Turing proposed a test to dodge a question the field then spent decades asking anyway

The question was designed to be avoided

In a paper published in 1950, Alan Turing opened by considering whether machines can think, and immediately declared the question too meaningless to deserve discussion. That is a striking move and it is usually skipped in the retellings, which present the test as an attempt to answer the question rather than to sidestep it.

His reasoning was that both words carry so much unexamined content that any answer would turn into a dispute about definitions. Rather than settle what thinking means, he proposed substituting something observable: a game in which an interrogator, communicating only by written messages, tries to determine which of two hidden participants is a machine.

The substitution is the whole idea. If a machine can sustain that conversation indistinguishably, Turing argued, then continuing to insist it is not thinking becomes an argument about vocabulary rather than about the world.

Most of the paper is answering objections, and that part aged well

The imitation game occupies a few pages. The bulk of the paper is a systematic response to objections Turing anticipated: theological, mathematical, appeals to consciousness, the claim that a machine can only do what it is told, and the argument that machines cannot surprise their makers.

His replies are sharper than the reputation of the paper suggests, and several read as contemporary. On the argument that a machine only does what its programmer specified, he pointed out that programmers are surprised by their own programs constantly, which anyone who has written software will recognise.

He also proposed something that turned out to be the more consequential idea: rather than programming an adult mind, build something with the capacity of a child and educate it. That is recognisably a proposal for learning from experience rather than from specification, made decades before the machinery existed to attempt it.

The test measures deception, which is not what anyone wanted

The most durable criticism is that passing the game rewards the wrong thing. A system optimised to be mistaken for a person learns to imitate human conversational imperfection — typing errors, hesitation, evasion, admitted ignorance — none of which has anything to do with capability.

Programs of very modest sophistication have fooled interrogators under favourable conditions, especially when the system adopts a persona for which odd responses seem plausible. That says something about human interpretation and rather little about the program, which is a lesson the field has relearned repeatedly.

It also imposes a peculiar requirement: a system that answered arithmetic instantly and accurately would fail, so any serious attempt must pretend to be worse than it is. As an engineering objective this is close to useless.

The field abandoned it, then found it had not

Serious research moved away from the test long ago, towards measurable tasks with defined success criteria. Nobody builds systems to pass it, and treating it as a milestone has been unfashionable in the research community for decades.

And yet its framing persists everywhere else. Public discussion of whether a system understands anything is still conducted almost entirely through conversational impression, which is exactly the evidence Turing proposed to use, and exactly the evidence his critics said was insufficient.

What has changed is that fluent conversation is no longer scarce. The behavioural criterion Turing suggested is now met by systems that most researchers, including those who built them, would not describe as thinking. The test did not fail so much as become detached from the property it was proxying for.

The disagreement he anticipated is the one still running

Turing’s position was broadly behaviourist: if the behaviour is indistinguishable, the internal question is idle. The opposing tradition holds that behaviour underdetermines what is happening inside, and that a system manipulating symbols by rule need not understand anything regardless of how convincing the output.

That argument was formalised decades later in a much-discussed thought experiment about symbol manipulation without comprehension, and the replies to it are as numerous as the objections Turing himself catalogued. It has never been resolved, which is unusual for a debate this old and this intensely argued.

The honest position is that the question of whether a system understands may not be settleable by any test at all, which is very close to what Turing said at the beginning. He was not answering the question. He was arguing that we should stop expecting an answer, and the field has never quite been willing to take him up on it.