI might reduce some different story.
of these existing imperfections, and even supervised AI
could still significantly boost human workers’ output.

Generative AI has the potential to handle a broader, more general range of tasks compared to conventional AI, but it is still not artificial general intelligence
2  

(AGI). AGI, an unattained concept, refers to machine intelligence capable of performing any intellectual task humans can do.

J.P. Morgan Asset Management 3

Exhibit 2: Generative AI has the potential to accelerate efficiency, quality improvement and innovation across industries

Writing Software Engineering Health care

Everyday workers using generative AI Software developers can complete Generative AI can help nurse practitioners
chatbots can reduce time spent on manual and repetitive coding tasks up to in clinical processes and decision-making,
common writing tasks by as much as twice as fast when using generative AI enabling them to take on advanced
40% and enhance output quality.3 tools, increasing their ability to tackle new tasks from primary care physicians;6
Even economists can be 10-20% more and more complex challenges.5 AI chatbots have even been shown to
productive.4 outperform average doctors in answering
real patient questions.7

Architecture and Construction Marketing Education

Architects can rapidly output designs Marketers can leverage generative AI to Educators and students can employ
for new buildings subject to precise brainstorm creative ideas, including generative AI to develop tailored and
constraints, including optimal energy suggesting new brand names and logos interactive content and exercises for
usage. based on descriptions and criteria. personalized learning experiences.

Source: J.P. Morgan Asset Management.

This disconnect may be due to a few factors. Foremost, be due to mismeasurement in government statistics of
although advances like smartphones and online media the real value of new forms of software and human and
have made huge impacts on our daily lives, since they organizational capital.9
do so at a relatively low cost to consumers, they have
Generative AI, by contrast, may be the advancement that
limited impact on the market economy (various studies
finally ushers in a large, sustained boost to productivity.
show a willingness among many consumers to pay for
First, the broad-scale automation of existing activity
such services they receive for almost free).8 Additionally,
– producing similar outputs with less labor input –
by distracting workers and delivering “information
should, essentially by definition, result in a more directly
overload,” they may detract from productivity in other
measurable productivity impact.
activities. Some growth disappointment may also simply

3  
Using ChatGPT for mid-level professional writing tasks allowed workers to reduce time spent by 37% and improve output quality by 0.4 standard deviations.
See Noy, Shakked, and Whitney Zhang. “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Science 381, no. 6654, July 2023.
4  
Korinek estimates, based on 25 use cases for language models, that economists can be 10-20% more productive using large language models. See Korinek,
Anton. “Language Models and Cognitive Automation for Economic Research,” February 1, 2023.
5  
A McKinsey study found that developers using generative AI can increase task completion speed by 35-50% for lower complexity tasks and were 25-30% more
likely to complete higher complexity tasks with time savings. Similarly, research by GitHub found that 88% of surveyed developers felt more productive, 73% felt
more “in the flow” and 87% spent less mental effort on repetitive tasks when using the AI-powered GitHub Copilot. See Deniz et al. “Unleashing Developer
Productivity with Generative AI,” McKinsey & Company, June 27, 2023; and Kalliamvakou, Eirini. “Research: Quantifying GitHub Copilot’s Impact on Developer
Productivity and Happiness” The GitHub Blog, March 17, 2023.
6  
Louis Raymond et al., “Nurse Practitioners’ Involvement and Experience with AI-Based Health Technologies: A Systematic Review,” Applied Nursing Research
66, August 2022.
7  
In one study on patient questions randomly drawn from a social media forum, chatbot responses were preferred over physician responses and rated
significantly higher for both quality and empathy. See John W. Ayers et al., “Comparing Physician and Artificial Intelligence Chatbot Responses to Patient
Questions Posted to a Public Social Media Forum,” JAMA Internal Medicine 183, no. 6, April 2023.
8  
See, for instance, Brynjolfsson, Erik, Avinash Collis, and Felix Eggers, “Using Massive Online Choice Experiments to Measure Changes in Well-Being,”
Proceedings of the National Academy of Sciences of the United States of America 116, no. 15, March 26, 2019.
9  
Brynjolfsson, Erik, Daniel Rock, and Chad Syverson, “Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics,”
November 1, 2017.

4 The transformative power of generative AI

Even more profound implications for productivity, output Innovating innovation itself
and welfare gains could come if generative AI is the
Expanding on that last point, generative AI’s greatest
tipping point that enables AI to become the ‘general-
potential might not be in merely automating what
purpose technology’ of the 21st century.10 Like earlier
humans do, but in enhancing human efforts to create
general-purpose technologies, such as electricity, the
novel solutions to all sorts of real-world problems. Such
steam engine and the internet, generative AI could
efforts could lead to considerable productivity and
fundamentally change how a wide range of goods and
welfare gains beyond automation.
services are produced, transform industries and create
entirely new jobs, owing to its potential to: Simply making workers more efficient could perpetually
accelerate technological progress.13 For generative AI
• Be pervasive. Its general capabilities mean generative
specifically, further upside likely lies in its ability to:
AI can be integrated in many different contexts to
supplement or replace many activities currently done • Quickly sift through vast datasets. In a world of
by humans. “information overload,” generative AI can be a potent
filtering tool, automating many of the time-consuming
• Spawn complementary technologies and
tasks in research and development.
infrastructure. Companies across industries
are rushing to adopt AI in their fields, and the • Unlock new ideas and insights that inform researchers
development of ancillary business applications is of where to concentrate their efforts. Generative
necessary to fully leverage AI’s benefits. As we discuss AI can analyze vast troves of unstructured data,
later in this paper, generative AI may also enhance the something that is virtually impossible for humans to
performance of existing “traditional AI” technologies do, and in doing so can identify new patterns, reveal
and vice versa. insights and discover better ways of doing things.
• Experience exponential growth and economies of • Conduct comprehensive predictive and evaluative
scale. AI’s computing workload has been doubling analyses on new ideas. AI can improve the accuracy
every three to four months since 2012 and is likely to of our predictions and models, or even provide a
accelerate even further. OpenAI’s GPT-3 and GPT-411 sounding board for new ideas. Try prompting ChatGPT
were released just two years apart, and the latter is to list the pros and cons of your latent business idea,
significantly more complex, can interpret images for instance.
received as inputs, is 40% more accurate in its
responses and scores significantly higher percentiles Joining forces with “traditional” AI: greater
on many standardized tests.12 than the sum of its parts
• Reshape industries. Broad-scale automation will While generative AI technologies are currently in
reshape the nature of jobs and business models, with vogue, recent years have seen the proliferation and
transformative implications across industries. refinement of many “traditional” AI technologies that
• A
 ccelerate innovation. AI has the potential to have been trained to perform specific tasks very well.
accelerate research and development and unlock These tasks tend to be the kind where generative
new insights that inform and inspire innovation systems still fall short, particularly in the performance
efforts. Many leaders in the field think this may be AI’s of accurate predictive modeling, numerical calculations
paramount application. and optimization. McKinsey estimates that these
applications will account for a majority of the overall
potential economic value added from AI.14
10  
Bresnahan and Trajtenberg first coined the term “General Purpose Technology” (GPT) to describe technologies that drove new eras of technological
progress and growth. GPTs are characterized by pervasiveness, inherent potential for technical improvements and “innovational complementarities,”
giving rise to increasing returns-to-scale. See Bresnahan, Timothy F., and Manuel Trajtenberg, “General Purpose Technologies ‘Engines of Growth’?,”
Journal of Econometrics 65, no. 1, January 1, 1995.
11  
Open AI’s GPT-3 and GPT-4 refer to generative pre-trained transformers (not to be confused with “general-purpose technology”), which are the family of
neural network models that power generative AI applications like ChatGPT.
12  
GPT-4 has 170 trillion parameters compared to GPT-3’s 175 billion parameters, enabling many of the improvements in GPT-4 on processing and generating
text with greater accuracy and fluency. See OpenAI, “GPT-4” https://openai.com/research/gpt-4.
13  
Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, “Generative AI at Work”, NBER, April 2023.
14  
McKinsey & Company. “The Economic Potential of Generative AI: The Next Productivity Frontier,” June 14, 2023.

J.P. Morgan Asset Management 5

 Beyond generative AI, traditional AI is still delivering major solutions: some examples

Predicting the complex folding structure of proteins is one of the most exciting use cases of
non-generative AI. In the last 60 years, scientists have determined the structure of 180,000
proteins, a small number in proportion to the millions yet undiscovered. This arduous task is
an important part of drug discovery, but it can take years to execute. DeepMind’s AlphaFold is
now carrying out the same task in minutes with unprecedented accuracy,15 a milestone in the
application of AI to scientific research with immediate potential to advance drug development,
biological research and our understanding of diseases at a molecular level.

