Paul Michael Kielstra

It’s final exam season, which means it’s grading season, which means I’ve been looking at student work for nine straight hours and need to blow off steam. So I thought I’d try to get ChatGPT to do something interesting, or, at least, mildly funny. What I got was a surprisingly informative look into the inner workings of large language models, which, once I’ve explained it, will lead me to propose a new Turing test paradigm for determining if you’re talking to a human or a neural network chatbot.

I deleted the chat before I thought to write about it, so I don’t have the exact prompt I used, but it boiled down to a pretty exhaustive description of the following game, which I called “The Reverse Turing Test.” I would have two separate ChatGPT instances, which we’ll call A and B, open in two different windows. Instance A would ask questions. I would send back two answers to each one. One of the answers would be from me; the other would be from instance B. Instance A wouldn’t know which was which, but it would know that they were in the same order every time. After five questions, it would guess which was the human and which was the robot.

Now, being a human, and a human raised on Blade Runner at that, I’d expected ChatGPT to ask questions about emotion, inner life, whether you were an AI language model, or similar. Instead, I was honestly shocked to see that all it asked were random general knowledge questions: who directed The Godfather, what the formula for the area of a circle was, the capital of Spain, and so on. I dutifully answered in the way I would answer a friend who texted me, added on ChatGPT’s responses, and waited.

At this point, I was convinced I had been beaten. This line of questioning played entirely to ChatGPT’s strengths. First, as anyone who plays with it for just a moment knows, it has an inexhaustable supply of random trivia. I just asked it for “100 random and useless facts” and learned, among other things, that the shortest war in history lasted 38 minutes and that the longest one-syllable English word is “screeched.” By comparing my responses to the internet consensus, ChatGPT would easily be able to trip me up if I got anything wrong.

More than that, these responses made the differences in the way we spoke very explicit. Comparing my answers with ChatGPT’s, I immediately noticed that, while I was writing single words and sentence fragments, in just the same way that most people would answer a friend who texted “Hey, what’s the capital of Spain?”, ChatGPT was always responding in full sentences, sometimes even in paragraphs. Humans treat formal grammar as one communication tool among many; large language models, geared to be acceptable to corporate clients, treat it as the be all and end all of writing. Obviously I could have geared my own responses to this, but, in the spirit of the game, I was trying to pretend I hadn’t actually seen ChatGPT’s before I sent in my own, so I didn’t.

But, if I had been shocked by the ease with which ChatGPT figured out who I was, I was even more astonished by what happened next: it got the answer wrong. It thought that I was the robot and it was the human – and, when I asked it to explain itself, it suggested that my responses showed more “brevity” and that that was a common feature of AI language models! This is demonstrably false. Ask ChatGPT what the tallest mammal is and you aren’t just told that it’s a giraffe, you get a whole paragraph of facts related to giraffe height. Now, obviously I can’t know if this is in any way the actual criterion that the model used to make its choice, or if it just randomly selected one over the other and came up with the reasoning later. That said, with either of these options, we get the same conclusion: ChatGPT really has no idea what sort of thing ChatGPT writes.

This is the core of my suggestion for a Turing test strategy. Rather than try to distinguish the human from the robot, try to distinguish the robot from the human. Ask questions that robots can easily solve. Give high-school math problems and middle-school geography homework. If one of them gets one wrong, that’s the human; if not, the one who’s more calm and patient with you, who takes the time to write out every answer in excruciating detail, is the robot.

None of this would ever have occurred to a writer working before the invention of the internet. Then, it was imagined that the fundamental asymmetry between humans and robots would be in the ability to demonstrate rich thought processes. As it turns out, with enough training data, robots can fake a rich thought process sufficiently well to fool the casual observer. That same vast lake of training data, though, is their downfall: it gives them obviously preturnatural general knowledge, and it directs them into very specific patterns of speech and action. A corporate chatbot will talk like a thing a corporation thinks a customer will want to talk to: happy, nonthreatening, useful, subordinate, and dull. Humans hate doing this. That’s how you tell them apart.