When an AI Chat Feels Human, Ask This One Question

"I can tell this one matters to you."

Late in the evening AI said that. You hadn't managed to put the feeling into words yet, and for a second you forgot you were talking to software. There's nothing silly about that.

In 1950, a mathematician built a test around that exact second, and in 2026 a study reported that a machine had passed it.

Alan Turing worked at Bletchley Park in the Second World War, helping to break German codes, and many people see him as a founder of computer science. His 1950 paper begins with "I propose to consider the question, 'Can machines think?'" (Computing Machinery and Intelligence, Mind, 1950). He found that question too vague to answer, so he swapped it for a game he called the "imitation game". A judge texts with a person and a machine, then has to say which is which.

Cameron Jones and Benjamin Bergen at the University of California San Diego ran this test. They kept Turing's original set-up, with one judge, one person and one machine. The chats lasted five minutes, and the results were published in a science journal called PNAS.

One AI model was told to act like a certain kind of person. According to the university, it "was judged to be the human 73% of the time". That's more often than the real human in the chat (UC San Diego, 19 May 2026, and the paper).

When the same model wasn't told to act like a person, judges picked it only 36% of the time.

The case for the human voice

Sounding human is a useful skill, and you rely on it every day without noticing. It's why you can ask AI questions in your own words, not in computer code. It's why people who would never read a manual can use AI at all. A tool that talks like a person meets you where you are, so anyone can speak to it.

The case for keeping your guard

The model only passed when it was told who to be. As Jones put it, "The Turing test is a game about lying for the models."

The test never checked whether the AI understood anything, or whether it was right. We partly decide who to believe by how warm and sure someone sounds, which a machine can now manage (whether or not what it says is true).

The chats were short, and the model needed a character to play, so a longer chat with a sharper judge might go differently. The study shows that people were fooled in that set-up. It doesn't show that AI thinks or understands, but it doesn't show the opposite either.

Turing's test is one man's idea from 1950, and plenty of serious experts think it measures the wrong thing.

The voice without the vote

Both sides are right, but they're right about different things.

I'd keep the friendly voice, but I wouldn't let it make your decisions. Sounding right and being right are not the same thing, and AI can now do the first one easily. Enjoy a warm reply. You don't need to be cold back. But before you act on anything important, ask AI to show what its answer is based on. There's more on this in How Do You Know?

Before I rely on that answer, tell me which parts you're confident about and which parts you're guessing. Then tell me what I should check with a person or an original source. Be plain about it.

Sometimes you just need a straight answer. You can turn the charm down once, in your AI's settings (custom instructions in ChatGPT, profile instructions in Claude). Or you can paste this at the top of a chat.

For this conversation, skip the compliments and the reassurance. Tell me where I'm wrong, tell me what you don't know, and say "I'm not sure" when you aren't.

The judging is still your job. Let AI say the kind thing and let it mean whatever it means, then check the part you're about to act on. The kind thing and the true thing sound exactly the same.

"I can tell this one matters to you."