You Ask, I Answer: Generative AI Hallucinations?

In today's episode, Brian asks regarding inherent risks. You talked about privacy bias and copyright when it comes to large language models. What about hallucinations or potential incorrect responses? Well, yes, of course. That's that's always a risk.

Um models, large language models in particular, can hallucinate. They can come up with incorrect information. And the reason for this is because they don't have any reasoning capability. Not really. Um, there is reasoning that kind of happens as a result when a model is very large.

The just the nature of the interlinking probabilities creates a reasoning-like emergent property. But for the most part, at the end of the day, all these models are just doing is predicting the next word, right? That is all they are doing. Um how long they've trained for, how many parameters, what their weights are, all that stuff, it's just a word guessing game for them internally. And so when they are given a response, um, they're given a prompt that doesn't make sense, they will hallucinate.

Or they do have a prompt that makes sense, but they don't know the answer, they will hallucinate. They will just make stuff up. Um, one of the most famous tests uh for this is to ask a model who was president of the United States in 1566. The the way these models work, they look at the words and phrases, they break them up, and they look at what has proximity to those terms. And early, uh early in the GPT models, they would say things like Christopher Columbus, because it was the name that was most closely associated with maybe early time periods and the United States, what is eventually the United States.

And that's a completely wrong answer. Um today's models don't make those mistakes because they've been trained better and bigger and stuff, but they that is always a risk. So there's two things you need to do to reduce the likelihood of risks. Number one, don't let models behave and act and do stuff unsupervised, right? You should always be checking their work and saying, oh, you know, is it still doing what it's supposed to be doing?

That's number one. And number two, uh, whatever tasks you're having the model perform, you should have some subject matter expertise in those tasks so that you can judge whether the output is correct or not. If I ask a model to look at gastrosophical reflux disease, acid reflux disease, it can give me some answers, and I haven't the foggiest clue whether it is correct or not, because I don't specialize in that. That is not what I do, right? I'm not a doctor.

Um I don't even play one on YouTube. And so it could tell me things that are blatantly wrong, and I won't know unless I have, you know, I take the time to corroborate that to go good do Google search on the answer and validate it from reliable sources that what it told me is correct. Um under no circumstances, particularly for high stakes stuff, should you ever just be, you know, using responses from large language models willy-nilly with no fact checking, right? In the same way that you wouldn't do that from a search engine. This is not new.

This is just a different technology now. You would not just copy paste something from the first result on Google for your query without looking at it, without reading it, without going, ah, that doesn't make sense, or ooh, I don't trust that source, you know. I was I I asked um Bing a question the other day, and it gave me a response and the citation, which is very important. Uh, the citation it gave was to a uh a known disinformation source. And I'm like, that's wrong.

Um, and I gave feedback. I said, you know, thumbs down, this is an incorrect response, it's factually incorrect. Whether Microsoft uses that information or not, I don't know. But even regular old fashioned search engines can give you incorrect responses, right? They can come up with something, they can find something that is factually just flat out wrong.

There's a greater risk in large language models because they don't do citations nearly as well as search engines do, right? When you ask Chat GPT for an answer and then you ask it to cite its sources, sometimes those sources are just made up. There was a very famous case, uh legal case not too long ago where uh a lawyer got in a lot of trouble because ChatGPT cited cases that don't exist. It looks good when he he published it, but didn't exist. So you've got to fact check these things.

Humans should be fact checking what AI produces for the foreseeable future, right? For the foreseeable future, because there's just too many ways for these tools to go off the rails. And it's much easier and safer to fact check them yourself. And if you don't have subject matter expertise in the things you're having it generate content for, A, I wonder why you're generating content on those things, and B, find someone who does have the expertise so that they can correct what the models are spitting out. It's a good question.

It's an important question, so thank you for asking it. I'll talk to you next time. If you like this video, go ahead and hit that subscribe button.