Environmental sustainability is another notable application of AI systems. AI systems are

increasingly helping optimize energy production, storage, distribution and use. In 2016, Google’s
DeepMind developed an AI framework which reduced energy usage for data center cooling by
40%.16 More recently, AI systems are aiding clean energy transitions. Whereas traditional weather
models fare poorly at predicting clouds, AI systems trained on satellite and weather data could
help solar grid and wind turbine operators optimize power generation and reduce fossil fuel
energy held as reserve. In the United Kingdom, Open Climate Fix is currently working with the
country’s electric grid operator to better forecast cloudy British weather.17

Source: J.P. Morgan Asset Management.

Integrating generative and traditional AI systems could But why stop there? Generative AI chatbots could draw
yield value far beyond what each alone could deliver, on the vast library of specialized traditional AI tools, from
since each has its own strengths and weaknesses. mathematics engines to commute time-forecasting
models, that have already been quite capable for over
No generative AI systems could achieve the accuracy
a decade – one by one, expanding their capabilities.
of AlphaFold’s predictions or estimate the exact
Already, OpenAI is privately testing several such
hours of sunshine tomorrow. Both abilities required
additions to ChatGPT.19
specialized training on structured datasets. A generative
AI chatbot like ChatGPT even struggles with some Some of these applications might be highly specialized.
simple quantitative reasoning.18 Ask it to multiply two For instance, Bloomberg’s approach to integrating
large numbers and it is likely to produce a close but generative AI into its terminal allows users to prompt a
incorrect answer. However, ChatGPT is fully capable system that is especially fluent in matters of finance,
of writing computer code to perform the very same tapping into decades of financial data collection and
calculation. Simply granting such chatbots access to development of specifically trained models that tackle
code interpreters might be one way to supply the correct matters of financial complexity.20 Indeed, we often hear
answer – not unlike calculators help humans solve now that “English will be the coding language of the
math problems that most of us couldn’t solve in our future,” and it seems likely to be in many cases.
own heads.

“AlphaFold,” DeepMind, accessed August 15, 2023, https://www.deepmind.com/research/highlighted-research/alphafold.

15 

“DeepMind AI Reduces Google Data Centre Cooling Bill by 40%,” Google DeepMind, accessed August 17, 2023, https://www.deepmind.com/blog/deepmind-
16 

ai-reduces-google-data-centre-cooling-bill-by-40.
“Sun in Their AIs: Nonprofit Forecasts Solar Energy for UK Grid,” NVIDIA, https://resources.nvidia.com/en-us-energy-utilities/ai-forecasts-solar-e?xs=363482.
17 

Even Google’s Minerva, a language model that achieves improved performance by gathering training data primarily from scientific papers, still makes simple
18 

mistakes with high frequency. See “Minerva: Solving Quantitative Reasoning Problems with Language Models,” Google Research (blog), June 30, 2022.
“ChatGPT Plugins,” OpenAI, March 23, 2023, https://openai.com/blog/chatgpt-plugins.
19 

Shaohua Wu et al., “BloombergGPT: A Large Language Model for Finance,” ArXiv (Cornell University), March 30, 2023.
20 

6 The transformative power of generative AI

Beyond written English, all these capabilities could A digitalized speed of adoption
be made more accessible by incorporating speech
AI’s implications for economic growth and societal
recognition and synthesis, areas where traditional
change can be profound, but the other factor to
AI already excels (along with handwriting and image
consider is timing. Although generative AI has suddenly
recognition; see Exhibit 3). We can imagine, for instance,
become dinner-table conversation, its ultimate power
verbally requesting a conference room smart assistant
and impact will not be seen for some time, though this
to draw a new logo idea and display it on a screen,
may take place faster than with earlier transformative
without the need for typing. In comparison to the
technologies.
current generation of “smart assistants” that rely on
users’ remembering pre-trained command phrases, Technological breakthroughs can take considerable
generative AI could make interacting with all these time to raise productivity, with the peak impact of
modules truly conversational experiences. Considering many industrial and post-industrial era technological
that we already spend an estimated 25% of our total breakthroughs, including general-purpose
work time communicating with one another, being able technologies, often only coming after 20-30 years. In
to communicate just as seamlessly with machines 1987, Robert Solow famously quipped, “You can see
opens the door to working alongside them. the computer age everywhere but in the productivity
statistics;” in that case, a significant impact did
ultimately show up in productivity statistics, albeit over a
decade later (Exhibit 4).

Exhibit 3: Traditional and generative AI capabilities are increasingly comparable to those of humans
Test scores of AI relative to human performance
Initial performance for each AI capability set to -100

20 AI systems
perform better
Human performance, as the benchmark, is set to 0 than humans
0
AI systems
Handwriting recognition perform worse

-20

-40
Reading
Comprehension

-60
Speech recognition

-80
Language
Image recognition Understanding

-100
1998 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020

Source: Douwe Kiela, Max Bartolo, Yixin Nie et al., “Dynabench: Rethinking Benchmarking in NLP,” Proceedings of the 2021 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Technologies, Association for Computational Linguistics, June 2021;
J.P. Morgan Asset Management. Published online at OurWorldInData.org. Licensed under CC-BY under the author Max Roser.

J.P. Morgan Asset Management 7

This lag owes to the considerable time it can take to: • Generative AI is very accessible and easy to use for
(1) establish wide-scale familiarity and access to the the average person, and its rapid accession to the
technology, (2) reshape business models to integrate mainstream is a testament to this fact. ChatGPT
the technology, (3) achieve a sufficiently large capital shattered records by amassing 100 million monthly
stock of it and (4) develop complementary innovations users in just three months, compared to the time it
and infrastructure that allow for full benefits of the took TikTok (nine months) and Instagram (two and a
technology to be harnessed. half years) to reach the same milestone.21
• Generative AI is decades in the making, with
Exhibit 4: Past general-purpose technologies have taken
considerable progress already made. While much
considerable time to deliver gains in labor productivity,
which has been the main engine of U.S. GDP over the of the public hadn’t heard about generative AI until
last century this year, its most notable underlying innovations
were developed in 2014.22 Meanwhile, decades
Labor productivity growth
Rolling 10-year annualized rate of advancement in cloud infrastructure and an
explosion of data and computing power23 have helped
4.5%
Development of Personal computer train these systems.
4.0% electric motor: invented: 1981
~1890
• Massive business investment has already been
3.5% made... In the five years ending in 2021, global
3.0% business investment in all types of AI grew more than
sixfold in real terms, with the United States leading
2.5%
the pack at $73bn invested.24 From 2017 to 2022, the
2.0% share of businesses that have adopted AI, and the
1.5%
number of AI capabilities used, more than doubled.25
Moreover, compared to some earlier technologies,
1.0% generative AI infrastructure and service providers
0.5% are bearing a larger share of the necessary capital
Resulting
productivity boom investments potentially increasing adoption rates by
0.0%
1885 1900 1915 1930 1945 1960 1975 1990 2005 2020 lowering the financial barrier for end users.
• ...and more is underway to integrate AI or develop
Source: BLS, NBER, J.P. Morgan Asset Management. Data from 1888-1957
reflect productivity data for the total private economy from John Kendrick, applications for business use. Company management
“Productivity Trends in the United States,” NBER. Data from 1958-2022 teams are increasingly focused on AI, with 40% of
reflect non-farm productivity data from the BLS. Data are as of August 31,
2023. S&P 500 management teams mentioning AI in their
2Q 2023 earnings calls, up from 19% a year earlier
We think that AI adoption could be faster. Over time, an (Exhibit 6). These businesses are rushing to develop AI
increasingly digitized world has helped accelerate the infrastructure and applications across a wide range
pace of technological adoption (Exhibit 5), and there are of domains, with many launched this year or in the
some reasons to believe generative AI could be adopted development pipeline.26 The surge in investor interest
faster still: has propelled hefty gains for stocks; the global AI
market is currently valued at $150bn and projected to
grow to $1.3tn by 2030.27

21 
Statista, Reuters.
22 
The introduction of generative adversarial networks (GANs) in 2014 marked a breakthrough in generative AI. GANs quickly became one of the most
influential generative AI models, allowing for the high-quality generation of images, audio, text and other types of content.
23 
The amount of computing power used to train AI systems has been doubling every six months over the past decade. See Jaime Sevilla et al., “Compute
Trends across Three Eras of Machine Learning,” ArXiv (Cornell University), February 11, 2022.
24 
Netbase Quid Companies dataset, 2022, as cited in “The 2023 AI Index Report,” Stanford Institute for Human-Centered Artificial Intelligence, 2023.
25 
McKinsey & Company. “The State of AI in 2022—and a Half Decade in Review,” December 6, 2022.
26 
Sequoia Capital’s “AI 50 2023” identifies the emerging trends in privately held AI companies, stemming across large-language models, infrastructure for
model training, generative AI applications and predictive AI applications.
27 
“Artificial Intelligence Market Size & Trends, Growth Analysis, Forecast,” MarketsandMarkets, June 2023.

