Produced in partnership with Financial services firms have

started to adopt generative

AI, but hurdles lie in their path
toward generating income from
the new technology.

Finding value in
generative AI for
financial services
2  MIT Technology Review Insights

Preface
“Finding value in generative AI for financial services” is an MIT Technology Review Insights
report developed in partnership with UBS Group. This report is based on six in-depth
interviews with senior executives and experts conducted in June to September 2023.
The report looks at the early impact of generative AI within the financial sector, where it
is starting to be applied, and the barriers that need to be overcome in the long run for its
successful deployment. Paul Kielstra was the author of the report, KweeChuan Yeo was the
editor, and Nicola Crepaldi was the publisher. The research is editorially independent and
the views expressed are those of MIT Technology Review Insights.

We would like to thank the following individuals for their time and insights:

Michael Briest, Head of European Technology Research, UBS

Jason Napier, Head of European Banks Research, UBS

John Mileham, Chief Technology Officer, Betterment

Chia Hock Lai, Co-Founder, Global Fintech Institute

Rama Cont, Chair of Mathematical Finance and Head of the Oxford Mathematical and
Computational Finance Group, Oxford University

Lito Villanueva, Chief Innovations Officer/Executive Vice President, Rizal Commercial

Banking Corporation (RCBC)
 MIT Technology Review Insights 3

CONTENTS

01 Executive summary.........................................................................4

02 The promise of generative AI...................................................6

Time for a more measured assessment...............................7
It’s not ‘magic’..........................................................................................7
The road ahead.....................................................................................9

03 Reality check: Recent deployments of...........................10

generative AI
Cost cuts for now; income generation.................................10
will have to wait
Eyeing higher-value work...............................................................11
Uses of generative AI in the finance sector......................12

04 Only half-speed ahead: Wariness about more..........15

extensive innovation
Great expectations...........................................................................15

05 Two general challenges for new..........................................16

technology adoption
Legacy technology............................................................................16
A tight talent market.........................................................................17

06 Tech-specific challenges and the.......................................19

regulatory hurdle
Importance of customization.....................................................19
Reliance, bias, and accountability............................................19
Intellectual property rights and hallucinations.............. 20
Regulatory risks of a new technology................................. 20

07 Conclusion: Valuable tool, but yet to.................................21

be fully disruptive
4  MIT Technology Review Insights

01 Executive
summary

W
ith tools such as ChatGPT, DALLE-2, and For businesses from every sector, the current challenge
CodeStarter, generative AI has captured is to separate the hype that accompanies any new
the public imagination in 2023. Unlike technology from the real and lasting value it may bring.
past technologies that have come and This is a pressing issue for firms in financial services.
gone—think metaverse—this latest one The industry’s already extensive—and growing—use
looks set to stay. OpenAI’s chatbot, ChatGPT, is perhaps of digital tools makes technology advances particularly
the best-known generative AI tool. It reached 100 million likely to affect the sector’s companies.
monthly active users in just two months after launch,
surpassing even TikTok and Instagram in adoption speed, According to a recent UBS report on the impact
becoming the fastest-growing consumer application in of generative AI, statistics from the U.S. Bureau of
history.1 Labor suggest that banks and insurance are among
the industries with the greatest proportion of their
Several innovations inherent within generative AI workforces exposed to potential automation.3 This
have obvious potential benefits for businesses. Its MIT Technology Review Insights report delves further
large language models (LLM) can learn from even into that possibility, examining the early impact of
bigger quantities of data than classical AI, including generative AI within the financial sector, where it is
incorporating information from unstructured inputs. starting to be applied, and the barriers that need
Users can interact with generative AI tools and receive to be overcome in the long run for its successful
responses in natural language. And finally, according to deployment.
user needs, generative AI can—as the name indicates—
generate a range of new outputs including text, The following are the report’s key learnings:
pictures, computer code, and data streams.
Corporate deployment of generative AI in financial
Such breakthroughs have led to high expectations. services is still largely nascent. The most active
According to a McKinsey report, generative AI could use cases revolve around cutting costs by freeing
add $2.6 trillion to $4.4 trillion annually in value to the employees from low-value, repetitive work. Companies
global economy.2 The banking industry was highlighted have begun deploying generative AI tools to automate
as among sectors that could see the biggest impact (as time-consuming, tedious jobs, which previously required
a percentage of their revenues) from generative AI. The humans to assess unstructured information. Employees
technology “could deliver value equal to an additional are thereby freed for more creative work, and in some
$200 billion to $340 billion annually if the use cases cases, the tools outperform people. The following are
were fully implemented,” says the report. common areas of deployment:
 MIT Technology Review Insights 5

• Customer service: Generative AI chatbots are As is the case with any new technology, talent with
already deployed to assist human customer service expertise specifically in generative AI is in short supply
agents and may soon directly advise customers on basic across the economy. For now, financial services
questions. An academic study found that they increase companies appear to be training staff rather than
customer service agent productivity and consumer bidding to recruit from a sparse specialist pool. That
happiness. said, the difficulty in finding AI talent is already starting
to ebb, a process that would mirror those seen with the
• Fraud prevention and risk management: The rise of cloud and other new technologies.
technology allows software to incorporate a wider and
richer set of data into risk and fraud detection. More difficult to overcome may be weaknesses in
the technology itself and regulatory hurdles to its
• Coding: Tools that can produce bespoke code for rollout for certain tasks. Financial services companies
specific tasks have begun to appear. In a controlled aiming to benefit from generative AI face several
experiment, an Australian bank found that these raised hurdles. First, general, off-the-shelf tools are unlikely to
programmer productivity by 46%. adequately perform complex, specific tasks, such as
portfolio analysis and selection. Companies will need
• Information analysis and summarization: to train their own models, a process that will require
Generative AI tools are already creating summaries of substantial time and investment.
business conversations—a previously human task. On a
larger scale, BloombergGPT provides the company Once such software is complete, its output may
another channel to monetize its archives by letting be problematic. The risks of bias and lack of
subscribers search them using natural language accountability in AI are well known. Finding ways to
questions. validate complex output from generative AI has yet to
see success. Any new tool has to be designed to avoid
There is extensive experimentation on potentially violating other actors’ intellectual property (IP) rights;
more disruptive tools, but signs of commercial and generative AI algorithms may act unpredictably—
deployment remain rare. Academics and banks are even illegally—especially when interacting with other
examining how generative AI could help in impactful ones. Finally, so-called “hallucinations,” or confident
areas including asset selection, improved simulations, wrong answers, are a worry with any use of generative
and better understanding of asset correlation and tail AI. These risks mean that financial services companies
risk—the probability that the asset performs far below must be cautious in how they deploy generative AI.
or far above its average past performance. So far,
however, a range of practical and regulatory challenges Financial regulators are certainly wary about how
are impeding their commercial use. these tools may be used. Companies are hoping
for guidance, but this will take time. Authorities
Legacy technology and talent shortages may slow acknowledge that they need to study the implications
adoption of generative AI tools, but only temporarily. of generative AI more, and historically they have rarely
Many financial services companies, especially approved tools before rollout. Meanwhile, the possible
large banks and insurers, still have substantial, benefits of any extensive use of generative AI need to
aging information technology and data structures, be weighed against regulatory dangers.
potentially unfit for the use of modern applications.
In recent years, however, the problem has eased with Adoption of generative AI is growing in the financial
widespread digitalization and may continue to do so. services sector, but it is so far limited to automation of
In addition, newer companies, notably fintechs, do not low-value activities where its use can be monitored by
face this issue. The burden of legacy technology will humans or would not pose unacceptable risks.
therefore likely diminish, a process that generative AI, if
anything, can accelerate.
6  MIT Technology Review Insights

