text
stringlengths
0
182
challenges and the
regulatory hurdle
The difficulties seen in market simulation software
illustrate some of the likely, wider challenges for
generative AI. According to Cont, while generative AI
tools can create new data sets, they will not include
any new information if, for example, a company wants
better insight into tail risk. More broadly, he adds, users
cant just hope that a generic algorithm will extract
what is needed for a niche use. If you dont target
what you want to learn, you may not learn it, Cont
explains. You have to ask what you want to use output
for and then create a bespoke algorithm tailored to the
intended use case. When you generate cat pictures
from generative AI, you dont want a mutant cat, but
in financial applications you are typically interested in
extreme events.
Reliability, bias, and accountability
Additionally, checking the reliability of the output
presents important challenges for activities that use
substantial amounts of generated data. There is no
easy way to validate the output, says Cont. If you
want to feed these into a risk management framework
or a portfolio optimization problem, youd better make
sure that the output correctly captures the risk of the
portfolio. That isnt 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
Other challenges with the technology, according to
Mileham, include the classic issues of AI such as bias
and accountability. According to a recent IMF study,
generative AI, if anything, exacerbates these problems.
The far greater breadth of data used to train LLMs, for
example, leads to a greater theoretical possibility of
bias. Similarly, the higher complexity of its architecture
and decision-making processes compared to previous
AI makes the reasoning behind given output more
opaque.33 Very often, the sophisticated models of deep
learning are black boxes, says Chia.
Generative AI also poses its own specific challenges.
These may arise even when the technology is working
as planned. A recent study co-authored by Cont shows
that algorithms that have learned from a common set
of datasuch as the history of asset pricesmay end
up synchronizing as if they were a cartel even though
they are not communicating. If an algorithm learns to
manipulate prices, can you sue anybody? asks Cont.
It is a legal nightmare. This is one of the questions
we are just beginning to study.34 Worse still, the
potential for a herd-like response inherent in such an
unconsciously coordinated response could present a
threat to financial stability under a worst-case scenario,
according to IMF research.35 Meanwhile, if generative AI
sees increasing use for automated decision-making, it
is likely to attract a high number of adversarial attacks.36
Intellectual property rights and
hallucinations
Useful answers produced by generative AI may violate
the IP rights of other actors, depending on the inputs
used for training the model. Already, several artists
have launched lawsuits based on the inclusion of their
works in training data.37 Meanwhile, if a generative AI
tool is trained on licensed software and generates new
code, that may violate the IP rights of the licensor. At
the same time, it is not a straightforward question in law
if a company can license software created within its
computer systems by generative AI with minimal or no
human intervention.38
Finally, things do not always work as planned.
Hallucinations are the key bit, says Mileham, referring
to incorrect content that can be generated confidently
by AI, which brings substantial risk. A lawyer in U.S.
federal court relying on ChatGPT recently submitted
an affidavit in a personal injury lawsuit that included six
fake cases. This led to substantial news coverage for
the firm and potential sanctions.39 Hallucinations are a
known problem but, so far, research to address them
has focused on specific cases rather than the general
issue.40
These kinds of issues make rapid adoption of
generative AI tools across a wider range of functions
irresponsible. Institutional leaders must first assess
the safety and security of these tools, and address
any forthcoming challenges or risks, says Villanueva.
Regulatory risks of a new technology
Accountability is at the core of industry thinking on the
rollout of generative AI. UBS research points to potential
regulation as the main barrier to adoption of generative
AI in the fintech space. Others argue that the same
could be said of the entire financial sector. AI is not
an easy button to use to bypass the accountability
that we have to our customers, says Mileham.
Regulators strongly agree. In a July 2023 speech,
the chief executive of the UKs Financial Conduct
Authority (FCA) reiterated: While the FCA does not
regulate technology, we do regulate the effect on
and use oftech in financial services.... With these