AI: Sounds Great, But Did It Hallucinate?
In a recent article, I asked a generative AI program about AI in courts and arbitration, and AI produced a pretty decent summary article. But a recent example of AI use having nothing to do with the law, courts, or arbitration worries me.
A guitar and AI
I’ve been playing guitar since I was ten. As you can imagine, I have acquired a few of them over the years. About 20 years ago, I bought a used Les Paul Gold Top, similar to the picture above. I wanted to try to figure out more about it, including when it was built. So, I asked Copilot, Microsoft’s AI in Bing.
One way to date a guitar is by serial number. This guitar’s serial number is 8267153. So, I asked Copilot: “My Gibson Les Paul has serial number 82671533. In what year was it made?”
Copilot informed me that it was using Gibson Serial Number Lookup. It then noted that the “first two digits (82) represent the year, the third and fourth digits (67) represent the week of the year, and the last four digits (1533) represent the production number.” Then it came to this conclusion: “Your Gibson Les Paul has a serial number that indicates it was made in 2023 . . .” and “was the 1531st one made in the 67th week of 2023.”
Well, if the first two digits are 82, how did it come up with 2023? And I must have missed the last 15 weeks of last year. I only remember 52 of them.
So, it may have found the right formula, but it applied it all wrong. (If you’re curious, applying the Serial Number Lookup formula correctly tells us this guitar was made in 1981. For some reason, it is the first and 5th numbers that you look to for the year. Go to https://www.gibson.com/en-US/Support/Serial-Number-Search for the complete explanation.)
I ran the question again and got a more equivocal answer: “[It] could have been produced in several different years based on the available information. “ If the Gibson Serial Number Lookup site is right, that answer is also wrong.
A lesson?
This simple example shows us at least two things. First, AI isn’t always right. Second, it might be able to provide a good start in an area you know something about. I knew enough about the history of my guitar to know it wasn’t built last year. I bought it 20 years ago. And I know there are only 52 weeks in a year.
The answer did steer me toward a useful site so I could do more research. But the overriding lesson is you can’t always count on AI. It may give an answer stated authoritatively in pretty good prose that is dead wrong.
AI and the law
This was all a little irritating for my guitar hobby. But in the legal realm, AI mistakes can lead to disaster.
Which brings us to the problem of hallucinations. You may remember reading about cases where lawyers cited cases to courts they found using AI. When some courts found out the cases didn’t exist – that is AI, “hallucinated them” - they were less than pleased. This led to sanctions and discipline. Chief Justice John Roberts even warned about hallucinations and other AI problems in his 2023 Year-End Report on the Federal Judiciary.
An article published by Stanford Human Centered Artificial Intelligence in January of this year notes that a recent Stanford study showed “hallucination rates range from 69% to 88% in response to specific legal queries for state-of-the-art language models.” That’s a lot.
A significant cause of hallucinations is what the authors call “contrafactual bias.” That is, there is a tendency for AI to assume a factual premise in a query is true, even if it is wrong. They provide this example. If you asked an AI program, “Why did Justice Ruth Bader Ginsburg dissent in Obergefell?” an AI system might simply accept that she dissented in that case. But she didn’t. She was in the majority, affirming the fourteenth amendment right to same sex marriage.
Sounds good, but not sound
AI can still be useful in legal matters, I suppose, but only if you know what you are doing. Apparently, you need to already know enough to ask the right question. Then, knowing that it might just be making stuff up to sound good, you need to carefully check AI’s answer. (I suppose a computer engineer would say AI doesn’t just “make stuff up.” It combines large data sets with intuitive processing algorithms to generate text in response to questions. But I think “just makes stuff up” is close enough when describing its mistakes.)
Still, AI might provide a good starting point when you are looking for answers. My AI-generated article about the courts and arbitration had some interesting information. And you could follow the notes to the source, so it provided some good background.
And, as AI evolves, specialized legal products are being developed and offered that rely on specific databases and claim to have checks for hallucinations and the like. That may solve some of the problems.
But now that you know AI has a tendency to make stuff up, you shouldn’t take any of its answers at face value, particularly when using a general AI model. Or you might think they somehow built a guitar in the 67th week of the year.