02 The promise of
generative AI

T
o much of the public, the release of ChatGPT in
late 2022 was the first shot in a new “We tend to overestimate
technological revolution. Since then, interest in
generative AI has soared. For example, Google
the impact of any new
searches for the term have risen dramatically
in the past year (see Figure 1).
technology in the short
run and underestimate it
With this interest has come rapid adoption by users
across societies and economies. International Monetary in the long run.”
Fund (IMF) research shows that uptake of ChatGPT
alone—just the most prominent of many generative AI
tools—has been extremely rapid even by the standards Chia Hock Lai, Co-Founder, Global Fintech
of popular technology adoption (see Figure 2).4 Institute

Figure 1: The internet discovers a new interest

Searches for the term “generative AI” by Google users have surged since late 2022, reflecting rapidly growing
interest in the technology.

Interest over time

100

75

50
Oct. 15-21, 2023

Generative AI 80
25

0
Oct. 30, 2022 March 5, 2023 July 9, 2023

Note: Interest over time: A value of 100 is the peak popularity for the term. A value of 50 means that the term is half as popular.

Source: Compiled by MIT Technology Review Insights, based on data from Google Trends, 2023.
 MIT Technology Review Insights 7

Figure 2: The breakout technology of 2023

ChatGPT gained more than 100 million active users across the globe within a span of two months, a much
faster rate than that of other platforms.

Months to gain 100 million users

70

60

50

40

30

20

10

0
ChatGPT TikTok Uber Instagram Telegram Pinterest Facebook Spotify Twitter

Source: Compiled by MIT Technology Review Insights, based on data from “Generative Artificial Intelligence in Finance: Risk Considerations,” International
Monetary Fund, 2023.

As with every new technology, a degree of excessive Second, generative AI looks set to build on and
expectation—and fear—is inevitably driving the supplement existing technologies, rather than
discussion. Chia Hock Lai, co-founder of the Global necessarily replace them on a grand scale. As Chia
Fintech Institute, says he sees a lot of hype now: “We notes, “you definitely need a clear use case to start
tend to overestimate the impact of any new technology trying out any new technology.” Any such cases will
in the short run and underestimate it in the long run.” arise from what generative AI can do compared to
existing tools.
Time for a more measured assessment
Indeed, Gartner reported in August 2023 that These may be more limited than popular imagination
generative AI had peaked in its emerging technologies appreciates. Tools enabled by older versions of AI,
hype cycle, after which it would likely take two to five for example are now called “just software,” says John
years to become integrated into business productivity Mileham, chief technology officer at Betterment, a
tools.5 To an extent, the febrile current discussion robo-advisor that assists users in automated investing.
around generative AI, however, obscures some And such tools have already played a crucial role in
attributes that will shape its impact. the ongoing digitalization of financial services firms
and other companies. There’s little point in deploying
First, generative AI is a rapidly evolving field: it did not generative AI where less advanced technologies are
spring into existence with the release of ChatGPT doing the job as effectively and at low cost. Indeed,
and its development is far from complete. McKinsey some analysts project that, while generative AI will have
describes “a rush to throw money at all things generative a very large economic impact, it will be markedly less
AI” between 2017 and 2022. During those years, private than that of earlier iterations of AI.7
investments in the technology rose at an average annual
compound rate of 74%, far outstripping the equivalent It’s not ‘magic’
growth rate for AI as a whole of 29%.6 Similarly, tools and The use cases will arise from what generative AI
products incorporating generative AI have continued to can deliver that other tools cannot. The striking new
develop rapidly in recent months (see Figure 3). capacity of generative AI is best illustrated by a brief
8  MIT Technology Review Insights

Figure 3: Six months in the life of generative AI

Starting in late 2022, new iterations of generative AI technology have been released several times a month.

Timeline of major large language model (LLM) developments following Chat GPT’s launch

Nov. 2022 Dec. Jan. 2023 Feb. March April

1 Nov. 30, 2022: Open AI’s 4 Feb. 2, 2023: Amazon’s 7 March 7: Salesforce 10 March 6: Microsoft
ChatGPT, powered by multimodal-CoT model announces Einstein GPT announces the
GPT-3.5 (an improved incorporates (leveraging OpenAI’s integration of GPT-4 into
version of its 2020 GPT-3 “chain-of-thought models), the first its Office 365 suite,
release), becomes the prompting,” in which the generative AI technology potentially enabling
first widely used text- model explains its for customers broad productivity
generating product, reasoning, and relationship management. increases.
gaining a record 100 outperforms GPT-3.5 on
million users in 2 months. several benchmarks. 8 March 13: OpenAI 11 March 21: Google
releases GPT-4, which releases Bard, and AI
2 Dec. 12: Cohere releases 5 Feb. 24: As a smaller offers significant chatbot based on the
the first LLM that model, Meta’s LLaMA is improvements in LaMDA family of LLMs.
supports more than 100 more efficient to use than accuracy and
languages, making it some other models but hallucinations mitigation, 12 March 30: Bloomberg
available on its enterprise continues to perform well claiming 40% announces a LLM trained
AI platform. on some tasks compared improvement vs GPT-3.5. on financial data to
with other models. support natural-language
3 Dec. 26: LLMs such as 9 March 14: Anthropic tasks in the financial
Google’s Med-PaLM are 6 Feb 27: Microsoft introduces Claude, an AI industry.
trained for specific use introduces Kosmos-1, a assistant trained using a
cases and domains, such multimodal LLM that can method called 13 April 13: Amazon
as clinical knowledge. respond to image and “constitutional AI,” which announces Bedrock, the
audio prompts in addition aims to reduce the first fully managed
to natural language. likelihood of harmful service that makes
outputs. models available via API
from multiple providers in
addition to Amazon’s own
Source: Compiled by MIT Technology Review Insights, based on data from “What’s the future of generative AI? An early
Titan LLMs.
view in 15 charts,” McKinsey, 2023

comparison with older forms of AI. Put simply, older generation of new text, pictures, computer code, or
AI tools can train themselves on huge amounts of even large data sets.
structured data and can answer specific questions
asked in programming-like language. The potential value of this creativity is substantial but
should not be overestimated. For example, generative
Generative AI can learn from an even larger body of AI can generate a data set based on existing ones, says
information—including unstructured data—and it can Rama Cont, chair of mathematical finance and head of
create apparently novel content in response to natural the Oxford mathematical and computational finance
language questions. Older AI tools, for example, can group at Oxford University. However, while it may be
tell users if something is a cat. Generative AI can even “100 times larger, it won’t have more information,”
generate a new image of a cat. he explains. “It can extrapolate to certain situations,
provided you have similar data sets on which to train it.”
In short, generative AI’s strength is that it allows users
to ask questions in natural language and receive In the cat example, the picture will rely on existing
output that provides readily comprehensive answers knowledge of cats but will be unable to create something
in different formats based on a huge corpus of new. Such extrapolation of existing data “is a nice
information. This can involve, among other things, the feature” of generative AI, says Cont, “but not magic.”
 MIT Technology Review Insights 9

