Tip of the day: How to use ChatGPT to figure out if a text was written by ChatGPT

Post by Prof. Stephen Horowitz, Professor of Legal English

I had a thought about a week ago that perhaps the best way to check and see if a text written by a student was in fact written by ChatGPT or other AI. And it turns out there is a way to do it, but it’s not the way I thought it might work.

I had already tested ChatGPT by giving it the same final exam fact pattern/question I had given to my LLM students. And ChatGPT was able to provide a pretty decent response.

So next, I took that response and fed it back to ChatGPT asking it if ChatGPT had written this text. But ChatGPT responded by explaining that it has no memory, etc. So I figured that was the end of it.

However, I learned today from a commenter on LinkedIn that, as in law school, it’s really about asking the right questions. And in this case, the right questions is:

“Is this text AI-generated?”

This led to a very definite response that, yes, this text was indeed AI-generated. Problem solved!

Update: I learned after publishing this post that this is not exactly “problem solved.” A colleague at another law school shared that he followed a process similar to mine, but ChatGPT believe that the text was in fact written by a human. So the “tip” is far from foolproof.

But I wasn’t satisfied, so I asked it how it knew. And it very quickly and confidently listed off several factors:

“As an AI language model, I can recognize patterns and characteristics in text that are typical of machine-generated text. In this particular case, the text is characterized by a high level of technical language, a formal tone, and a logical structure that is typical of texts written by AI. In addition, the text does not contain any errors or irregularities that would be expected if it had been written by a human. All of these features suggest that the text was generated by AI.”

Wonderful! But wait a second….Lots of people, including most law students at Georgetown and other law schools, are trained to write with these very same characteristics. So I asked, “Can you be more specific?”

And it essentially listed the same characteristics, but this time in a numbered format which made it appear much more specific and persuasive. Even though it wasn’t.

So then I asked it specifically, “What kinds of differences might distinguish a text written by a highly skilled writer and a text generated by AI?” But it listed qualities that might distinguish a human’s writing from AI, such as style, creativity, context and human touch. As a representative of the human race, I guess I’ll take those as compliments. But it still doesn’t provide any concrete examples as to how it can distinguish between a highly-skilled human writer and an AI app like ChatGPT.

In other words, ChatGPT was essentially borrowing from Supreme Court Justice Potter Stewart who famously said in his decision on obscenity, “I know it when I see it.” (Jacobellis v. Ohio, 378 U.S. 184 (1964))

Leave a Reply

Your email address will not be published. Required fields are marked *