ChatGPT, the chatbot that Microsoft made available to the public on Nov. 30, has drawn unprecedented popularity and media attention. It attracts an estimated 96 million users per month, who ask it questions of all kinds. At first, users were wowed. It was able to write limericks about Google and rap songs about cats. Students got it to write essays for them. Microsoft founder Bill Gates said AI chatbots like ChatGPT will become “every bit as important” as the PC or the internet. Google leaders reportedly issued a “code red” over ChatGPT and redirected engineers to a competing AI project called Lamda.
Then users started pointing out flaws: Some of ChatGPT’s answers are wrong. Its knowledge base has not been updated since 2021. It crawls the internet, absorbing content from any source, and doesn’t reveal its sources. And some reporters have found some interactions with it disturbing.
When you peel away the hype, ChatGPT is not ready to be used in finance today. But it could be in the future, if the right direction and guardrails were applied.
“I try to keep an open mind and I also have a healthy degree of skepticism because I know that something new comes on the scene and people like to hype it and give it attributes that maybe it doesn’t deserve quite yet and downplay the negatives,” said Steve Rubinow, a faculty member in the Jarvis College of Computing and Digital Media at DePaul University and former chief information officer of the New York Stock Exchange. “I consider it a large language model, and like any model I think it’s best used in the hands of somebody who’s an expert. And they could be in financial services.”
Asked how it could be used in financial services, ChatGPT itself provided a bullet-point list that included customer service chat, fraud detection, personal financial management, automating routine tasks and improving customer experience, content generation (such as financial report summaries) and risk management (by analyzing vast amounts of data and supporting risk management processes).
Hot topic in banking
The technology has generated buzz in banking circles.
“I don’t know whether to be thrilled or terrified,” said Wade Peery, chief innovations officer at FirstBank in Nashville. “Like most folks, I have played around with it. I think it could be useful in providing information and answers if it is specifically focused on the right things.”
The media attention and the millions of people engaging with ChatGPT have been surprising to Chintan Mehta, CIO of strategy, digital and innovation at Wells Fargo.
“When you are in the middle of it, you don’t necessarily see some of the incremental stuff, but when you get exposed to it at one shot, it feels like it’s very groundbreaking and it’s very out there,” he said.
His own initial reaction to ChatGPT was muted.
“I’ve used large language models before,” Mehta said. “I don’t personally feel that much of a surprise when it’s a very well-formed, cogent set of sentences coming back. A large language model has always been able to do that.”
More important, in his view, is the actual substance of the responses and how they might be made better.
“I think that’s the evolution we have to go through,” Mehta said.
Not ready to be a bank chatbot
Banks have used AI-powered chatbots for years. Bank of America’s Erica virtual assistant uses home-grown machine learning to answer customers’ questions. TD Bank uses technology from Kasisto, a spinoff of the research lab SRI International, which also developed Apple’s Siri. Wells Fargo’s Fargo virtual assistant, which will be rolled out to customers next month, uses technology from Google called Dialogflow.
These and others like it are carefully constructed and draw answers from custom-built libraries of content. They have natural language understanding that can interpret the customer’s question, and entity extraction that can extract the key data elements and find the right curated answer. These chatbots do not randomly crawl the internet for information, and there’s no reason to think they ever will.
“Our teams are looking at ChatGPT, we’ve done a lot of experimentation with it,” said Ken Meyer, chief information and experience officer at Truist. “If you need to write a term paper right now, it’s pretty phenomenal. But you’re talking about a language-based platform that basically makes up stuff based on everything that’s feeding it. So in a highly regulated industry, I don’t suspect that ChatGPT is going to be answering questions for clients anytime soon.”
Meyer predicts a tech company like Amazon, Google or Microsoft will come out with a generative AI language-based product that companies can plug into a platform like Amazon Lex.
“There’s probably a lot of really talented engineers in a room somewhere figuring out how to make it a little bit tighter and a little less unpredictable,” he said.
Truist is about to relaunch its Truist Assist chatbot, which originally was a combination of Amazon Lex, the Amazon Web Service conversational AI platform behind Alexa, and technology from another vendor. The new version will be a full-stack AWS solution, according to Meyer, “which integrates really nicely with our online and mobile application.”
How large language models could work in banking
The technology that underlies ChatGPT, a large language model, does have a place in banking, Mehta said. A large language model is a deep learning algorithm that can recognize, summarize, translate, predict and generate text and other content based on knowledge gained from massive datasets. It’s designed to understand natural language and generate human-like responses to input text or speech.
A bank might use a large language model to analyze a set of documents, summarize them quickly, and identify anything adversarial in the content that lawyers should look at, Mehta said.
Some financial firms have been testing this, Rubinow said.
“Everyone’s experimenting to see how good the results are and if it really is a time saver for them,” Rubinow said. They have to factor in the amount of time it takes to edit the results and discern whether they’re factual or not, he noted.
Rubinow warns people not to copy and paste internal documents into ChatGPT, to summarize them, because they can immediately become part of its training set and cannot be taken back.
“People are a little bit zealous and over eager and they want to use it for work,” Rubinow said. “But in terms of security and privacy, it’s an open door and you have to be conscious of that.”
Developers already use ChatGPT-like technology to auto-complete code, Mehta said.
“You still are responsible as a developer to make sure that it’s working and it’s doing the right thing,” Mehta said. “These have productivity boosts that I think will get used across industries,” he said.
Wells Fargo doesn’t rule out the possibility of using ChatGPT in the future.
“We have a responsibility to have thoughtful AI in whatever form that it takes, including generative AI and large language models — to learn from them and build properly,” said Michelle Moore, head of consumer and wealth and investment management digital at Wells Fargo. “This is an evolution. ChatGPT is just the next generation and there will be something after ChatGPT. All of this is good if used wisely and built thoughtfully.”