The road ahead

Generative AI holds out substantial opportunities
“While later there will be
for financial services companies. Boston Consulting
Group, for example, projects that the sector is one
other, probably more
of three with the highest potential for growth arising important, deployments,
from the technology in the near term. In particular, the
consultancy argues that generative AI “can generate a lot of the potential of AI
personalized investment recommendations, analyze
market data, and test different scenarios to propose appears really nascent at
new trading strategies.”8 Certainly, interest exists
for applying generative AI in all these areas. “The this stage.”
financial sector has been very quick to understand the
usefulness of these tools,” says Cont. Jason Napier, Head of European Banks
Research, UBS
With limited and at most partial exceptions, however,
the industry has not gone beyond experimentation in
such areas. Jason Napier, head of European banks
research at UBS, says that while “later there will be
other, probably more important, deployments, a lot of
the potential of AI appears really nascent at this stage.”

Rather than generative AI revolutionizing the industry

over the short to medium term, companies are focused
on using it to address a range of specific tasks that
typically involve the automation of low value-added
jobs that previously required humans to assess
unstructured information. Interest in doing more exists,
but the still experimental state of generative AI-
powered tools and, equally important, regulatory issues
stand in the way.
10 MIT Technology Review Insights

03 Reality check:
Recent deployments
of generative AI

T
o be useful, any new technology or tool must
solve a problem. So far, generative AI is being “We will leverage on
applied to a range of specific tasks or
problems, rather than for the wholesale
generative AI to change
design of operating models across entire
companies. This is the case across much of the
the anatomy of work,
economy. McKinsey estimates that roughly three-
quarters of the value created by generative AI overall will
including our internal
arise from just four areas: customer operations, processes, by
marketing and sales, software engineering, and research
and development.9 automating typically
Cost cuts for now; income generation will tedious tasks and
have to wait
Where financial services differ from other sectors is workflows.”
that generative AI is largely not being used to drive
revenue. According to UBS research, so far “the real Lito Villanueva, Chief Innovations Officer
value being seen is from cutting costs.” (see Figure 4). and Executive Vice President, RCBC

Figure 4: Expected impact of generative AI in the financial sector

UBS analysts say generative AI’s main opportunity for the financial sector is lower costs, as staff expenses
represent a relatively high portion of total costs.
Revenues Costs Competition
Banks Neutral Reduction Increase

Exchange & financial business services Increase Reduction Increase

Fintech & payments Neutral Reduction Increase

Insurance Reduction Reduction Increase

Real estate Increase Reduction Neutral

Wealth & asset managers Neutral Reduction Increase

Source: Compiled by MIT Technology Review Insights, based on data from “Will Generative AI deliver a generational transformation,” UBS, 2023
 MIT Technology Review Insights 11

Figure 5: General Impact of generative AI

Overall, UBS analysts see cost savings as the most likely outcome from advances in generative AI, based on
analysis of 31 sectors.

Real estate Leisure

Revenue potential Cost saving potential
Mining

Food retail Telecoms Pharma

Capital goods
General retail Semiconductors
Steel
Technology hardware
Luxury goods Medical devices

Utilities

Media Software

Autos
Oil & gas Chemicals Banks

Wealth & asset

managers
Stock exchanges Staples Fintech/
payments
IT services

Internet* Aerospace/defense Support services

Building/
construction

Transport
Insurance**

Competition risk

*Internet identified an increase in costs as well as greater revenue and competition potential.
**Insurance saw a negative risk to revenues as well as cost-saving potentials and risk of increased competition

Source: Compiled by MIT Technology Review Insights, based on data from “Will Generative AI deliver a generational transformation,” UBS, 2023

According to its assessments of a range of industries According to a Fortune article published in June 2023,
(see Figure 5), the main impact of generative AI lies in “so-called generative AI is already helping to speed up
cost cutting, rather than on income generation. This is mundane tasks known to crush the spirit of junior Wall
true in every part of the sector, including in banking, Street employees, hedge funds say—from reviewing
wealth management, insurance, and payments. reams of market research to writing basic code and
summarizing fund performance.”10
These savings, in most cases, arise from using
technology to free human employees from the need to Eyeing higher-value work
analyze large amounts of often unstructured data and to These innovations are not typically designed to reduce
produce output in response to directions given in natural employment levels so much as to give time for staff to
language. “We will leverage on generative AI to change focus on higher-value work. Mileham notes that even
the anatomy of work, including our internal processes, by the limited applications deployed at Betterment so
automating typically tedious tasks and workflows,” says far represent “an unmitigated boon” in the day-to-day
Lito Villanueva, chief innovations officer and executive operations of the firm. “There are efficiency benefits
vice president at Rizal Commercial Banking Corporation across the board. The value is real, and we get a lot out
(RCBC), a Philippines-based bank. RCBC is not alone. of it,” he says.
12 MIT Technology Review Insights

“As of yet, it’s not like you’re going to be putting a

generative AI directly in front of customers and
setting it free.”
John Mileham, Chief Technology Officer, Betterment