8 The transformative power of generative AI

 Exhibit 5: Technological adoption has accelerated over time
U.S. Technology, Rate of Adoption*

80
Flush Toilet
70

60
Years to 75% adoption rate

Telephone Air Travel

50
ICE Automobile
Vacuum Cleaner Home Air Conditioning
40

30 Electric Power
Microcomputer (PC)

20 Household Refrigerator Internet Social Media

Radio Microwave Tablet
Cellular Phone
10
Smartphone
0
1840 1860 1880 1900 1920 1940 1960 1980 2000 2020
First year commercially available

Source: Asymco, compiled from various sources with support of the Clayton Christensen Institute, J.P. Morgan Asset Management.
*Estimated from current adoption trends.

Exhibit 6: 40% of S&P 500 companies mentioned AI in Sizing the potential AI productivity gain
2Q 2023 earnings calls
AI appears well positioned to significantly boost labor
Share of S&P 500 companies mentioning AI in earnings calls
productivity,28 but by how much? In our own analysis, we
45% estimate annual productivity gains between 1.4% and
2Q23: 40% 2.7% per year across developed markets over 10 years.
40%
This estimate, if realized, would be comparable to past
35% periods of technologically driven surges in productivity
30% (as shown earlier in Exhibit 4).

25% Importantly, our estimates quantify the impact of

Mentions of AI
automation alone. Such productivity enhancement
20%
would be in addition to any other productivity growth,
15% such as the acceleration of innovation, which we believe
10%
presents significant upside potential. On the other hand,
Mentions of crypto 2Q23: 1% external factors could partly limit productivity gains, for
5% instance, if an ineffective or overly restrictive regulatory
0% response impedes AI development, or social resistance
’10 ’11 ’12 ’13 ’14 ’15 ’16 ’17 ’18 ’19 ’20 ’21 ’22 ’23 stymies adoption.
Source: J.P. Morgan Asset Management. Mentions of AI include the
keywords: artificial intelligence (AI), deep learning, machine learning,
chatbots and natural language processing. Mentions of crypto include the
keywords: cryptocurrency, Bitcoin, Ethereum, blockchain, stablecoin and
altcoin. Data are quarter-to-date for 2Q23 as of August 25, 2023.

28 
In this publication, we focus on labor productivity, which is real economic output divided by the total number of hours worked. However, total factor
productivity (TFP) is often considered a more direct proxy for technological progress since it measures the efficiency with which all economic inputs,
including capital investments, are used to produce output. TFP growth represents the portion of output growth that is achieved above the accumulation of
these inputs. Since the 1970s, TFP has also seen relatively modest gains – just 0.6% per year in the U.S., according to commonly cited estimates by the
Penn World Tables. However, TFP is more difficult to measure and, as a result, historical estimates are less available; additionally, labor productivity may be
the more relevant measure from the perspective of individual human workers.

J.P. Morgan Asset Management 9

In line with several recent studies that come to similar Exhibit 7: Sizing the potential AI productivity boost
conclusions, our approximation of the potential
productivity impact of AI-related automation takes a 1. Direct labor cost savings
task-level approach. We estimate the aggregate time
spent on many types of tasks across the whole economy 2. Increased output from more productive workers
and judge the share of each of these tasks that might 3. Composition effects from re-employment of
be automated (for more details, see “Appendix: Sizing displaced workers
AI productivity gains”). Ultimately, we find that traditional
and generative AI applications could potentially
automate 14% to 27% of current work activities in the = Total productivity impact
United States over the next 10 years (we would expect
similar results across other developed markets). The Source: J.P. Morgan Asset Management.

wide range of these estimates owes to considerable

uncertainty around our assumptions outlined below. Ultimately, the greatest uncertainties in estimating the
labor productivity boost involve the capabilities of AI
Automation can materialize in productivity gains itself. Our more conservative estimates of AI capabilities,
through three channels (Exhibit 7). The first and most assuming a broader range of tasks are immune to
straightforwardly positive channel is direct labor cost automation, would see productivity gains of just 0.4%-
savings from fewer workers being needed to produce 0.9% per year over 10 years; on the other hand, our
the same amount of output. Alternatively, instead of most optimistic assumptions would see gains of as
reducing headcounts, companies can produce even much as 7.7% per year. To be sure, we think this upside
more output by retaining their more productive workers. is extremely unlikely, but its possibility does illustrate
The combined size of these two productivity impacts the significance automation could have dependent on
should equal the total productivity-weighted share of how powerful and pervasive AI ultimately becomes.
tasks automated. Finally, there is a composition effect Similarly, if our baseline scenario for AI automation were
that accounts for changes in the productivity of workers to take 20 rather than 10 years to take hold, then annual
displaced from automation. Our projections assume productivity gains would be proportionally smaller, i.e.,
that this effect is zero, or that, on average, displaced 0.7%-1.4% per year.
workers are reemployed in new jobs where they are
equally as productive as they were in their former ones.
Our 2024 forthcoming Long-Term Capital Market
We believe this is a conservative assumption, given the
Assumptions (LTCMAs) will incorporate a small first
potential for AI to lower the barrier to entry of many jobs
step in accounting for the role of AI technologies,
and because AI seems likely to generate entirely new
particularly generative AI, in automating current
jobs with high productivity. However, if displaced workers
economic activity. Although we feel it is too early
are employed in jobs where they are less productive,
to fully embed the plausible productivity upsides
overall productivity benefits would be lower; if many are
estimated in this paper into our base-case
not reemployed or work less, then productivity benefits
long-run assumptions, we plan to re-evaluate
would not fully accrue to real GDP.
this impact in the coming years and size the
productivity gain depending on the progress made
on automation.

10 The transformative power of generative AI

Work in an age of AI automation
AI may bring on considerable productivity gains, but by Transforming existing jobs
doing so through automation, the idea that robots will
In the vast majority of cases, AI will augment but not
“take our jobs” is becoming a popular concern. Are we
entirely replace human capabilities. While AI’s abilities
in for mass unemployment?
are impressive, there are many domains where AI
We don’t think so – at least not in the foreseeable future technologies still fall short compared to humans or
– but the future of work will likely look quite different. benefit from a layer of human supervision and feedback.
As repetitive and time-consuming “grunt work”
becomes automated, workers can spend more of their
Automating tasks, not jobs
time on higher complexity tasks, meaningful critical-
AI seems unlikely to automate many entire jobs, but it thinking or creative endeavors. As such, automation can
does have significant potential to automate many of also provide humans with the opportunity to deepen
the tasks involved in those jobs, with most estimates their skills, thereby expanding their overall potential and
of aggregate task exposure to automation ranging even increasing happiness at work.31
from 20% to 30%.29 Such exposure will be broad-based
Consider a couple of instances where we already see
across industries; OpenAI estimates that LLMs could
this dynamic playing out.
affect 80% of the U.S. workforce in some form.30
• Financial “robo-advisers” can provide customized
The degree of exposure, however, varies considerably by
investment advice and algorithmic portfolio
job type. Highly exposed jobs include those responsible
management. A human financial adviser, though, is
for documentation and review in legal professions,
still needed to provide appropriate advice on financial
providing administrative support in businesses and
matters involving complexity, counsel against
customer service representatives. Given generative AI’s
impulsive trading behavior during market crashes or
advanced abilities to understand language and draw
bubbles and offer empathy in times of crisis. Combine
upon vast bodies of information, exposed tasks include
those that involve a degree of knowledge or expertise. a human financial adviser with the tools of a robo-
As such, higher-skilled jobs, such as STEM professionals adviser, and this new “team” has both the benefit of
and health care providers, are also exposed. At the advanced technology, as well as a layer of emotional
intelligence for when contextual understanding is
opposite end of the spectrum, where tasks seem the
necessary.
least exposed to automation, are jobs involving physical
work or where the human component is invaluable • Similarly, for software engineers, generative AI can
(i.e., construction workers and daycare providers). significantly reduce time spent on research and
trial and error, especially when working with a new
programming language or software framework. With
these time savings, one study found that developers
were 25-30% more likely to complete higher
complexity tasks within the same time limit.32

McKinsey estimates that generative AI and other technologies have the potential to automate 30% of current hours worked today in the United States by
29 

