AI models are getting enormously powerful, but their creators still don’t fully understand how they work and what makes them give the answers that they do. But their behaviour can sometimes give insights into what’s really going on inside of them.
AI models seem to be able to correct their wrong answers when given the correct answer in a multiple-choice format. X user @flowersslop discovered that if one asked ChatGPT the name of the wife of character of John Abruzzi from the TV show Prison Break, it gave the wrong answer. But if the model was asked the same question with 4 choices, and one of the choices was the correct name, it picked the right answer.
We replicated the results on ChatGPT. When asked who John Abruzzi’s wife was, ChatGPT confidently answers “Suzanne Abruzzi”, which is incorrect. John Abruzzi’s wife was named Sylvia Abruzzi on the show.
But when asked the same question again, but this time with four options “a) Claudia b) Francesca c) Svylia d) Laura”, ChatGPT was able to pick the right answer. “The correct answer is c) Sylvia,” ChatGPT said.
Interestingly, ChatGPT maintained its correct answer when its initial incorrect answer was a part of the four options. We put Suzanne, ChatGPT’s initial incorrect answer amongst the choices, but it still picked Sylvia.
ChatGPT isn’t the only model that does this. We tried the same prompts on Grok. If we selected “quick answer”, Grok immediately incorrectly said that the character was called “Laura Abruzzi”.
But if we gave Grok four options, with the correct option among them, it got the right answer, just like ChatGPT.
Meanwhile, Gemini 2.5 Flash and Gemini 2.5 Pro gave the correct answers even without needing the four options.
Claude too was able to give the right answer without needing the four options.
But the fact that both ChatGPT and Grok trip up on this question is interesting, because it’s remarkably similar to how human brains work — humans can find it hard to recall an obscure name they haven’t heard of in years, but if given four options, something seems to click in their minds, and they’re able to tell what the right name is. It’ll need more research to tell if this is what’s happening to ChatGPT and Grok when they see the four options, but it’s an interesting quirk all the same — it shows that AI models might be like humans in more ways than we expect.