Another potential improvement is that generative our customer service team drastically,” Mileham says,
AI applications may sometimes outperform people even before the integration of generative AI. That step
when evaluating unstructured data. A University of is being considered carefully.
Chicago business school study, for example, found
that a generative AI tool could quickly and accurately The addition of generative AI elements to chatbots
process the implications of bloated financial disclosure brings obvious benefits. These include interaction
information, which might seek to hide the truth through in natural language and the possible integration of
verbiage.11 potentially huge repositories of unstructured data into
the tool’s knowledge base. Accordingly, Villanueva says
In contrast, academic research into the ability of a that a generative AI–based chatbot is one of RCBC’s
generative AI tool to help customer service operations leading digital priorities. That’s because it enables
found that it increased average productivity in terms of “real-time quality customer service interactions and
issue resolution per hour. However, it had “the greatest contributes to a seamless customer experience” by
impact on novice and low-skilled workers, and minimal facilitating the filing of complaints, accommodation of
impact on experienced and highly skilled workers.”12 In client requests, and collection of relevant customer data.
other words, it helped new employees get up to speed,
but was less useful to those who already knew the RCBC is not alone. In May 2023, tech company Kasisto
ropes. launched KAI-GPT, the first banking-specific LLM,
and the initial program built to use KAI-GPT was KAI
Uses of generative AI in the finance sector Answers, which aids employees in replying to customer
Financial services firms have focused generative AI queries.14 Meanwhile, various banks have announced
deployment on several specific tasks that share the development of generative AI-powered information
kind of attributes described above. tools for the direct use of customers or customer
service agents.15 Similarly, Cowbell Insurance has
Customer service: Chia speaks for many when he launched MooGPT to provide answers to its customers
notes “a very strong use case for generative AI in and agents.16 Most of these tools appear designed
customer service—going 24/7 for customer service, to improve customer-employee interactions rather
even to the extent of providing advice for customers.” than replacing them, except in cases where the aim is
Indeed, the academic research noted above on use providing basic information.
of the technology in customer service found that it
increased the number of issues solved per hour by 14%, This is intentional. Two considerations are at play.
improved client sentiment, and reduced requests for First, notes Mileham, any integration of generative AI
managerial intervention.13 into existing chatbots needs to be done safely. “As of
yet, it’s not like you’re going to be putting a generative
The most common customer service–related innovation AI directly in front of customers and setting it free,”
is the creation of chatbots for either direct use by he says. Second, some financial services firms have
customers or company service agents. These tools reported an increased need for human customer
represent not so much a new development as an service staff. “We’ve seen people wanting more human
evolution. At Betterment, for example, a chatbot based interactions for retail transactions instead of fewer,”
on predictive AI has already “reduced the workload on says Villanueva. As a result, the Philippines-based
 MIT Technology Review Insights 13

bank’s generative AI chatbots are being designed to Coding: The financial services industry is, in many
“allow our talents to maximize their time and move ways, a knowledge-based one. Michael Briest, head of
toward more highly specified customer concerns that European technology research at UBS, notes that “as a
require more detailed and sensitive attention,” he adds. sector, banks spend more as a percentage of revenues
on IT than any other, and a lot of banks still do software
Fraud prevention and risk management: For some design and maintenance in house.” Accordingly, ways
years, financial services companies have been using to improve the efficiency of such activities will interest
advanced technologies, including predictive AI, to companies. If properly trained, a generative AI tool can
improve risk management and fraud prevention. produce requested computer code as easily as others
Generative AI will allow the sector to go further, can answer questions or generate pictures.
including through greater integration of unstructured
data into these efforts. Chia says that using such A leading U.S. bank has started using generative AI
information will allow companies to identify new tools to help its code developers.19 It is not alone.
patterns and anomalies with associated risks at both Westpac, a large Australian bank, ran a trial with
a micro level—such as the potential for an individual to generative AI to assist its coders and found a 46%
default—and a broader one—such as market trends. increase in productivity against a control group.20

More specifically, according to UBS research, reducing Mileham says his company, Betterment, uses
fraud is key for the payments industry in general. generative AI software to help with debugging. They
Visa and PayPal have both deployed generative AI to have also procured GitHub Copilot, a cloud-based AI
prevent fraudulent transactions by blocking suspicious tool, to help with code generation and auto-completion.
ones.17 More generally, fintechs such as Datavisor, As with other uses, he stresses that this should happen
Feedzai, and Forter have all integrated generative AI only in the context of robust review and testing, and
into their off-the-shelf solutions to reduce payment a person taking ultimate responsibility for any new
fraud.18 At the same time, the new technology allows code. Even within these constraints, Mileham says it’s
the sector to go beyond reinforcing existing anti-fraud a worthwhile effort. “Everybody who has deployed it
defences, says Villanueva. He adds that RCBC is [at Betterment] spends less time banging their heads
hoping to use it to engage customers more actively in against the wall, not knowing the right answer [to
fraud prevention, through more targeted and effective relatively straightforward questions], and more time
awareness and education campaigns. being creative,” he explains.

“Definitely, there will be a lot of

automation of routine tasks like
report generation.”
Chia Hock Lai, Co-Founder, Global Fintech Institute
14 MIT Technology Review Insights

0
“Some fintech startups have started this model,
sharing or commercializing the generative model but
not the data.”
Rama Cont, Chair of Mathematical Finance and Head of the Oxford Mathematical and
Computational Finance Group, Oxford University

Diverse forms of information analysis and Such subscription tools, however, do raise questions
summarization: One of the strengths of generative around how generative AI will change thinking around
AI is its ability to use the information in its model to the role and value of data. Major news organizations
answer questions. The same capacity is already finding have taken steps, for example, to block ChatGPT’s
various uses within financial services. Mileham cites web crawler from accessing their websites.24 On the
the example of using generative AI to create the first other hand, Cont says generative models can now be
draft of summaries following a business meeting or shared without revealing the underlying data on which
phone conversation. This, he says, allows a minor, the model was trained. “Some fintech startups have
everyday task to “be compressed into a review rather started this model, sharing or commercializing the
than being a drafting exercise.” Similarly, at Man generative model but not the data,” he explains. “That
Group—a large hedge fund—managers have found is a new possibility.” Ultimately what data is commonly
that generative AI can speed up initial research by retained, what is shared, and how, remains a question
reviewing academic papers and spotting patterns.21 for markets—and regulators—to answer.
Meanwhile, a blue-chip Wall Street firm is rolling out
an app to act as a virtual assistant to help wealth
managers find client-relevant research or forms.22
Looking ahead, Chia expects such activity to grow,
saying, “Definitely, there will be a lot of automation of
routine tasks like report generation.”

Beyond this quotidian manipulation of information,

however, will be the opportunity for some businesses
to use generative AI to monetize data. One of the
most high-profile innovations using the technology
in financial services, BloombergGPT, falls into this
category. Subscribers could already access the
company’s large data archive, but this has now been
turned into a specialist LLM for answering questions
related to financial services.23
 MIT Technology Review Insights 15

04 Only half-speed ahead:

Wariness about more
extensive innovation

W
hile the buzz around generative AI has algorithms that beat passive asset selection.26 Risk
somewhat eased in the second half of management is another field where generative AI is
2023, optimism still surrounds its seeing use as an advanced research tool that goes
long-term impact. “Generative AI is an beyond its current rollout. Use cases here include trying
exciting development,” says Mileham. to better understand asset correlation27 and tail risk,
“There is a lot of opportunity.” among others.28

On the other hand, a distinct disconnect exists between While much of the abovementioned activity has
the current deployment of generative AI with its occurred in academia, some companies are
perceived potential. Today’s effort will have an effect, interested as well, at least in principle. UBS, for
but only on a limited number of functions. Moreover, the example, is researching the use of generative AI
innovation, so far, appears more to be improvements for trading applications. One compelling avenue is
to current practices than the kind of fintech-driven using the technology to express news, in the form of
disruption seen in payment services and wealth unstructured text, as a numerical vector in order to
management in recent years. assess its impact on asset prices.