2030. Goldman Sachs estimates that roughly two-thirds of the U.S. workforce may be exposed to some degree of automation from AI, with up to one-
fourth fully substituted. The OECD estimates that 27% of jobs in major countries are at high risk of automation.
OpenAI, OpenResearch and the University of Pennsylvania estimate that 80% of the U.S. workforce would have at least 10% of their work tasks affected by
30 

LLMs, with 19% of workers seeing at least 50% of tasks impacted. See Tyna Eloundou et al., “GPTs Are GPTs: An Early Look at the Labor Market Impact
Potential of Large Language Models,” ArXiv (Cornell University), March 17, 2023.
Survey results from the OECD’s 2023 Employment Outlook found that 63% of workers in finance and manufacturing said using AI in the workplace
31 

improved their enjoyment in the job. Similarly, Noy and Zhang (2023) found that exposure to Chat-GPT among participants using the tool for writing tasks
was associated with a substantial increase in job satisfaction of 0.40 standard deviations.
Deniz et al., “Unleashing Developer Productivity with Generative AI.”
32 

J.P. Morgan Asset Management 11

This pivot will require workers to focus on skills where Artificial intelligence has some clear advantages,
humans have a comparative advantage (Exhibit 8). including in its ability to process information at a speed
Human intelligence can still better understand context impossible for humans to match. In these cases, leaning
and matters of complexity, apply intuition and employ on artificial intelligence can expand our overall potential.
emotional intelligence in social settings or where
cultural norms are relevant. Place a human professional Exhibit 8: Humans and AI should still have distinct
in a room full of potential new clients, and within minutes comparative advantages
she can adapt and respond to social cues across all five
senses, adjusting everything from her handshake to the
level of detail in her presentation. AI now outperforms average humans on:
Humans also still have some edge in conceptual
thinking. Each human mind is intrinsically capable
of developing or working with an infinite number of Handwriting, speech and imagine recognition*
abstract representations and models of the world
Reading comprehension, language understanding*
– from solving physics problems to deciding how to
structure an organization. No form of AI boasts such Breadth of knowledge
general abilities. Traditional AI applications can work Computational power
with some such abstract representations, albeit only for Speed
the relatively narrow set of instances for which they are
designed. Generative AI chatbots, despite producing
novel content across a wide variety of domains, do so
by essentially echoing patterns in the troves of text on
Humans still have the advantage on:
which they are trained. Since this process involves no
underlying conceptualization, it falls short in purely
conceptual domains like mathematics, as we noted
earlier. Conceptual reasoning
Depth of understanding
Perhaps also owing to this edge in conceptual
understanding, humans should maintain an edge in Complex problem-solving
artistic creativity. While AI might soon be able to create Emotional intelligence
a lot of music and visual art that sounds or looks like
Morality and ethics
what has already been produced, it seems less likely to
generate entirely new genres on its own. Creativity
Intuition
Where AI’s capabilities fall short, humans will be needed
to fill in the gaps, and the jobs of tomorrow will be Source: J.P. Morgan Asset Management. *See Exhibit 3.
increasingly focused in these areas. Naturally, these
are also the sorts of skills we will want to emphasize in
education and job training, but many tough questions
remain. Can everyone be a great problem-solver? How
do we evaluate and train such skills? Do we need to
foster creativity in classes, instead of teaching students
how to write code?

12 The transformative power of generative AI

Exhibit 9: Technological innovation has coincided with the creation of new jobs that account for the bulk of
employment growth
Employment, by new and pre-existing occupations
Millions, 2018 employment compared to 1940 occupations

Axis
50 250

45 225

40 200

35 175

30 150
60%
25 125

20 100

15 75

10 50
40%
5 25

0 0
s

les

ns

in

ls

)
s
g

hs
ice

ice

er
io

tio

tio
ice

na
m
in

cia
Sa

ct

l (r
ag
in

ta

uc

ad
rv

rv

sio
rv

tru
ni

ta
m

or

an
se

se
se

od

l&

es
ch

To
sp
ns
&

M
ng

al

Pr
th

ica

of
Te
rm

an
Co
on
al

ni

Pr
er
He
Fa

Tr
rs
ea

Cl
Pe
Cl

Employment in new work Employment in pre-existing work

Source: Autor, David, Caroline Chin, Anna M. Salomons, and Bryan Seegmiller. "The New Frontiers: The Origins and Content of New Work, 1940-2018."
No.w30389. National Bureau of Economic Research, 2022 (and subsequent calculations by the authors as of August 2023); J.P. Morgan Asset Management.

Creating new jobs in the fields, people also had more time to spend on art
and science. Today, in a similar vein, an estimated 72% of
Although automation means that some of us will level
workers are employed in occupations that did not exist
up our daily tasks, increasing productivity also means
in 1940, implying that over 87% of employment growth
businesses need less workers to create the same
over the last 80 years has come from the tech-driven
output, which could, in turn, result in job cuts.
creation of new jobs (Exhibit 9).
However, if AI ushers in an economic boom, there
This proliferation of new jobs is ultimately because
should be all sorts of new jobs to which we can pivot.
productivity stimulates consumer demand (Exhibit 10).
The long history of technological advancement has
The ability to produce more output with fewer inputs
been associated with the continued process of job
inherently reduces production costs which tends to
displacement and re-instatement that has supported
drive down consumer prices, enriching consumer
an economy at full employment. For instance, in the
wallets and enabling the consumption of all sorts of new
early 1800s, over 80% of the U.S. workforce was in
goods and services. Higher consumer demand then
agriculture,33 yet even after improvements in everything
stimulates business demand for workers in new jobs.
from mechanization to crop rotation dramatically
Indeed, the rise in computerization has been associated
reduced the need for farm labor, a subsequent economic
with broad employment growth and the birth of many
boom re-employed many of those formerly working in the
jobs in computer science, software engineering, graphic
fields. Displaced farmers moved to the cities to find jobs
design, social media marketing and more.
in the industrial sector, where they then contributed to an
explosion in associated commerce. With less time spent

“The Story of U.S. Agricultural Estimates.” U.S. Department of Agriculture, April 1969.
33 

J.P. Morgan Asset Management 13

Exhibit 10: Higher productivity ultimately drives demand for new jobs

Automating once- Consumer savings

Lower production Growing aggregate
human tasks reduces from lower prices
cost reduces prices of demand stimulates
the cost of labor in redirected to additional
goods and services demand for labor
production goods and services

Source: J.P. Morgan Asset Management.

Generative AI ought to similarly increase our purchasing Exhibit 11: The expansion of bank automated teller
power, ultimately stimulating growth in new occupations machines (ATMs) coincided with growth in human bank
– the web designers and app coders of tomorrow. AI, in teller employment
some instances, may perform the work of a thousand Total employment of bank tellers and ATMs, thousands
humans for the cost of one. Such economies of scale
seem likely to generate new business models we can 600
only begin to imagine, along with demands for humans
to complement and manage them. Despite the many 500

headlines about job cuts at the hands of AI, the World

Economic Forum found that 50% of global employers 400

expect AI to create job growth versus just 25% who

300
expect it to create job losses.34
We may even see labor demand increase in the same 200
sectors undergoing automation. One telling example
is the impact that the rise of automated teller 100
machines (ATMs) had on the employment of bank
tellers (Exhibit 11).35 Although ATMs automated the cash 0
’70 ’74 ’78 ’82 ’86 ’90 ’94 ’98 ’02 ’06 ’10
handling tasks bank tellers had been doing, the number
of bank tellers grew concurrently with the rise in ATMs Number of Employed Bank Tellers Number of ATMs Installed

for about a decade. How did this happen? ATMs allowed Source: Bank for International Settlements, Bessen (2016), Bureau of Labor
banks to operate branches more efficiently, lowering Statistics, Occupational Employment Survey, J.P. Morgan Asset Management.
operating costs and prompting banks to open many
more branches. Increased accessibility of banking
services also spurred greater consumer demand for
them – more people wanted to use ATMs, and with it,
opted for additional banking services. Ultimately, growth
in consumer banking drove greater demand for bank
tellers, even if those new branches were staffed with
fewer bank tellers per branch. The type of work bank
tellers did also changed. With cash handling tasks
mostly automated, bank tellers could focus more
on customer service and sales, with more of them
receiving skills training and college education than
they did in the past.

 orld Economic Forum, “The Future of Jobs Report 2023,” May 2023.
W
34 

James Bessen, “How Computer Automation Affects Occupations: Technology, Jobs, and Skills,” Boston Univ. School of Law, Law and Economics, no. 15–49
35 

(October 3, 2016).

