The Turing test

What it tests Is behaving indistinguishably from a thinking person in conversation enough to count as thinking?

The case

In 1950 Alan Turing opened his paper “Computing Machinery and Intelligence” with the question “Can machines think?”, and then set it aside as too vague to be worth answering. In its place he proposed a game. An interrogator exchanges typed messages with two hidden players, a person and a machine, and must say which is which. If the machine is picked out no more reliably than chance, it has, in Turing’s terms, done well at the imitation game.

Turing predicted that in about fifty years machines would play well enough that an average interrogator would have no more than a 70 per cent chance of identifying them correctly after five minutes of questioning.

What it is meant to test

The test replaces an inner, hard-to-define property, thinking, with an outward and checkable one: conversation indistinguishable from a person’s. Its attraction is that we grant thought to other people on the same kind of evidence. We never observe their thinking directly; we infer it from what they say and do.

Has it been passed?

For decades no system came close, and contests run under looser rules produced only brief, contested successes. That changed with large language models. In a pre-registered study published in 2025, Cameron Jones and Benjamin Bergen at UC San Diego ran standard three-party tests: participants chatted for five minutes with a human and a machine at the same time and then chose which was the human. GPT-4.5, prompted to adopt a humanlike persona, was judged to be the human 73% of the time, significantly more often than the real humans. LLaMa-3.1-405B with the same prompt was chosen 56% of the time; the older chatbot ELIZA and GPT-4o without the persona prompt were chosen 23% and 21% of the time. The authors describe this as the first empirical evidence that any artificial system passes a standard three-party Turing test.

The objections

Passing is not thinking. John Searle’s Chinese room imagines a person who passes a Turing test in Chinese by following rules, without understanding a word. If he can pass without understanding, passing does not show understanding.

The blockhead. Ned Block imagined a machine that simply looks up a stored, sensible reply for every possible conversation of a given length. It would pass, yet it plainly does not think. So whether a system thinks depends on how it produces its behaviour, not only on the behaviour.

Turing’s own replies. Turing anticipated many objections, among them that machines cannot be conscious, cannot do anything new, or merely follow their programs, and answered each in the paper. His most important point was that we apply the same behavioural standard to other people, and refusing it to machines needs a reason.

The wrong target. Others object that the test measures one thing, the ability to pass as human in short text chats, which is neither necessary nor sufficient for intelligence. A system could be highly intelligent and unhumanlike, or humanlike and shallow.

What passing does and does not settle

The 2025 result settles the empirical question Turing posed: under standard conditions, people can no longer reliably tell some language models from people in short conversations. It does not settle whether those models think, understand or are conscious, because whether behaviour suffices was always the contested question. It does change the practical situation. If you cannot tell from a conversation whether you are talking to a person, then what you can learn from the conversation, and whom you can hold to what was said, has to be settled some other way. That is the subject of Can AI testify? and Should you trust what ChatGPT tells you?. You can test your own criteria for minds in the lab’s Mind Detector.