Great expectations The company’s preliminary results show promise in

That hasn’t stopped other companies and researchers improving the ability to forecast volatility changes
from looking further. In Cont’s view, the financial sector driven by incoming news. Meanwhile, a blue-chip Wall
has been quick to adopt new technologies at an Street firm has applied for a trademark for what it
experimental level. “Their deployment could be very hopes will be a tool that will advise customers on stock
easy as long as they [companies] are comfortable with selection.29 In practice, however, Briest observes that
the output,” he says. the banking industry’s restrained approach reflects the
one being taken across the financial services industry
An area of particular interest is asset selection. One as a whole. “The sector is relatively conservative in
aspect of this is finding tools that balance portfolio- adopting new technological trends,” he says.
wide risk and return.25 Other research seeks to develop

Risk management is another field where generative AI

is seeing use as an advanced research tool that goes
beyond its current rollout.
16 MIT Technology Review Insights

05
W
Two general
challenges for new
technology adoption

hen grappling with the challenge of

adopting new technologies,
companies often have to tackle the
confluence of legacy technology and
a tight labor market.
According to an Accenture survey of large banks, even
though the respondent pool consisted of companies
interested in cloud usage, only 31% had moved more
than half of their previous mainframe activity to the new
platforms.31 “A lot of banks maintain old IT systems,” says
Briest. “We’re hearing from technology companies about
Legacy technology a lot of pilot projects starting and companies moving
Financial services companies, especially banks, were quite quickly to the next step, but this is going to take
among the early adopters of IT decades ago. Choices some time.” It’s an observation shared by Chia. “Most
made then, though, have long resisted further change. financial services organizations have a lot of data that is
The most striking example of this phenomenon is that, usually poorly structured or even fragmented,” he says.
as late as 2017, 43% of banking systems relied on
a six-decade-old computer programming language, Despite this enduring challenge of legacy IT for many
COBOL, which was also behind 80% of credit card companies, the problem has been diminishing across
transactions and 85% of ATM activity. the industry because of extensive digitalization in recent
years. “A lot of financial services firms have invested
Typically, COBOL drove large mainframe computers heavily in digital transformation,” says Chia. “Most have
because it was the only option decades ago. Although gained a certain capability in data management and
such arrangements have provided substantial stability, there’s already a level of fundamental readiness in terms
they make it difficult to add new capabilities arising of technology investment.”
from more recent technological developments.30
One of them is RCBC, which was established in 1960.
COBOL encapsulates the broader legacy-technology “The past three years have been pivotal for our digital
deficit in the financial sector. It’s a problem that transformation,” says Villanueva. “The introduction and
encompasses old software and siloed data storage expansion of generative AI solutions will be smooth
arrangements that have evolved to meet challenges and easy.” Meanwhile, new entrants do not have a
across decades but are no longer fit for purpose. technological deficit to overcome. Mileham says that

More generally, companies have a big opportunity to

use “generative AI to accelerate the shift off some
legacy applications that maybe it was just cost-
prohibitive to consider previously.”
Michael Briest, Head of European Technology Research, UBS
 MIT Technology Review Insights 17

Betterment, as a cloud-native company, can deploy Despite the current shortage in generative AI talent,
generative AI as broadly as it sees a use for it relatively some see this as an expected problem, common to
quickly. “I’m confident that major cloud providers are every industry. “I don’t see it as a long-term problem,”
going to be able to productize these capabilities and says Briest. He explains that the same talent shortage
expose them to companies very efficiently,” he says. has accompanied the appearance of other new
Cont also says that he believes that financial companies technologies, such as cloud, but that a supply of the
are, overall, “pretty ready” to make use of generative AI. necessary talent has developed. An April 2023 survey
indicates that for generative AI, finding talent is already
Even those who currently are not in such a state may growing easier (see Figure 6).
be able to use the technology to help modernize
existing IT infrastructure. Generative AI itself, Experts say that as the technology evolves over time,
notably its ability to generate code, can help with the finding talent could become less of a problem. First,
transformation away from legacy systems and data new entrants into the workforce will increasingly have
storage.32 More generally, says Briest, companies have been educated with the technology in mind. At the
a big opportunity to use “generative AI to accelerate the same time, while generative AI is new, it overlaps with
shift off some legacy applications that maybe it was just other fields of AI and machine learning.
cost-prohibitive to consider previously.”
The past can offer some lessons, says Cont. Consider,
A tight talent market for example, financial services professionals using
Another challenge around adoption of new technology earlier AI to conduct simulations for investment banks.
is a lack of talent and expertise. “Currently, generative “A lot of quants switched to data scientist roles,” he
AI is so new that you can’t really hire a whole lot of says. “They just changed their business cards. There’s
experience,” says Mileham. Villanueva agrees that not a shortage of people. The tech is new, but the math
“it is challenging to find talent because of the highly and computational foundations are not.”
competitive labor market.”

Figure 6: Hiring for AI-related roles

Responses in McKinsey’s survey suggest that hiring tech talent has become somewhat easier since 2022.

Share of respondents reporting difficulty in organizations’ hiring of AI-related roles %

Less difficult More difficult

Source: Compiled by MIT Technology Review Insights, based on data from “The state of AI in 2023: Generative AI’s breakout year,” McKinsey, 2023
18 MIT Technology Review Insights

0
Meanwhile, companies like RCBC are looking to
develop internally the skills needed to use generative “Our thought is to get the
AI tools. Villanueva says the bank’s approach
contributes to employee satisfaction with the new
technology into people’s
technology. “RCBC has its Digital Academy providing
the best-in-class and relevant trainings for its human
hands so that they can
resources,” he adds. start to become the
At Betterment, letting people develop their own skills experts.”
is designed to help both employees and the company.
“Our thought is to get the technology into people’s John Mileham, Chief Technology Officer,
hands so that they can start to become the experts,” Betterment
says Mileham.
 MIT Technology Review Insights 19

06 Tech-specific
challenges and the
regulatory hurdle

G
enerative AI applications appear The difficulties seen in market simulation software
impressive, but they are general-purpose illustrate some of the likely, wider challenges for
tools that do not address most of the generative AI. According to Cont, while generative AI
specific needs of financial services tools can create new data sets, they will not include
companies. “There are different use cases any new information if, for example, a company wants
within financial services that will need to use some better insight into tail risk. More broadly, he adds, users
proprietary data and not the very general kind of data can’t just hope that a generic algorithm will extract
that ChatGPT is using,” says Chia. Companies will need what is needed for a niche use. “If you don’t target
to be prepared for the fact that they’re not going to get what you want to learn, you may not learn it,” Cont
immediate results when they start investing in the explains. “You have to ask what you want to use output
technology, he warns: “It takes a long time to have a for and then create a bespoke algorithm tailored to the
high-quality model.” intended use case. When you generate cat pictures
from generative AI, you don’t want a mutant cat, but
Importance of customization in financial applications you are typically interested in
Companies like Betterment will likely need to create extreme events.”
distinct tools to address different uses. “We wouldn’t
create an ‘everything machine’ that has all of our Reliability, bias, and accountability
customer information in it for anybody to draw from,” Additionally, checking the reliability of the output
says Mileham. “That wouldn’t be a good practice for presents important challenges for activities that use
a serious financial institution.” Even once those use- substantial amounts of generated data. “There is no
specific models are in place, Chia says the work doesn’t easy way to validate the output,” says Cont. “If you
end there. “In financial services, you will always have want to feed these into a risk management framework
new products and new processes, which means that or a portfolio optimization problem, you’d better make
there will always be a need to retrain the models,” he sure that the output correctly captures the risk of the
explains. portfolio. That isn’t obvious at all.”

