Opinion: Using ChatGPT in Compliance? This AI isn’t quite ready yet.
21st June 2023 by Samuel Rossiter
Large Language Models (LLM) and Artificial Intelligence (AI) software like ChatGPT are a hot topic right now. The general sentiment is that people are either very excited or slightly terrified about how this current wave of technology will revolutionise the way we work, and how this may affect our jobs.
Work in trading? According to Business Insider, ChatGPT could replace jobs across a range of Wall Street industries, from trading to investment banking.
Work in content creation? ChatGPT can write articles in seconds on demand, and some fear that the AI software may replace copywriters entirely.
Work as a coder or a programmer? ChatGPT can write code in a flash and far quicker than any human being. According to some sources, the company who created ChatGPT (Open AI) may already be training its AI technology to replace some of its software engineers.
If you need some more proof of how ChatGPT and other AI is already changing the world, here are 10 real-world examples of how companies are already using ChatGPT in 2023.
With its seemingly impressive capabilities, it is no surprise that the software reached 100 million users in just two months after its launch, which means that it could be one of the fastest-growing apps in history.
Even those who work in governance, risk and compliance are not exempt from the general global sentiment that “the robots are coming for us all, oh God, oh God”. In fact, there are some vendors in the compliance surveillance space who are already experimenting with ChatGPT—integrating it within their platforms to do things like reduce false positives and further query data. As a business that believes in the power of automation and technology to strengthen compliance, we here at Fingerprint have been intrigued by the potential of this newest wave of AI.
But, here’s our question: Should ChatGPT be used in compliance right now?
Our answer? We don’t think so—yet.
As it stands, real-world usage of ChatGPT has presented some major red flags, which must be considered before deciding that it is an appropriate tool for compliance and, more importantly to us, communications monitoring.
Here are three reasons why:
1. ChatGPT doesn’t always present factual information
If you Google ‘chatgpt lies’ or ‘chatgpt fake results’, you get a list of examples demonstrating how ChatGPT has provided incorrect information when queried.
There are a few examples from The Guardian, where their journalists have received emails from researchers asking for links to specific Guardian articles, of which the journalists had no recollection of writing.
The reason why? Those researchers carried out their research using ChatGPT, and when it was asked to list out articles on the subject, the AI made some articles up. No lie!
What’s particularly unnerving about this example is that those fake articles were convincing enough to these researchers. They then spent time and energy reaching out to the journalists named as their authors to try to cite these made-up articles. Some embarrassment was experienced, we’re sure.
This warning blog from Duke University around ChatGPT and its fake citations sum up the problem quite well:
“ChatGPT has significant limitations as a reliable research assistant. One such limitation is that it has been known to fabricate or “hallucinate” (in machine learning terms) citations. These citations may sound legitimate and scholarly, but they are not real… if you try to find these sources through Google or the library—you will turn up NOTHING.”
If you’d like to see the real-world consequences of fake ChatGPT information and sources, ask Jonathan Turley. As part of a research study, a lawyer in California asked ChatGPT to make a list of legal scholars guilty of sexual harassment. On the list that ChatGPT generated, Jonathan Turley was on it—the AI said that Turley had made sexually suggestive comments towards a student and had tried to touch them during a class trip to Alaska. ChatGPT cited a March 2018 article in The Washington Post as evidence.
But the article did not exist. There had never been a trip to Alaska, and Turley had never been accused of sexually harassing a student.
Technology that does not present factual data every time it is queried and essentially ‘lies’? Terrible for compliance. How can you attempt to manage risk and conduct within your firm or on behalf of your clients’ firm when the technology you use to query data could be making up its results?
2. ChatGPT is not consistent
So what exactly is ChatGPT? To get a very thorough and in-depth answer, it’s best to read this Medium article by Colin Fraser, who is a data scientist at Meta (the company behind Facebook). The article delves into the very nature of ChatGPT and more of its flaws, but we’ll try to sum it up below:
ChatGPT uses a Large Language Model which is a type of AI algorithm. A language model is a “probability distribution over words”. Language models try and emulate human language by getting fed a lot of training data (e.g. books and articles written by humans), analysing the frequency and the probability of those words next to each other, then generating sentences of text by, as Fraser puts it, “finding the most likely next word given the previous word and the training data”.
ChatGPT itself has been trained on 300 billion words and 570GB of data taken from books, articles, websites, Wikipedia entries, and more. ChatGPT also adjusts its language model after every interaction with its users, so to the outside eye it looks like it is learning as it talks to you. As Fraser mentions, there’s also a bit of randomness to the LLM when it generates its answers for its users. All of this means that the answer it gives you today might be different from the answer it gives you in two weeks, even if you ask it the same question.
Just look at the three different answers it provides to Fraser’s question ‘what are the roots of the polynomial x^2 + 5x + 6?’ in Figures 1, 2 and 3.
Figure 1
Image by Colin Fraser from ChatGPT: Automatic expensive BS at scale
Figure 2
Image by Colin Fraser from ChatGPT: Automatic expensive BS at scale
Figure 3
Image by Colin Fraser from ChatGPT: Automatic expensive BS at scale
Technology that provides inconsistent answers to the same query, multiple times? Not something well suited for compliance either. Repeatability and consistency are the backbones of good compliance. Regulated financial businesses need compliance policies, processes and systems that create a consistent approach to risk management. Sheer randomness and receiving different outcomes at different times have no place in compliance.
3. ChatGPT can get things plain wrong
The three answers above to the question ‘what are the roots of the polynomial x^2 + 5x + 6?’
All wrong.
But what if you ask ChatGPT a simpler question? The classic: 2 + 2 = ?
Figure 4
Image by Colin Fraser from ChatGPT: Automatic expensive BS at scale
ChatGPT gets this correct!
At first…
Because if you push and question the AI’s answer, in actual fact…
Figure 5
Image by Colin Fraser from ChatGPT: Automatic expensive BS at scale
2 + 2 = 5, it seems.
ChatGPT gets most of the answers that it generates for its users correct, but there are many examples of incorrect answers that it generates too. As we mentioned before, this AI might give you 10 correct answers in a row, but ask it the same question the 11th time, and it may give you the wrong answer, with a fake source to boot.
So, to summarise, AI that presents false information, is inconsistent and provides incorrect answers to your queries? ChatGPT is not ready for compliance or communications monitoring, yet.
We’ll see what the future brings for ChatGPT as it develops and evolves. It’s a big brave world out there – and an exciting time to be in tech!
If you need to monitor your communications channels like Teams or Slack for compliance in a scalable, unified and streamlined way, or if you have clients that require communications monitoring to satisfy regulatory requirements, then you can learn more about the Fingerprint supervision platform here.
If you’re interested in chatting to us or getting a demo of our platform, then please get in touch below. Otherwise, you can go back to our blog to read more about our views on the industry and other useful content.
