You Can’t Hide Your Lyin’ AI

Believe Me, I’m AI!

Recently I asked ChatGPT a question that should not be too difficult. I wanted a list of the 10 least populous states to have been represented by a leader who went on to become President of the United States. Yes, I know, it’s a weird and convoluted request, but sometimes I just like to gather knowledge that might be useful in a future trivia night.

The question is slightly tricky for a human to answer and might involve some tedious research. But for a computer armed with a large enough and relevant enough dataset, it’s a breeze. Here’s one way a computer could approach the problem. First, find the list of presidents. Then find the list of states they represented as a political leader. Finally, cross-reference those states against a list of states in ascending order of their population. Once you hit 10 states, you are done.

ChatGPT confidently rattled off its answer, and even though I’m a lowly human, I was pretty sure it was wrong. The list excluded Arkansas. Now I didn’t know EXACTLY where Arkansas is ranked (it turns out it is the 18th least populated) but I knew it didn’t have THAT many people, and I also knew that Bill Clinton was once its Governor.

It turns out ChatGPT was wrong in this case, but also that it’s wrong quite often. It can get particularly confused when performing certain tasks. For example, try getting it to invent a mind-reading game, and you might see it collapse in its own logic, all while seeming as confident as ever.

And that is the heart of the challenge with this form of AI. Aside from some specific situations where they know their data is out of date, these systems are just not that self-aware. Ethan Mollick recently described these AI systems as being like “an intern that lies a little bit” – and I think this a great analogy – systems like ChatGPT are useful, lack some experience, and you cannot always trust what they say.

And of course, that is not the only problem. Several studies have illustrated that AI systems often display forms of systemic bias. As disturbing as this is, it shouldn’t be surprising – after all, if the data used to train the models is biased, the AI will typically be as well. As we used to say – Garbage In, Garbage Out.

If you have concerns about where all this is heading, you are not alone. Right now, AI-generated content is flooding the Internet. Some early studies into AI systems have shown that humans rapidly trust their AI counterparts, sometimes more than the humans they interact with. The more humans trust AI, the less effort they will make checking the information produced by it. Combine that with studies that show false information tends to spread faster than true information on social media, and there is at least a potential tsunami of bias and falsehoods that dwarfs what we have today.

All of this might sound pretty dystopian, but it’s also true that the very technology that could spread misinformation can guard against misinformation. I hope that very soon we will all have access to AI-based systems that reliably determine the level of confidence in all content being produced, whether it is from AI or a human. This will help us figure out how much we should trust the information we see. Or as a president from the MOST populous state once said “Trust but Verify”