“In financial services, you will always have new

products and new processes, which means that there
will always be a need to retrain the models.”
Chia Hock Lai, Co-Founder, Global Fintech Institute
20 MIT Technology Review Insights

0
Other challenges with the technology, according to the firm and potential sanctions.39 Hallucinations are a
Mileham, include the “classic” issues of AI such as bias known problem but, so far, research to address them
and accountability. According to a recent IMF study, has focused on specific cases rather than the general
generative AI, if anything, exacerbates these problems. issue.40
The far greater breadth of data used to train LLMs, for
example, leads to a greater theoretical possibility of These kinds of issues make rapid adoption of
bias. Similarly, the higher complexity of its architecture generative AI tools across a wider range of functions
and decision-making processes compared to previous irresponsible. “Institutional leaders must first assess
AI makes the reasoning behind given output more the safety and security of these tools, and address
opaque.33 “Very often, the sophisticated models of deep any forthcoming challenges or risks,” says Villanueva.
learning are black boxes,” says Chia.
Regulatory risks of a new technology
Generative AI also poses its own specific challenges. Accountability is at the core of industry thinking on the
These may arise even when the technology is working rollout of generative AI. UBS research points to potential
as planned. A recent study co-authored by Cont shows regulation as the main barrier to adoption of generative
that algorithms that have learned from a common set AI in the fintech space. Others argue that the same
of data—such as the history of asset prices—may end could be said of the entire financial sector. “AI is not
up synchronizing as if they were a cartel even though an easy button to use to bypass the accountability
they are not communicating. “If an algorithm learns to that we have to our customers,” says Mileham.
manipulate prices, can you sue anybody?” asks Cont.
“It is a legal nightmare. This is one of the questions Regulators strongly agree. In a July 2023 speech,
we are just beginning to study.”34 Worse still, the the chief executive of the UK’s Financial Conduct
potential for a herd-like response inherent in such an Authority (FCA) reiterated: “While the FCA does not
unconsciously coordinated response could present a regulate technology, we do regulate the effect on—
threat to financial stability under a worst-case scenario, and use of—tech in financial services.... With these
according to IMF research.35 Meanwhile, if generative AI developments [the growing use of generative AI], it is
sees increasing use for automated decision-making, it critical we do not lose sight of our duty to protect the
is likely to attract a high number of adversarial attacks.36 most vulnerable and to safeguard financial inclusion
and access.”41 A month earlier, the U.S. Consumer
Intellectual property rights and Financial Protection Bureau warned that, with the
hallucinations applications already being put in place, “financial
Useful answers produced by generative AI may violate institutions risk violating legal obligations, eroding
the IP rights of other actors, depending on the inputs customer trust, and causing consumer harm when
used for training the model. Already, several artists deploying chatbot [sic] technology.”42 Meanwhile, the
have launched lawsuits based on the inclusion of their U.S. Securities and Exchange Commission has invited
works in training data.37 Meanwhile, if a generative AI discussion on proposed new regulations regarding
tool is trained on licensed software and generates new AI-enabled tools.43
code, that may violate the IP rights of the licensor. At
the same time, it is not a straightforward question in law Accordingly, financial services companies are looking
if a company can license software created within its to regulators for guidance on how they can use the
computer systems by generative AI with minimal or no technology. “We have to start the AI project with a
human intervention.38 view toward being regulatorily compliant,” says Chia.
“There will be growing regulatory content so it is
Finally, things do not always work as planned. imperative for financial services to be able to engage
“Hallucinations are the key bit,” says Mileham, referring regulators effectively on what kind of rules and
to incorrect content that can be generated confidently standards there should be.” Mileham adds, “It’s going
by AI, which brings substantial risk. A lawyer in U.S. to be helpful to see regulators looking at this seriously
federal court relying on ChatGPT recently submitted and ensuring that all of the actors in these fields with
an affidavit in a personal injury lawsuit that included six fiduciary responsibilities understand the depth of
fake cases. This led to substantial news coverage for those responsibilities.”
 MIT Technology Review Insights 21

07
T
Conclusion: Valuable
tool, but yet to be fully
disruptive
he arrival of generative AI as a viable
technology is important news for the financial
services industry. Its application, however, will
for now be limited to specific uses where
other, cheaper technologies are not already
sufficient for requirements and where the weaknesses of
Nevertheless, generative AI continues to evolve quickly,
and solutions to its drawbacks may rapidly emerge.
It is very likely, for example, that ways will be found
to reduce the number and extent of hallucinations,
although some errors appear to be inevitable.44 If such
problems do persist, the technology might still find
the technology do not put off executives and regulators. new uses in financial services, as long as the potential
The difficulty in providing a more detailed forecast is the drawbacks are fully understood by all parties. “We’ll
uncertainty in how today’s failings might be overcome. have to figure out as a culture, and within the regulatory
regime, what that should mean,” says Mileham. “It might
The strengths of generative AI shape its current be that people just get comfortable with the constraints
application. In particular, with the capacity to learn of AI and you’re able to, with a straight face, disclose
from large amounts of unstructured data and to create to a customer that this thing is useful, but it is also
output of various kinds, it is already beginning to play potentially flawed, so take it with a grain of salt.”
an important role in customer service support, fraud
prevention, risk management, code generation, and As generative AI and societal attitudes inter-evolve,
unstructured data analysis. Its impact in these use financial services companies should reap the substantial,
cases should not be underestimated simply because if constrained, benefits of the new technology while
the technology is not being deployed more widely. keeping an eye on where it might head.
As generative AI is rolled out further in these areas,
substantial cost savings—up to $340 billion annually
across the financial services industry—are likely from
the shift in human employment away from repetitive,
low-value tasks toward more creative and profitable
The use of generative AI in
ones. In that sense, it is no exaggeration to say, in
Mileham’s words, “This is a big moment.”
more sensitive tasks closer
to the core of financial
Compared to what might eventually be possible,
however, “we are only scratching the surface at the service business models—
moment,” says Briest. The use of generative AI in more
sensitive tasks closer to the core of financial service such as making decisions
business models—such as making decisions on, or
executing, investment strategy for clients or companies
on, or executing, investment
themselves—remains some way off. Too many questions
hang over the technology itself for companies or
strategy for clients or
regulators to give it free rein any time soon. Villaneuva companies themselves—
also urges caution. “As exciting as this technology is, we
must be wary before fully realizing the benefits,” he says. remains some way off.
22 MIT Technology Review Insights