14 The transformative power of generative AI

Easing labor shortages... could be especially helpful in the United States, where
the most acute labor shortages today include many
Generative AI may be coming at an opportune time for
occupations where AI technologies are likely to have a
the global economy when aging populations in most
significant impact, such as healthcare providers and
developed regions stand to meaningfully slow growth
skilled software engineers.
in the decades ahead. Globally, the ratio of young and
elderly people who are economically “dependent”
on those of working age will gradually increase. … And even allowing us to work less
This dependency ratio will rise especially quickly in With freed up time from automation, and machines
developed markets, where it has been climbing from doing more of the heavily lifting on driving the economy,
a near-bottom of 48% in 2010 and is projected to rise workers might also enjoy some more time for family, rest
to 82% by 2080.36 In other words, by 2080 there will be and fun.
nearly two consumers for every working-age person,
and each year until then, these workers will need to Indeed, over time and across countries today, rising
support about 0.5% more people. All else constant, productivity has coincided with fewer hours worked.
workers would face pressure to work longer hours or 150 years ago, workers in today’s richest countries
postpone retirements in order to support the same per used to work a lot, but average working hours declined
capita economic output.37 significantly in the wake of the Second Industrial
Revolution (Exhibit 12a). In 1930, after a period of
Against this backdrop, AI presents a major opportunity particularly large productivity gains, John Maynard
to counterbalance building labor shortages. Every job Keynes suggested that further advances in technology
automated by AI is also one more person who can retire and productivity might lead to a 15-hour workweek. In
without reducing overall economic output. This dynamic the decades since that prediction, though, declines

Exhibit 12a: Over time, people have worked less as productivity has risen
Average weekly working hours per worker

70

60

50

40

30

20
1870 1880 1890 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 2020

United States United Kingdom Germany France Japan

Source: Huberman & Minns (2007), Our World in Data, PWT 9.1 (2019). Published online at OurWorldInData.org. Retrieved from: ‘https://ourworldindata.org/
time-use’. Licensed under CC-BY by Esteban Ortiz-Ospina, Charlie Giattino and Max Roser.

T he United Nations defines the “dependency ratio” as the ratio of young population (under age 15) and elderly population (aged 65 and over) to the
36 

working-age population (aged 15 to 64). See “World Population Prospects” dataset, United Nations Department of Economic and Social Affairs Population
Division, 2022.
37 
These developed regions could also offset their demographic headwind by increasing net immigration from, or expanding trade deficits with younger,
less developed regions.

J.P. Morgan Asset Management 15

 Exhibit 12b: Today, people work less when they work in more productive economies
Weekly working hours vs. labor productivity

50

Cambodia
Myanmar
Bangladesh
Singapore
45
China
India South Africa Malaysia
Indonesia
Weekly working hours per worker

Taiwan
Peru Mexico
40 Vietnam Thailand Greece Poland
Russia South Korea
Colombia
Israel
Chile
Pakistan Turkey
United States
35
Italy Australia
Brazil Japan Canada
Argentina United Kingdom Spain Sweden
France
30 Ecuador
Uruguay Switzerland
Iceland
Netherlands
Germany Norway
Denmark
25
0 20 40 60 80 100
Labor productivity (GDP in USD per hour of work)

Asia Africa Europe North America Oceania South America

Source: Feenstra et al. (2015), Our World in Data, Penn World Table (2021), J.P. Morgan Asset Management. GDP is adjusted for inflation and for differences in the
cost of living between countries. Feenstra, R. C., Inklaar, R. and Timmer, M.P. (2015), “The Next Generation of the Penn World Table”. American Economic Review,
105(10), 3150-3182. Published online at OurWorldInData.org. Retrieved from: ‘https://ourworldindata.org/time-use’ [Online Resource] Licensed under CC-BY.

in working hours have leveled off for major developed Considerations for income inequality
economies despite further technological progress. It
The automation of routine and manual work, or the
may be that some cultures intrinsically value hard work,
potential to work less, are exciting prospects. For many,
while consumerism may keep us from ever feeling like
the cost could be increasing income inequality, which
we have enough. However, across countries today,
for many reasons has been rising across developed
the relationship between labor productivity and hours
markets in recent decades. In the United States, the
worked is consistently negative, suggesting we could
share of pre-tax national income accounted for by
still get closer to Keynes’ vision (Exhibit 12b).
the top 10% of earners has grown from 34% to 57%
For developed regions, we estimate that a hypothetical since 1951, leaving the bottom 50% with only 10% of
AI-driven 30% increase in labor productivity over the national income.38 A similar trend is seen in wealth
coming decade could drive a 5%-10% reduction in inequality, with the impressive growth in financial
average hours worked. However, in order to realize assets39 concentrating economic gains among those
this outcome, individual workers will also need to earn with the means to invest. Some argue that technological
enough income that they are able to give up potential advancement has played a significant role in these
working hours in exchange for leisure. That, in turn, trends, with one study estimating that automation
requires mitigating further pressure on income explains 50 to 70% of the increase in wage inequality
inequality. from 1980 to 2016.40

 re-tax national income represents total labor and capital income before taxes and excludes government transfers. Prior to 1976, income is defined as
P
38 

market income and excludes government transfers but includes capital gains and is sourced from “Income Inequality in the United States, 1913-1998” by
Thomas Piketty and Emmanuel Saez, updated to 2021. Top decile includes all families with annual income above $135,000. Data for 2022 are J.P. Morgan
Asset Management estimates utilizing data sourced from realtimeinequality.org.
The ratio of U.S. financial assets relative to nominal GDP has grown from 2.3x in the late 1970s to about 4.3x today.
39 

40 
Acemoglu and Restrepo argue that a significant portion of the rise in US wage inequality over the last four decades has been driven by automation (and to
a lesser extent offshoring) displacing certain workgroups from employment opportunities for which they had comparative advantage. See Acemoglu,
Daron, and Pascual Restrepo. “Tasks, Automation, and the Rise in US Wage Inequality,” NBER, June 1, 2021.

16 The transformative power of generative AI

A key pitfall of automation is that it can lead to the These skill-leveling effects might slightly offset inequality
concentration of gains in the holders of capital, at least among workers. Recent studies on the impact of
initially. If some part of what a worker does is replaced ChatGPT on customer service workers43 and on college-
with an AI program, then the owner of that AI capital will educated professional performing writing tasks44 found
receive the “wages” the worker used to earn. While this that the greatest productivity gains came from novice
may propel gains for technology companies, and their and low-skilled workers. Higher performers saw less
investors, this dynamic doesn’t bode well for labor’s benefit, perhaps because they were already delivering
share of income, particularly in an economy where results closer to their peak potential, while lower-skilled
worker bargaining power has already dwindled.41 Over workers were not only able to complete tasks faster but
time, these inequalities can fade, as cost savings from also perform tasks with greater complexity, “leveling up”
automation pass through to consumer prices and in their responsibilities. Indeed, advanced technologies
as new jobs emerge that reemploy displaced labor. are already enabling nurse practitioners to take on
However, while the U.S. economy has maintained full more tasks usually performed only by primary care
employment, there is some evidence that automation physicians.45
has been outpacing the creation of new tasks and jobs
The net impact of AI of inequality – the inequality-
in recent decades.42 If so, generative AI could complicate
increasing effects of benefitting those with
this challenge by further narrowing the set of skills that
complementary skills and the inequality-reducing
are uniquely human, increasingly including those of
effects of leveling the playing field among workers –
higher-skilled white-collar professionals who largely
will likely vary considerably by industry, and ultimately
escaped the effects of prior waves of automation.
depend on how it is developed and deployed.

