What Is the Turing Test — and Has Any AI Actually Passed It?

The Turing Test is a way of asking whether a machine can hold a conversation convincingly enough that a person can’t tell it apart from a real human. A judge chats in text with two hidden partners — one human, one machine — and simply has to guess which is which. If the judge can’t reliably tell, the machine is said to pass. It’s not a measure of a computer’s raw processing power or its ability to do math; it’s a test of whether its conversation feels human.

The imitation game

The test comes from a 1950 paper by British mathematician Alan Turing, titled “Computing Machinery and Intelligence.” Turing was trying to sidestep the fuzzy question “Can machines think?” — a question he considered too vague to answer usefully — and replace it with something you could actually run as an experiment.

His original version, which he called the “imitation game,” had three participants: a man, a woman, and an interrogator who could only communicate with the other two through written notes. The interrogator’s job was to figure out which was which; one player tried to help, the other tried to deceive. Turing then asked: what happens if a machine takes the deceiving player’s place, and tries to convince the interrogator it’s the woman? If the machine succeeds about as often as the man did, Turing argued, that success is worth taking seriously as evidence of “thinking.”

Over time, the test was simplified into the version most people picture today: one evaluator, one hidden human, one hidden machine, and a plain question — which one is the computer?

Has any AI actually passed it?

For decades, the answer was no — or at least, not convincingly. Early chatbots occasionally fooled judges in informal settings, most famously a program called Eugene Goostman in 2014, but critics pointed out that short conversations, distracted judges, and low-effort test designs made those wins easy to dismiss.

That changed with the rise of modern large language models. In a controlled study published in March 2025, researchers Cameron Jones and Benjamin Bergen at UC San Diego ran a rigorous three-party Turing test on four systems: the 1960s chatbot ELIZA, GPT-4o, Meta’s LLaMa-3.1-405B, and OpenAI’s GPT-4.5. Each system was given a persona prompt — instructions to act like a specific kind of person — and judged against a real human in five-minute conversations.

The results: GPT-4.5 was judged to be the human 73% of the time, meaning interrogators picked it over the actual human participant more often than not. LLaMa-3.1 scored 56%, roughly a coin flip. Without the persona prompt, both models did far worse — GPT-4.5 dropped to 36% — showing that how a model is instructed to behave matters as much as its raw capability. The researchers called it the first empirical evidence of any system passing a standard three-party Turing test.

Why critics say that doesn’t settle much

Passing the test has never meant “the machine thinks” to everyone’s satisfaction, and the Turing Test has been contested almost since it was proposed.

The best-known objection is philosopher John Searle’s 1980 Chinese Room argument: imagine a person who doesn’t speak Chinese, sitting in a room, mechanically matching Chinese symbols to instructions in a rulebook well enough to produce fluent-looking answers. From outside, the room “understands” Chinese. Inside, no one does. Searle used the thought experiment to argue that manipulating symbols convincingly — which is what a language model does — is not the same as understanding meaning.

A related problem is the “ELIZA effect,” named after a 1966 chatbot that mimicked a therapist by reflecting users’ own statements back as questions. ELIZA understood nothing, yet people confided in it. The lesson researchers drew was that humans are quick to project understanding onto a system that merely sounds right, which makes “fooling a judge” a weaker bar than it first appears.

The test also has structural blind spots: it only measures conversational behavior in one language and format, it rewards a machine for hiding capabilities that would give it away as superhuman (like instant arithmetic), and a skilled or skeptical interrogator gets very different results than a casual one.

Why it still matters

Despite the criticism, the Turing Test has outlasted almost every other benchmark proposed for machine intelligence, because it captures something real: the point at which a machine’s output stops feeling mechanical. It’s become the common cultural shorthand for “an AI good enough to pass as a person” — even researchers who reject it as a scientific measure of intelligence still use “it would pass the Turing Test” as an informal yardstick.

It’s also a useful contrast to more ambitious ideas like artificial general intelligence (AGI), which asks whether a machine can match human ability across many different kinds of tasks, not just conversation. A system can pass a Turing Test — convince a judge it’s human in chat — without being anywhere close to AGI, and AI research increasingly treats the two as separate questions: one about believable behavior, the other about general competence. As artificial intelligence systems keep getting better at sounding human, the test’s real legacy may be less “did it pass” and more the vocabulary it gave everyone for talking about that moment at all.