Footnotes

1. Krystal Hu, “ChatGPT sets record for fastest-growing user base – analyst note,” 23. “Introducing BloombergGPT, Bloomberg’s 50-billion parameter large language
Reuters, February 3, 2023, https://www.reuters.com/technology/chatgpt-sets-re- model, purpose-built from scratch for finance,” March 30, 2023, Bloomberg,
cord-fastest-growing-user-base-analyst-note-2023-02-01/. https://www.bloomberg.com/company/press/bloomberggpt-50-billion-parameter-
2. Michael Chui, Eric Hazan, Roger Roberts, Alex Singla, Kate Smaje, Alex llm-tuned-finance/.
Sukharevsky, Lareina Yee, and Rodney Zemmei, “The economic potential of 24. Ariel Bogle, “New York Times, CNN and Australia’s ABC block OpenAI’s GPTBot
generative AI: The next productivity frontier,” McKinsey, June 14, 2023, https:// web crawler from accessing content,” Guardian, August 25, 2023, https://www.
www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-po- theguardian.com/technology/2023/aug/25/new-york-times-cnn-and-abc-block-
tential-of-generative-ai-the-next-productivity-frontier. openais-gptbot-web-crawler-from-scraping-content.
3. Michael Briest, Lloyd Walmsley, Karl Keirstead, Timothy Arcuri, Francois-Xavier 25. Yada Zhu et al., “PAGAN: Portfolio Analysis with Generative Adversarial Networks,”
Bouvignies, Gerry Fowler, Nicolas Gaudois, Victoria Kalb, Navin Killa, David Vogt, IBM Research, December 26, 2020, https://ssrn.com/abstract=3755355; and
and Kathinka de Kuyper, “Will Generative AI deliver a generational transforma- Chi Seng Pun et al., “Financial Thought Experiment: A GAN-based Approach
tion?” UBS, May 26, 2023, https://www.ubs.com/global/en/investment-bank/ to Vast Robust Portfolio Selection,” IJCAI’20: Proceedings of the Twenty-Ninth
in-focus/2023/generative-ai.html. International Joint Conference on Artificial Intelligence, 2021, https://www.ijcai.
4. Ghiath Shabsigh and El Bachir Boukherouaa, “Generative Artificial Intelligence in org/proceedings/2020/0637.pdf.
Finance: Risk Considerations,” International Monetary Fund, August 2023, https:// 26. Aliaa Bassiouny, Omar Khater, and Eskandar Tooma, “Enhancing Value Investing
www.imf.org/-/media/Files/Publications/FTN063/2023/English/FTNEA2023006. Using an AI Genetic Algorithm Model for Dynamic Security Selection,” MS provid-
ashx. ed by author.
5. “Gartner Places Generative AI on the Peak of Inflated Expectations on the 2023 27. Gautier Marti, “CorrGAN: Sampling Realistic Financial Correlation Matrices Using
Hype Cycle for Emerging Technologies” (press release), Gartner, August 16, 2023, Generative Adversarial Networks,” 2020 IEEE International Conference on Acous-
https://www.gartner.com/en/newsroom/press-releases/2023-08-16-gartner-plac- tics, Speech and Signal Processing (ICASSP), https://arxiv.org/pdf/1910.09504.
es-generative-ai-on-the-peak-of-inflated-expectations-on-the-2023-hype-cycle- pdf.
for-emerging-technologies. 28. Rama Cont et al., “Tail-GAN: Learning to Simulate Tail Risk Scenarios,” arXiv.org,
6. Michael Chui, Eric Hazan, Roger Roberts, Alex Singla, Kate Smaje, Alex revised March 25, 2023, https://arxiv.org/pdf/2203.01664.pdf.
Sukharevsky, Lareina Yee, and Rodney Zemmei, “The economic potential of 29. Hugh Son, “JPMorgan is developing a ChatGPT-like A.I. service that gives invest-
generative AI: The next productivity frontier,” McKinsey, June 14, 2023, https:// ment advice,” CNBC, May 25, 2023, https://www.cnbc.com/2023/05/25/jpmor-
www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-po- gan-develops-ai-investment-advisor.html.
tential-of-generative-ai-the-next-productivity-frontier.
30. Dave McKay, “What Is COBOL, and Why Do So Many Institutions Rely on It?”
7. Ibid. How-To-Geek, April 15, 2020, https://www.howtogeek.com/667596/what-is-co-
8. “Generative AI,” Boston Consulting Group, https://www.bcg.com/capabilities/artifi- bol-and-why-do-so-many-institutions-rely-on-it/.
cial-intelligence/generative-ai. 31. “The great cloud mainframe migration: what banks need to know,” Accenture,
9. Michael Chui, Eric Hazan, Roger Roberts, Alex Singla, Kate Smaje, Alex Banking Cloud Altimeter, April 2022, https://bankingblog.accenture.com/wp-con-
Sukharevsky, Lareina Yee, and Rodney Zemmei, “The economic potential of tent/uploads/2022/04/Accenture-Altimeter-PDF-02-1.pdf.
generative AI: The next productivity frontier,” McKinsey, June 14, 2023, https:// 32. “AI: The missing piece for Cobol modernisation in finance,” FinTech Futures,
www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-po- September 9, 2019, https://www.fintechfutures.com/2019/09/ai-the-missing-
tential-of-generative-ai-the-next-productivity-frontier. piece-for-cobol-modernisation-in-finance/; and Michael Chui, Eric Hazan, Roger
10. Justina Lee, Saijel Kishan, and Bloomberg, “Hedge funds are using ChatGPT to Roberts, Alex Singla, Kate Smaje, Alex Sukharevsky, Lareina Yee, and Rodney
do all their mundane grunt work usually relegated to junior staffers,” Fortune, Zemmei, “The economic potential of generative AI: The next productivity frontier,”
June 2, 2023, https://fortune.com/2023/06/01/hedge-fund-chatgpt-grunt-work- McKinsey, June 14, 2023, https://www.mckinsey.com/capabilities/mckinsey-dig-
mundane/. ital/our-insights/the-economic-potential-of-generative-ai-the-next-productiv-
11. Alex Kim et al., “Bloated Disclosures: Can ChatGPT Help Investors Process ity-frontier.
Financial Information?” University of Chicago, April 2023, https://bfi.uchicago.edu/ 33. Ghiath Shabsigh and El Bachir Boukherouaa, “Generative Artificial Intelligence in
wp-content/uploads/2023/04/BFI_WP_2023-59.pdf. Finance: Risk Considerations,” International Monetary Fund, August 2023, https://
12. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work,” www.imf.org/-/media/Files/Publications/FTN063/2023/English/FTNEA2023006.
National Bureau of Economic Research, October 2023, https://www.nber.org/ ashx.
system/files/working_papers/w31161/w31161.pdf. 34. Rama Cont and Wei Xiong, “Dynamics of market making algorithms in dealer mar-
13. Ibid. kets: Learning and tacit collusion,” Mathematical Finance, May 30, 2023, https://
onlinelibrary.wiley.com/doi/full/10.1111/mafi.12401.
14. Douglas Blakey, “Kasisto launches KAI-GPT, the first banking industry-specific
large language model, Retail Banker International, June 1, 2023, https://www. 35. Ghiath Shabsigh and El Bachir Boukherouaa, “Generative Artificial Intelligence in
retailbankerinternational.com/news/digital-experience-platform-kasisto-launch- Finance: Risk Considerations,” International Monetary Fund, August 2023, https://
es-kai-gpt/. www.imf.org/-/media/Files/Publications/FTN063/2023/English/FTNEA2023006.
ashx.
15. Aoibhinn McBride and Jobbio, “FinTech Futures Jobs: Three ways generative AI
is being used in financial services,” FinTech Futures, June 2, 2023, https://www. 36. Clark Barrett et al., “Identifying and Mitigating the Security Risks of Generative AI,”
fintechfutures.com/2023/06/fintech-futures-jobs-three-ways-generative-ai- October 17, 2023, https://arxiv.org/pdf/2308.14840v3.pdf.
is-being-used-in-financial-services/; and Tatiana Bautzer and Lananh Nguyen, 37. Gil Appel et al., “Generative AI Has an Intellectual Property Problem,” Harvard
“Morgan Stanley to launch AI chatbot to woo wealthy,” Reuters, September 7, Business Review, April 7, 2023, https://hbr.org/2023/04/generative-ai-has-an-in-
2023, https://www.reuters.com/technology/morgan-stanley-launch-ai-chatbot- tellectual-property-problem.
woo-wealthy-2023-09-07/. 38. Chris Baker, “Intellectual property and generative AI: Five IP risks for users and
16. “Cowbell introduces MooGPT,” TechsterHub, June 21, 2023, https://www.techster- developers,” Deloitte (blog), September 2023, https://www2.deloitte.com/uk/en/
hub.com/news/cowbell-introduces-moogpt/. blog/ai-institute/2023/intellectual-property-and-generative-ai.html.
17. Samuel May, “Generative AI and the Fraud Examiner,” Association of Certified 39. Lyle Moran, “Lawyer cites fake cases generated by ChatGPT in legal brief,” Legal
Fraud Examiners, May 9, 2023, https://www.acfe.com/fraud-resources/fraud-ex- Dive, May 30, 2023, https://www.legaldive.com/news/chatgpt-fake-legal-cas-
aminer-archives/fraud-examiner-article?s=may-2023-generative-ai. es-generative-ai-hallucinations/651557/.
18. “The Next Evolution in Finance: Leveraging Generative AI for Enhanced Risk 40. Ziwei Ji et al., “Survey of Hallucination in Natural Language Generation,” ACM
Management and Fraud Detection,” iffort, April 7, 2023, https://www.iffort.com/ Computing Surveys, March 3, 2023, https://dl.acm.org/doi/10.1145/3571730.
blog/2023/04/07/generative-ai-for-enhanced-risk-management-and-fraud-de- 41. “Our emerging regulatory approach to Big Tech and Artificial Intelligence” (press
tection/. release), Financial Conduct Authority, July 12, 2023, https://www.fca.org.uk/news/
19. Ryan Browne, “Goldman Sachs is using ChatGPT-style A.I. in house to assist speeches/our-emerging-regulatory-approach-big-tech-and-artificial-intelligence.
developers with writing code,” CNBC, March 22, 2023, https://www.cnbc. 42. “Chatbots in consumer finance,” Consumer Financial Protection Bureau, June
com/2023/03/22/goldman-sachs-experiments-with-chatgpt-like-ai-to-help-devs- 2023, https://s3.amazonaws.com/files.consumerfinance.gov/f/documents/cfpb_
write-code.html. chatbot-issue-spotlight_2023-06.pdf.
20. Kate Weber, “Westpac sees 46 percent productivity gain from AI coding 43. “SEC Proposes New Requirements to Address Risks to Investors From Conflicts
experiment,” IT News, June 1, 2023, https://www.itnews.com.au/news/west- of Interest Associated With the Use of Predictive Data Analytics by Broker-Deal-
pac-sees-46-percent-productivity-gain-from-ai-coding-experiment-596423. ers and Investment Advisers” (press release), U.S. Securities and Exchange
21. Justina Lee, Saijel Kishan, and Bloomberg, “Hedge funds are using ChatGPT to Commission, July 26 2023, https://www.sec.gov/news/press-release/2023-140.
do all their mundane grunt work usually relegated to junior staffers,” Fortune, 44. Kyle Wiggers, “Are AI models doomed to always hallucinate?” TechCrunch,
June 1, 2023, https://fortune.com/2023/06/01/hedge-fund-chatgpt-grunt-work- September 4, 2023, https://techcrunch.com/2023/09/04/are-language-mod-
mundane/. els-doomed-to-always-hallucinate/.
22. Tatiana Bautzer and Lananh Nguyen, “Morgan Stanley to launch AI chatbot to
woo wealthy,” Reuters, September 7, 2023, https://www.reuters.com/technology/
morgan-stanley-launch-ai-chatbot-woo-wealthy-2023-09-07/.
 MIT Technology Review Insights 23