Mixed effects among workers Effective policy management

Among workers, the greatest beneficiaries are likely
Importantly, the speed at which adoption is taking
to be those whose skills are complementary to AI,
place suggests governments, businesses and workers
rather than replaced by it. Those who work in such
will need to act swiftly to reshape education and skills
complementary roles already tend to earn relatively
training and implement fiscal policies to smooth the
more, and increased demand for their skills could add
transition for labor. Skills mismatches might be offset
to inequality among workers. Consider a hypothetical
by investing in education and reskilling programs to
customer service center that is made significantly more
ensure workers are keeping pace with the new skills
efficient by generative AI. Customer-facing workers may
demanded in an AI economy, while proper public
experience a direct productivity enhancement, but
safeguards will be needed in cases of job displacement.
efficiency gains mean needing fewer of them to produce
A greater concentration of wealth may call for further
the same output, leaving them exposed to replacement.
redistribution of economic income; reducing inequality,
By contrast, a manager of this center who effectively
moreover, should help drive the demand boost that
develops and maintains systems that integrate the work
creates new jobs and new incomes, promoting a
of humans and AI would be performing work that is
virtuous cycle that helps reduce inequality sustainably.
more complementary to AI.
If accompanied with the right policy approach, an
At the same time, generative AI technologies can level economic boom from AI automation should be a
the playing field between lower-skilled and higher- “win-win” that ensures all income levels stand to benefit
skilled workers, by “lending” expertise to those who lack for the foreseeable future.
it, without the need for formal training and investment.

 ccording to BLS data, the share of the U.S. workforce represented by unions has fallen by more than half since the early 1980s, amounting to just 11.3%
A
41 

in 2022.
Acemoglu and Restrepo find that automation corresponded to greater displacement effects and weaker reinstatement effects, or the acceleration of
42 

automation compared to the creation of new tasks, over the last three decades than the preceding decades. See K. Daron Acemoglu and Pascual
Restrepo, “Automation and New Tasks: How Technology Displaces and Reinstates Labor,” Journal of Economic Perspectives 33, no. 2, May 2019.
43 
Brynjolfsson, Li, and Raymond, “Generative AI at Work.”
44 
Noy and Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.”
45 
Raymond et al., “Nurse Practitioners’ Involvement and Experience with AI-Based Health Technologies: A Systematic Review.”

J.P. Morgan Asset Management 17

With considerable promise comes considerable risk
While the potential economic fruits of AI are bountiful, • Data privacy is also a major concern. AI systems
the technology brings with it several sociological and often require vast amounts of personal information
ethical concerns that we will need to navigate and to function effectively, raising concerns about the
confront in the coming years. Challenges include collection and storage of sensitive data without
both those intrinsic to the underlying technology and proper consent or security measures. AI technologies
its economic implications, as well as the potential for can also inadvertently facilitate cross-border
manipulation and misuse of powerful AI technologies by data transfers, resulting in potential violations of
bad actors: international data privacy laws. Few regulations exist
so far to ensure responsible use of both national and
• Social instability is one potential consequence of
cross-border data sharing.
rapid development and deployment of AI technologies
that could manifest in a few ways: However, AI is perhaps also uniquely positioned to
address some of the same societal challenges it could
– A
 I-generated disinformation like falsified photos
potentially worsen. Properly designed and trained,
and videos intended to deceive, otherwise known
AI may prove more objective and less biased than
as “deepfakes,” might soon flood the internet.
human counterparts. We can imagine generative AI-
Such a development could make it even harder
powered tools that present information in ways that
for members of the public to discern and agree
help ideologically opposed individuals understand
on facts, ultimately amplifying ideological and
and relate to one another, and perhaps even serve as a
affective political polarization.
real-time mediator or fact-checker for online discourse,
– T he concentration of AI ownership among a few encouraging objectivity and even social stability.
large corporations or countries could likewise
concentrate power among them. At what point do
AI corporations become “too big” for the greater Many fears around AI focus on the potential for a
“doomsday” scenario that puts all of humanity at
good? Competition over such power could also
risk. As AI systems become more autonomous and
lead to an unmitigated “arms race” between
capable of making decisions, many experts forecast a
competing AI superpowers that might have significant risk that humans fail to contain a powerful
second-order consequences, including for the AI system that is not aligned with our values. This sort
careful assurance of safety of more powerful AI of risk, however, is a longer-term consideration that
technologies. depends on considerable further advancement in AI
technology that could be many decades away and is
– E
 conomic hardship from transitional
thus beyond the scope of this paper.
unemployment and increased inequality could
encourage further political extremism.
• AI bias is one key ethical concern, in that AI can
perpetuate and amplify existing biases present in
the data on which it is trained. As one old computer
science adage says, garbage in, garbage out. The
incidence of racial discrimination of facial recognition
technology has been studied extensively and images
produced by generative AI evidently amplify existing
stereotypes.46

Alexandra Sasha Luccioni et al., “Stable Bias: Analyzing Societal Representations in Diffusion Models,” ArXiv, March 20, 2023.
46 

18 The transformative power of generative AI

Key asset class implications
The broad implications of AI for global economies also In our view, performance so far does not nearly
leave much for investors to consider. resemble a “bubble.” Price multiples are not yet
significantly stretched as enthusiasm has also been
accompanied by strong upward revisions to earnings
Higher equity prices
outlooks for stocks with the most AI exposure, and that
For equities, if AI delivers on its promises, the implications multiple expansion has been relatively broad based, in
should be straightforwardly positive. Our work on Long- contrast to the narrow leadership of the dot-com bubble
Term Capital Market Assumptions suggests that an (Exhibit 13).
acceleration in potential GDP growth, all else the same,
The future potential, however, for AI to drive something
is likely to drive an acceleration in earnings by a similar
more like a bubble presents some upside risk
degree. Moreover, a greater share of national income
to equities. Historically, bubbles have commonly
flowing to capital – the owners of AI technologies – could
involved some new technology with no direct
give an added boost to equity returns.
historical comparisons, making the impact hard
While these impacts could take several years to to precisely quantify. When a consensus emerges
materialize, markets – and equities in particular – are that this technology is the “next big thing,” investors’
likely to price in AI optimism long before then. Indeed, imaginations tend to run wild. AI certainly seems
strong global equity performance in 2023 so far, capable of satisfying this criterion. On the other hand,
particularly in U.S. large caps, has been largely influenced most bubbles have also developed in periods of highly
by excitement around generative AI technologies. Such available credit, whereas today’s environment is one of
performance has inspired comparisons to the early restrictive monetary policy and tightening bank lending
2000s dot-com “bubble,” where enthusiasm about the standards, which should keep investors’ optimism at
internet propelled blind exuberance that drove stock least somewhat in check.
prices significantly above intrinsic values.

Exhibit 13: Rise in valuations has so far been broad based, unlike in the early 2000s dot-com bubble
12-month foward price-to-earnings (P/E) mutiple, S&P 500

26

24

22

20

18

16

14

12

10
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020 2022

Index Mean Index Median

Source: FactSet, J.P. Morgan Asset Management. Data as of August 2023.

J.P. Morgan Asset Management 19

AI excitement has already led to considerable gains for
technology companies in the S&P 500. Among individual
Conclusions
equities, Big Tech – now including chipmaker Nvidia – When we discuss AI with our clients, many of them are
has obvious exposure to AI and especially generative AI, more concerned than excited. Fear can be a good thing,
but as in past technology cycles, tomorrow’s winners but it’s easy to simply fear what we don’t understand, so
from AI may include relatively new players that are not part of our goal with this publication is to help identify
in vogue today. For our U.S. Equity Group’s thinking on what we do know and where the real challenges are
navigating AI investment opportunities, see “Artificial with this emerging technology. AI certainly does present
intelligence: Powering the next wave of technological many challenges, especially for labor, but an era of
innovation.” mass unemployment seems highly implausible. After all,
we wouldn’t bet against our ability as humans to always
find new ways to challenge ourselves.
Potentially higher yields
Managed properly, we do think AI has the potential
For government bonds, an AI productivity shock will
to make us all more productive, lower the real costs
likely contribute to modestly higher yields across
of many goods and services, reignite economic
developed markets. Faster growth in productivity and
growth and offset aging demographics. Beyond
thus real economic activity is likely to push cycle-neutral
economic growth, AI could also help solve some of
real yields higher by roughly the same degree, as has
our hardest societal challenges, such as in medicine
been the case over the long term (especially in the
and energy sustainability, and even accelerate the
United States since the middle of the 20th century).
pace of innovation itself. With all of this potential, AI
However, greater income inequality and downward
may prove to be the major transformative technology
pressure on wages could reduce inflationary pressures
of the 21st century, a rare occurrence that has
and thus breakeven inflation rates, partially offsetting
historically preceded significant industry and societal
upward pressure on nominal yields. One key point of
change. Generative AI, and its rapid accession to the
reference is the late 1990s, the most recent period of
mainstream, may be the tipping point.
strong productivity growth; over this period, nominal
yields did rise modestly – with the U.S. 10-year up by For investors, all of these outcomes provide significant
over 2% between late 1998 and early 2000, a period that multi-asset investment opportunities, but with
also roughly corresponded to the strongest NASDAQ generative AI still in its early innings, we would also
appreciation. emphasize the importance of humility and discretion.
It is in such uncertain environments, however, where
A similar dynamic may play out this time, but we are
we believe active management ultimately excels at
mindful of the possibility that greater inequality is met
identifying the winning companies of tomorrow, and the
with increased pressure to fund policies like universal
paths various asset classes may take along the way. As
basic income (UBI) that are not met with greater
the importance of and attention to AI continues to grow,
tax revenues. Such policies would result in greater
we are working hard at J.P. Morgan Asset Management
government debt issuance, which tends to result in
to develop and share our insights on the many questions
higher long-term yields.
surrounding this technological wave. As always, we
One consequence for monetary policy is that structurally welcome your feedback.
higher yields could reduce the need for unconventional
policies like quantitative easing, which have placed
downward pressure on longer-dated yields. An audio accompaniment of this paper is featured
on the “Insights Now” podcast series, available
on Spotify, Apple and Google Podcasts. The
episode is entitled “The economic implications of
generative AI.”

