When ChatGPT Goes Astray: Understanding AI Errors

0:00ChatGPT is incredibly powerful when it comes to writing content. It's super fast. And in fact, if you upgrade to their premium version, I've heard people say it's up to 60 % faster. So this thing is incredibly powerful, but it does have a big flaw that a lot of people talk about, and that is accuracy. So ChatGPT is notoriously bad at writing fake facts. And the way that this typically works is that someone will go to it and say, ChatGPT, write me a story, or it'll say, like, here's my, like, theory on my, you know, paper I'm writing. Write me about why this theory is correct and give me, like, five stories to back it up.

0:38or tell me a story about this person in this situation that backs that up. And ChatGPT will just, you know, it's doing what you told it to do, so we'll just go for it even if that is completely fake. The news or the information that you're giving to it is completely fake. And so the way that ChatGPT works is essentially it is guessing, it's writing what you tell it to write first off, but it's guessing what the next word in a sentence or the next phrase is supposed to be based off of just context and what it's been trained on. And so, yeah, it's just notoriously bad at coming up with fake information.

1:09So today on the podcast, we are going to dive into why it does that, how to fix that, some of the implications of ChatGPT generating fake information and everything you need to know about that. So one of the main reasons that it generates this kind of info is because it just fundamentally, it lacks the ability to discern truth from fiction. So it's not trained on factual information necessarily. It's more just trained on how to create content, how to talk, how to say things. It's a language learning model. So if you tell it to say something, it will usually repeat that. Another reason is because it's designed to prioritize generating responses that are likely to be engaging to the user.

1:51So this means that it might sometimes sacrifice accuracy in favor of producing the results that are more interesting or attention grabbing, or honestly, more just more likely for the user to like them. So if the user asks for something, it's going to give them that. Finally, ChatGPT is going to generate some fake responses because it has not been programmed to recognize certain types of misinformation. So for example, if you ask ChatGPT for information about something fake, it's going to generate a response that appears to support your theory, even if that theory has completely been debunked by experts, blah, blah, blah.

2:28lot. And another problem I would say with it is that it has, it's not attached to the internet. So it's been shut off from, um, it's been shut off from like current information. And so another problem is whether you're talking about medical things or, um, uh, you know, facts, science, et cetera, if there's like a new theory or a new discovery that has come out in recent years, um, chat GPT doesn't have the access to find that and get updated information. It just knows whatever information it was trained on that it was originally plugged into. So obviously, that is pretty problematic. So I'd say it's pretty important just to acknowledge the limitations of ChatGPT.

3:04It's not a perfect tool. And it's not going to generate 100 % accurate information all the time. So instead of relying on ChatGPT to provide you with 100 % accurate information, I would say, use it to kind of give you an outline and make sure that you are fact checking the responses that it generates. I've done this with different papers, different essays, different long form blogs, and those types of pieces of content. And it's done a really good job of writing me stories that go along with my theory or thesis. But it wasn't very good at coming up with very factual pieces of information is specifically when it comes to numbers, dates, and locations.

3:44I found it wasn't super good at that. This obviously is not a supplement for critical thinking. So I would say when it creates content, make sure you read it, make sure you fact check it. It can give you really good ideas and direction overall. But for me in particular, I've had it create content and then I go back and make sure that all the dates, all the times, and even the story is factual. And it saves me tons of time creating my content, but I do need to go make sure that all the particulars are like actually true. So as we go forward, I think it's gonna be pretty important to continue to invest in research and development, specifically aimed at improving the accuracy of these language learning models like ChatGPT.

4:28I think this is going to involve first off developing new techniques for training and evaluating the models, as well as probably building new tools and systems to really integrate fact checking and source verification. And we're seeing tools like Google Bard and Microsoft Bing trying to integrate more of those source sourcing kind of tools. So like showing you where they got the information from. So at the same time, I think it's going to be pretty important to recognize that the development of more accurate language models is only one piece of the puzzle when it comes to building a more reliable and trustworthy information ecosystem.

5:04I think we also need to invest in media literacy education. And I mean, okay, I guess that's a fancy way of saying we really just need to invest in making sure people understand to be critical and to use critical thinking when they're receiving any bit of information, whether that's from ChatGPT or anywhere else. Make sure you fact check things, make sure you verify things, make sure you're getting it from credible sources and whatnot. So to conclude everything here, ChatGPT is very powerful. It creates great outlines. It's not perfectly factually accurate on everything it says, but with some fact checking and some verifying, you can get some pretty incredible content out of it.

5:45Nash Falls is the relentless new thriller from number one New York Times bestselling author David Baldacci. When mild-mannered business analyst Walter Nash is recruited by the FBI to help bring down a global crime network, his life has turned completely upside down. Publishers Weekly says, in Baldacci's long list of heroes, Walter Nash is among the most memorable. Experience the full cast audio production of Nash Falls, led by McLeod Andrews. Available from Hachette Audio wherever audiobooks are sold. Also available in hardcover and e-book.