About MIT Technology Review Insights

MIT Technology Review Insights is the custom publishing division of MIT Technology Review, the world’s
longest-running technology magazine, backed by the world’s foremost technology institution—producing
live events and research on the leading technology and business challenges of the day. Insights conducts
qualitative and quantitative research and analysis in the US and abroad and publishes a wide variety of content,
including articles, reports, infographics, videos, and podcasts. And through its growing MIT Technology Review
Global Insights Panel, Insights has unparalleled access to senior-level executives, innovators, and
entrepreneurs worldwide for surveys and in-depth interviews.

From the sponsor

UBS is a leading and truly global wealth manager and the leading universal bank in Switzerland. It also provides
diversified asset management solutions and focused investment banking capabilities. With the acquisition of
Credit Suisse, UBS manages 5.5 trillion dollars of invested assets as per second quarter 2023. UBS helps
clients achieve their financial goals through personalized advice, solutions and products. Headquartered in
Zurich, Switzerland, the firm is operating in more than 50 countries around the globe. UBS Group shares are
listed on the SIX Swiss Exchange and the New York Stock Exchange (NYSE).

Illustrations
Illustrations assembled by Tim Huxford with elements from Shutterstock.

While every effort has been taken to verify the accuracy of this information, MIT Technology Review Insights cannot accept any responsibility or liability for reliance by any person
on this report or any of the information, opinions, or conclusions set out in this report.

© Copyright MIT Technology Review Insights, 2023. All rights reserved.

MIT Technology Review Insights
www.technologyreview.com
[email protected]