20 The transformative power of generative AI

Appendix: Sizing AI productivity gains
Our methodology Our assumptions on automation exposure include
both generative AI and “traditional” AI technologies but
In line with a number of peer estimates, we took a task-
exclude robotics. Our analysis only accounts for the
based approach to approximating the impact of AI
potential effects of AI automation; further sources of
automation based on the task dataset from O*NET. This
productivity gains could skew these results higher.
dataset provides many details on the task composition
of different jobs, including task difficulty and relative
importance, based on large-scale surveys of U.S. workers. Analytical limitations
We then multiply the O*NET task content for each job by The O*NET dataset on task content presents some
U.S. Bureau of Labor Statistics (BLS) data on the share inherent limitations and challenges. Most notably,
of this job in the economy, providing an estimate of the reported task difficulty corresponds to difficulty for
national aggregate share of each task, and from there, the humans rather than for AI, requiring our own estimates
total potential impact of automation on economic activity. on the latter, and the reported figures for relative
We identify 33 out of the 41 catalogued occupational “importance” reflect a subjective measure of the task’s
tasks as likely to be exposed to automation from AI respective importance to the overall job rather than time
technologies in some form over the next decade, spent. Presumably, these two metrics are related, but
leaving eight tasks immune (mainly due to physical- there are certainly cases where the tasks that take up the
world requirements, as we exclude robotics from our most time in a worker’s day aren’t the most value-added
assumptions). We then scored those tasks based on (i.e., so-called “grunt work”). Our sample captures 94% of
(1) the share of the task we expect to be potentially the labor force as measured by BLS, as some roles in the
automated (i.e., AI automates up to 80% of the BLS dataset did not exist in the O*NET database and were
tasks bucketed under “interpreting the meaning of therefore excluded from our analysis.
information for others”), (2) the skill intensity of the
We also note limitations in the extrapolation of automation
automation (i.e., interpreting Python code would have a estimates of specific O*NET tasks across all professions,
higher difficulty than interpreting customer feedback) since the same “task” can vary between occupations.
and (3) the impact of such automation scaled across the For example, our analysis found that fashion models,
entire economy (i.e., some tasks, such as “documenting a profession that should arguably be immune to
information,” are more prevalent in the economy than
automation, had over 30% exposure to automation (in
others, such as “repairing mechanical equipment”).
theory, image generating AI could probably automate
some fashion models, if societal norms permitted it). This
Key assumptions example underscores the difficulty in translating the “task
difficulty” measure provided by O*NET to the tasks AI
Our framework involves a few key assumptions:
could realistically automate.
1. AI automation is widespread across tasks and
industries with most impacts taking hold in the next
Sensitivities
10 years.
The above limitations result in estimation error in our
2. AI can automate up to a difficulty level of 3 on the 0-7
projections, and indeed, minor adjustments in the
O*NET difficulty scale (see Exhibit 15 for examples);
parameters of our analysis can lead to varied results. If
varying this figure produces quite a wide range of
we assumed AI technologies could only automate up to
alternative outcomes.
a level 2 on the seven-point scale employed by O*NET
3. Displaced workers are re-employed in jobs where and took a more conservative cut of tasks subject to
they are equally as productive as they were before, automation, with 15 tasks considered immune, estimated
translating to a “composition effect” of zero, as we productivity gains would then be cut to 0.4%-0.9% per
reviewed in the prior section “Sizing the productivity year. The most aggressive assumptions, assuming AI
impact.” automates all tasks up to a difficulty of 5, would see

J.P. Morgan Asset Management 21

productivity gains of up to 7.7% per year. This wide range Potential limitations to productivity upside
of outcomes underscores the high uncertainty behind
While the potential boost to productivity is significant,
any projections of the magnitude of productivity gains.
several factors may constrain its effects. The
Regardless, we do believe such exercises are worthwhile
composition effect hinges on the ability for workers
as we imagine the extent of potential automation
to pivot to new tasks and occupations where they are
across the economy, and the channels through which
similarly productive. An AI boom that does not coincide
productivity gains could materialize.
with the creation of new tasks and jobs would hinder
overall productivity benefits. In addition, AI technologies
Comparisons to other estimates are still very expensive to develop and integrate into
Many peer estimates on the potential productivity business use cases; if AI is not met with monetizable
gain are within the range of what we find plausible demand drivers, its uptake could be more limited.
(Exhibit 14). For instance, McKinsey came to similarly While regulation is necessary to ensure a sustainable
high estimates of 0.2%-3.3% per year through 2040 rollout of AI, overly restrictive or ineffective policies
including a broad set of AI technologies. This kind of could impede the development and deployment of AI
productivity gain would be a significant feat, translating technologies. Moreover, a broader resistance to change
to $2-$4tr annually added to the world economy from and technological limitations may further constrain
generative AI use cases alone. Brynjolfsson et al. found productivity gains.
that generative AI alone could raise productivity by 18%
over 10 years, or 1.8% per year, and acknowledge further
potential upside from the acceleration of innovation.47

Exhibit 14: Others’ estimates on the productivity gain from AI

Source Estimated annual AI productivity gain Region Period

Alderucci et al. 6.8%* United States next 5 years

Brynjolfsson et al. 1.8%** United States next 10 years

Goldman Sachs 1.5% (range: 0.3% to 2.9%) United States next 10 years

McKinsey 0.2% to 3.3%*** World 2023 to 2040

J.P. Morgan Asset Management 1.4% to 2.7% Developed markets next 10 years

Source: J.P. Morgan Asset Management; estimates compiled as of August 31, 2023.
* Alderucci et al. (2022) found that on a firm-level, manufacturing firms with AI-related inventions experienced a 7% increase in TFP, along with an 8.3%
increase in total revenue per employee and an 8.9% increase in value-added per employee. Across the economy, the impact of AI-related invention is
associated with a 6.8% increase in revenue per employee in the following 5 years.
** Brynjolfsson et al. (2023) estimate that generative AI will raise productivity by an added 18% over ten years (or 1.8% per year) above the current
Congressional Budget Office (CBO) projection of 1.5% productivity growth.
*** McKinsey (2023) estimates that generative AI alone could enable labor productivity growth of 0.1% to 0.6% annually through 2040. Combined with a
broad set of other AI technologies, work automation could add 0.2% to 3.3% points annually to productivity growth.

 rookings. “Machines of Mind: The Case for an AI-Powered Productivity Boom,” May 10, 2023. https://www.brookings.edu/articles/machines-of-mind-
B
47 

the-case-for-an-ai-powered-productivity-boom/.

22 The transformative power of generative AI

Exhibit 15: A task-based approach to sizing the productivity impact of AI technologies

Task difficulty (1-7 scale) 2 4 6

Getting information Follow a standard blueprint Review a budget Study international tax laws

Monitoring processes, Check to see if baking bread Check the status of a patient
Test electrical circuits
materials or surroundings is done in critical medical care

Identifying objects, actions Test an automobile Judge the suitability of food Determine the reaction of a
and events transmission products for an event virus to a new drug

Estimating the quantifiable Estimate the size of Estimate the time required to Estimate the amount of
characteristics of products, household furniture to be evacuate a city in the event of natural resources that lie
events or information shipped a major disaster beneath the world’s oceans

Calculate the costs for Calculate the adjustments for Compile data for a complex
Processing information
shipping packages insurance claims scientific report

Evaluating information to Evaluate a complicated

Review forms for Make a ruling in court on a
determine compliance with insurance claim for
completeness complicated motion
standards compliance with policy terms

Analyze the cost of medical

Skim a short article to gather Determine the interest cost
Analyzing data or information care services for all hospitals
the main point to finance a new building
in the country

Keep current on changes in Learn information related

Updating and using relevant Keep up with price changes
maintenance procedures for to a complex and rapidly
knowledge in a small retail store
repairing sports cars changing technology

Make appointments Prepare the work schedule Schedule a complex

Scheduling work and
for patients using a for salesclerks in a large conference program with
activities
predetermined schedule retail store multiple, parallel sessions

Organize a work schedule Plan and adjust a personal

Organizing, planning and Prioritize and plan multiple
that is repetitive and easy to to-do list according to
prioritizing work tasks several months ahead
plan changing demands

Record the weight of a Maintain information about

Documenting/recording Document the results of a
patient during a routine the use of satellites for
information crime scene investigation
health exam industry communications

Interpret a complex
Interpreting the meaning of Interpret a blood pressure Interpret how foreign tax law
experiment in physics for
information for others reading applies to U.S. exports
general audiences

Serve as the benefits director

Performing administrative Complete tax forms for a
Complete routine paperwork for a large computer sales
activities small business
organization

Source: O*NET, J.P. Morgan Asset Management. Table is for illustrative purposes.