Zor Gorelov, founder and CEO of the financial conversational AI company Kasisto, sees ChatGPT as transformational.
“From a financial services industry perspective, I think a lot of things still need to be figured out: accuracy, timeliness, training costs, general costs and truthfulness,” he said. “These are not trivial issues that need to be worked out. But if you look at the bigger picture, ChatGPT passed U.S. medical licensing exams and another exam at the University of Pennsylvania’s Wharton School of Business. So how long before it starts passing Series 7 exams? I do think that things will change faster than many people expect.”
A virtual assistant using ChatGPT technology, trained on unbiased and proprietary (not public) data, could potentially advise investors more accurately than human financial advisors, Gorelov said.
“There are a lot of gray areas, but we envision a rapidly emerging market for new tools that address those ambiguities. And I think overall, it is a watershed event in the industry,” he said.
For chatbot providers like Kasisto, ChatGPT may be a threat, but Gorelov sees it as an opportunity.
“We’ve been working with large language models and generative AI for several years, and we’re excited about what this next generation can do to improve our products and provide more contextual answers for customers,” he said.
A chatty search engine
ChatGPT can also be used as a conversational search engine.
“ChatGPT takes your search results and wraps them up in a nice package and puts a bow around them and gives them to you,” Rubinow said. “But you have to be a discerning consumer of the data because we all know it makes mistakes. So we have to be able to spot the mistakes.”
The outputs also need polishing because they are machine generated, he noted.
“I look at it as a tool to help prompt my thinking in the same way that I would do a search if I went to Google or Bing,” he said. “The first page of hits is probably going to be the most relevant, because that’s what the algorithm is pulling out and then ranks in order of relevance. You, being the expert, have to figure out if it makes sense. So I think it’s a good tool for priming the creative pump. But you don’t want to attribute abilities to it that it can’t do.” Also, it’s not always good at math, Rubinow said. “And math is kind of important for finance.”
It will undoubtedly get better, he said.
Companies could potentially install a version of ChatGPT on their own servers and keep it off the public-facing internet, Rubinow mused.
ChatGPT could be useful to IT staff who have to find how to fix technology vulnerabilities or problems.
“In some ways, this is just a souped-up search engine,” Rubinow said. If a user types in an error message or a description of a problem, it could quickly find an answer. “Whether I can rely on that and whether it’s actually accurate and whether it’s actionable or whether it’s really the thing that I need to act on, I can’t know without further thought.”
Potential for misinformation and bias
As it stands today, ChatGPT has obvious shortcomings.
“These systems have bias and they can provide some well-articulated but inaccurate answers,” said Gorelov. “We know too well that the banks and their customers cannot tolerate this.”
The free-ranging nature of ChatGPT is off-putting to regulated companies and investors.
“We don’t know where the guardrails are,” said Michael Loukas, principal and CEO of TrueMark Investments, an ETF advisory firm; one of its funds invests in AI tech companies. “I think the concept is awesome. And the uses are limitless, whether it comes to education or efficiency or delivery of data. And of course the uses can also be quite nefarious.”
ChatGPT shares a weakness common to all machine learning models: They are trained on whatever examples and data they are given, and they have no ability to discern fact from fiction or useful data versus racist rant. In 2016, Microsoft introduced a Twitter bot called Tay that learned to be racist through conversations with Twitter users.
“Training algorithms is almost like training a dog,” Loukas said. “Some of your biases can make their way into that training.”
Microsoft, which bought an exclusive license to the underlying technology behind GPT-3 in 2020 after investing $1 billion in OpenAI in 2019, did not respond to a request for an interview. But in a blog, the OpenAI team acknowledged that ChatGPT “sometimes writes plausible-sounding but incorrect or nonsensical answers. Fixing this issue is challenging, as: (1) during training, there’s currently no source of truth; (2) training the model to be more cautious causes it to decline questions that it can answer correctly; and (3) supervised training misleads the model because the ideal answer depends on what the model knows, rather than what the human demonstrator knows.”
Microsoft is also trying to make the model refuse inappropriate requests, but “it will sometimes respond to harmful instructions or exhibit biased behavior,” the company wrote in a blog. “We’re using the Moderation API to warn or block certain types of unsafe content, but we expect it to have some false negatives and positives for now.”
Loukas’s firm looks at AI companies that produce algorithms, chip makers that supply processing power, data companies and companies that are sophisticated users of AI and use it as a competitive advantage or to solve a pain point.
“Our attitude has been that as an investment, ChatGPT doesn’t hold water,” he said. “As an eye-opening buzz for the AI, deep learning space, it’s fantastic.”
Vincent Cerf, a computer scientist who is considered one of the founders of the internet, has warned there are ethical issues with ChatGPT. Rubinow, who has taught ethics and technology, agrees.
“Here’s the problem,” Rubinow said. “You have machine-generated responses that have an air of authority about them because they came from a computer, so there must be some validity to them. And we know that really smart people have spent tons of money and tons of time developing these models. And so it might have a false sense of credibility for some issues.”
With large language models, “we can’t really be sure all the time why we get the results we get,” Rubinow said. “We’re not talking about a handful of developers that might have implicit bias. We’re talking about the entire body of knowledge that’s available on the web that has implicit bias in it, some intentional, some unintentional. And the naive user, hopefully there aren’t too many, might look at this and forget to ask the right questions or be swayed incorrectly.
“There are ethical implications there and that’s why you have to be really cautious,” Rubinow said. “We want to ask the right questions and always look at things with a critical eye so we’re not misled or we could help others not be misled. And we don’t know where those issues lie.”