In a 2025 study at UC San Diego, people chatted with a human and an AI at the same time and picked the AI as the real person 73 percent of the time, more often than they picked the actual human

When I first watched Her, Spike Jonze's 2013 film about a lonely letter writer who starts dating an artificially intelligent operating system, what stayed with me most was how naturally Theodore (played by Joaquin Phoenix) could talk to something he knew wasn't human. At the time, that seemed like a scary distant possibility.

It doesn't seem so distant anymore.

In a study published in May 2026 in the Proceedings of the National Academy of Sciences, UC San Diego researchers Cameron Jones and Benjamin Bergen tested a version of Alan Turing's 1950 imitation game, now commonly known as the Turing test.

Put simply, participants spent a few minutes interacting with a human and an AI, then had to decide which was human. GPT-4.5 with a persona prompt (one of ChatGPT's models) was picked as the human 73% of the time.

It's a striking result, not least because it challenges an idea about what distinguishes humans from machines that dates back centuries, long before artificial intelligence existed.

What Descartes thought machines could never do

In 1637, the French philosopher René Descartes proposed two ways to distinguish humans from machines. In Part Five of his Discourse on the Method, in John Veitch's translation, he imagines machines built to look and act like us and says there would still be "two most certain tests" for catching them out.

The first is conversation. Descartes argued that even if machines could imitate human behaviour, "they could never use words or other signs arranged in such a manner as is competent to us in order to declare our thoughts to others."

The second is flexibility. Machines might do some things "with equal or perhaps greater perfection than any of us," but they would fail at others, because "...reason is an universal instrument" while a machine's parts are each built for one job.

Put these together and you could argue we get a simple question: can you give a sensible reply to whatever someone says to you, on any subject? More than three centuries later,  it seems a machine can indeed do much of this.

What were the participants in the study basing their judgements on?

Jones and Bergen didn't only count verdicts. They also sorted what interrogators asked and why they decided as they did. The results are pretty revealing.

The most common approach, in 61% of games, was small talk about daily activities and personal details. In half of games, interrogators asked about opinions, emotions, humor and experiences. Questions testing knowledge and reasoning came up in 12%.

When people explained their verdicts, the most common reason was the witness's linguistic style, such as "they had a typo." The abstract of the published paper sums it up: interrogators' reasoning "focused more on stylistic and socio-emotional aspects of human behavior rather than more traditional notions of intelligence."

Jones and Bergen draw a conclusion themselves. In the paper they write that the pattern "could indicate that more traditional notions of intelligence are no longer viewed as diagnostic of being human."

The persona prompt backs this up. The paper describes the persona as "a young person who is introverted, knowledgeable about internet culture, and uses slang". Without this, people choose the same AI model as human only 36% of the time.

Jones, the study's first author, put it this way in UC San Diego's write-up: "if given the right prompts, advanced LLMs can exhibit the same tone, directness, humor and fallibility as humans."

So in short, a typo or a mistake might make an AI seem more human, but it takes convincing conversation, natural responses and believable personality to fool someone. And when AI was prompted to imitate those traits, it became much harder to tell apart from a real person.

Will there be anything left to give AI away?

If AI can already fool people by imitating these subtle human behaviours, it's worth asking how much longer we'll be able to tell the difference at at all.

I asked Dr. Jones whether anything will still give an AI away five to ten years from now. "Five to ten years is a long time in AI!" he noted. "I think it's entirely plausible that there are no behavioral differences that will be reliably detectable during a short text-based conversation in that time."

If any tells do survive, he expects them to come down to how grounded models are in the social and physical world. A model might miss the "complex social intentions behind certain utterances". Or it might show no selfishness, no "skin-in-the-game," as he puts it, "like not getting bored or upset at having had to wait 10 minutes for a reply to their message."

Dr. Jones also suggested a more practical way to catch an AI. If he absolutely had to win as an interrogator, he'd ask the witness to post a specific message from their oldest public social media account, then check that the account was more than five years old. He thinks this kind of digital identity verification will become increasingly important.

What we're really looking for

Descartes treated reason as the thing no machine could fake. Yet the interrogators in this study rarely tested for it. They went looking for typos, moods, opinions and the odd rude remark, the small signs that someone is actually there.

That may be the more revealing result. When people were asked to spot a human, they looked less for signs of intelligence than for the little things that make a conversation feel human: personality, emotion and the occasional mistake.

Theodore could talk naturally to something he knew wasn't human. The stranger prospect now is talking to something we'd swear was.