J.P. Morgan Asset Management 23

Authors

Michael Albrecht Stephanie Aliaga

Global Strategist, Executive Director, Research Analyst, Senior Associate,
Multi-Asset Solutions Global Market Insights

24 The transformative power of generative AI

Acknowledgements
We would like to express our gratitude to several insights on AI advances beyond generative AI, corporate
individuals across J.P. Morgan Asset Management who applications of AI and equity investment opportunities
generously contributed their time and expertise to the within AI: Robert Bowman, a semis and technology
development of this paper – through their insights and hardware analyst; Timothy Parton, a portfolio manager,
constructive feedback, help with research and broad and James Connors and Doug Stewart, both investment
support for our efforts. specialists. We are also grateful to Dillon Edwards, an AI
strategist in our Intelligent Digital Solutions Team, for his
For his sponsorship, we would first like to thank Dr. David
partnership and feedback on technology specific issues,
Kelly, Chief Global Strategist and Head of the Global
and to our Data Science team for supplying bespoke
Market Insights Strategy Team, who was the first to
data on mentions of AI in earnings calls. Finally, from
support our foray into the subject as part of our Long-
Market Insights, Tai Hui, Asia Chief Market Strategist, and
Term Capital Market Assumptions, and for providing
David Lebovitz, Global Market Strategist, and from Multi-
feedback on multiple iterations of this publication,
Asset Solutions, Sylvia Sheng and Thushka Maharaj, both
especially on economic implications.
Global Strategists. All of these colleagues gave useful
We are likewise grateful to Jed Laskowitz, Chief feedback in their respective areas of expertise.
Investment Officer and Global Head of Asset
Last though certainly not least, we would like to thank
Management Solutions, and John Bilton, Head of Global
direct contributors to this paper, foremost from Michael
Multi-Asset Strategy, who both supported our efforts,
Di Pentima, a portfolio management associate in Multi-
provided constructive feedback and helpfully suggested
Asset Solutions, who helped develop calculations
others whose inputs proved invaluable.
to quantify the economic impact of AI automation.
Special thanks also go to Michael Cembalest, Chairman Additionally, he, as well as Brandon Hall and Mary
of Market and Investment Strategy, who offered Park Durham, research analysts in Market Insights,
particularly extensive and detailed feedback on our and Ayesha Khalid, a Global Strategist in Multi-Asset
paper – given his remarkable depth of knowledge in Solutions, helped us collect and organize various exhibits
AI, among many other fields. Michael also brought our and data used in this publication. This publication,
attention to additional relevant research that we have and all of our published content, could not have been
since incorporated into ours. possible without the incredible support and assistance
We are also grateful for feedback from Jeffrey Geller, a from our marketing partners, namely Maria Bernice
Chief Investment Officer in Multi-Asset Solutions; Jared Halaguena, a marketing strategy associate, and Haley
Gross, Head of Institutional Portfolio Strategy; Evan Baron, marketing strategy lead.
Grace, Head of Multi-Asset Portfolio Management for the This paper is a joint effort between Market Insights and
International Private Bank; and Anthony Werley, Chief Multi-Asset Solutions with input and support from across
Investment Officer of the Outsourced CIO/Endowments the firm. We hope our insights help inform our clients’
and Foundations Team. We would like to thank several long-term investment planning and asset allocations.
members of the U.S. Equity Group, particularly for their

J.P. Morgan Asset Management 25

The transformative power of generative AI

J.P. Morgan Asset Management

277 Park Avenue I New York, NY 10172

The Market Insights program provides comprehensive data and commentary To the extent permitted by applicable law, we may record telephone calls and
on global markets without reference to products. Designed as a tool to help monitor electronic communications to comply with our legal and regulatory
clients understand the markets and support investment decision making, the obligations and internal policies. Personal data will be collected, stored and
program explores the implications of current economic data and changing processed by J.P. Morgan Asset Management in accordance with our privacy
market conditions. policies at https://am.jpmorgan.com/global/privacy.
For the purposes of MiFID II, the JPM Market Insights and Portfolio Insights This communication is issued by the following entities:
programs are marketing communications and are not in scope for any
In the United States, by J.P. Morgan Investment Management Inc. or J.P. Morgan
MiFID II/MiFIR requirements specifically related to investment research.
Alternative Asset Management, Inc., both regulated by the Securities and
Furthermore, the J.P. Morgan Asset Management Market Insights and
Exchange Commission; in Latin America, for intended recipients’ use
Portfolio Insights programs, as non-independent research, have not been
only, by local J.P. Morgan entities, as the case may be.; in Canada, for
prepared in accordance with legal requirements designed to promote the
institutional clients’ use only, by JPMorgan Asset Management (Canada)
independence of investment research, nor are they subject to any prohibition
Inc., which is a registered Portfolio Manager and Exempt Market Dealer in all
on dealing ahead of the dissemination of investment research.
Canadian provinces and territories except the Yukon and is also registered
This document is a general communication being provided for informational as an Investment Fund Manager in British Columbia, Ontario, Quebec
purposes only. It is educational in nature and not designed to be taken as and Newfoundland and Labrador. In the United Kingdom, by JPMorgan
advice or a recommendation for any specific investment product, strategy, Asset Management (UK) Limited, which is authorized and regulated by the
plan feature or other purpose in any jurisdiction, nor is it a commitment from Financial Conduct Authority; in other European jurisdictions, by JPMorgan
J.P. Morgan Asset Management or any of its subsidiaries to participate in Asset Management (Europe) S.à r.l. In Asia Pacific (“APAC”), by the following
any of the transactions mentioned herein. Any examples used are generic, issuing entities and in the respective jurisdictions in which they are primarily
hypothetical and for illustration purposes only. This material does not contain regulated: JPMorgan Asset Management (Asia Pacific) Limited, or JPMorgan
sufficient information to support an investment decision and it should not Funds (Asia) Limited, or JPMorgan Asset Management Real Assets (Asia)
be relied upon by you in evaluating the merits of investing in any securities Limited, each of which is regulated by the Securities and Futures Commission
or products. In addition, users should make an independent assessment of of Hong Kong; JPMorgan Asset Management (Singapore) Limited (Co. Reg. No.
the legal, regulatory, tax, credit and accounting implications and determine, 197601586K), which this advertisement or publication has not been reviewed
together with their own professional advisers, if any investment mentioned by the Monetary Authority of Singapore; JPMorgan Asset Management
herein is believed to be suitable to their personal goals. Investors should (Taiwan) Limited; JPMorgan Asset Management (Japan) Limited, which is a
ensure that they obtain all available relevant information before making member of the Investment Trusts Association, Japan, the Japan Investment
any investment. Any forecasts, figures, opinions or investment techniques Advisers Association, Type II Financial Instruments Firms Association and
and strategies set out are for information purposes only, based on certain the Japan Securities Dealers Association and is regulated by the Financial
assumptions and current market conditions and are subject to change Services Agency (registration number “Kanto Local Finance Bureau (Financial
without prior notice. All information presented herein is considered to be Instruments Firm) No. 330”); in Australia, to wholesale clients only as defined
accurate at the time of production, but no warranty of accuracy is given and in section 761A and 761G of the Corporations Act 2001 (Commonwealth), by
no liability in respect of any error or omission is accepted. It should be noted JPMorgan Asset Management (Australia) Limited (ABN 55143832080) (AFSL
that investment involves risks, the value of investments and the income 376919). For all other markets in APAC, to intended recipients only.
from them may fluctuate in accordance with market conditions and taxation
For U.S. only: If you are a person with a disability and need additional support
agreements and investors may not get back the full amount invested. Both
in viewing the material, please call us at 1-800-343-1113 for assistance.
past performance and yields are not reliable indicators of current and future
results. © 2023 JPMorgan Chase & Co. All rights reserved.
J.P. Morgan Asset Management is the brand name for the asset MI_MAS-AI bulletin
management business of JPMorgan Chase & Co. and its affiliates worldwide.
09nr231409193311