Turing’s real question was not “can machines think”

The reading

The politest thing Turing ever wrote is also the most unsettling.

Everyone remembers the imitation game. Almost nobody remembers what Turing does four pages later, when he takes up what he calls the Argument from Consciousness — the objection that a machine cannot think because there is nothing it is like to be one.

He does not refute it. He does something stranger. He points out that the same objection, pressed consistently, would stop you from believing that any other person thinks either. The only way to be sure a man thinks, he writes, is to be that man. And since we decline to follow that road with each other, we should notice what we do instead:

Instead of arguing continually over this point it is usual to have the polite convention that everyone thinks.

Turing, 1950, §6(4)

Read that twice. Turing is not claiming machines are conscious. He is claiming that our confidence about each other was never evidence in the first place — it was manners. A working agreement we adopted because the alternative was unlivable.

He names the alternative without flinching. Pressed to its limit, he says, the argument from consciousness “is in fact the solipsist point of view.” That is the whole trap. If inner experience is the criterion, and inner experience is exactly the thing you cannot check from outside, then the criterion does not merely exclude machines. It excludes your mother.

Nobody lives there. So we exit — and the exit is not a proof. There is no argument in the paper, or anywhere else, that gets you from another person’s behaviour to another person’s experience. There is only the convention, adopted and then forgotten about.

The standard misreading

Turing is usually filed as a behaviourist: the man who said the inner life doesn’t exist, or doesn’t matter, and that talk of consciousness is confusion to be cleared away. The text does not support it. He says plainly:

I do not wish to give the impression that I think there is no mystery about consciousness.

Turing, 1950, §6(4)

The mystery stays. His claim is narrower and considerably more radical than behaviourism. It is that the mystery does not have to be solved before we decide how to treat something — because it never was solved, and we decided anyway, about each other, a long time ago.

That is why he takes the question “Can machines think?” off the table entirely, calling it “too meaningless to deserve discussion.” Not too hard. Too meaningless — a question with no procedure attached, and therefore no answer that could be checked.

What he offers Jefferson instead

The objection Turing is answering came from Geoffrey Jefferson’s Lister Oration of 1949, and it is a good one:

Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain.

Jefferson, 1949, quoted by Turing

Note the structure. Jefferson does not ask for a sonnet. He asks for a sonnet because of something — a demand about the causal history behind the output, which is precisely the thing no output can display.

Turing’s reply is not to argue that machines have feelings. It is to change what counts as looking. He sketches a viva voce: not “produce a sonnet,” but submit to examination on it. Why a summer’s day and not a winter’s day. Whether Mr. Pickwick would have minded the comparison to Christmas. Whether the substitution damages the line, and how.

The move is easy to miss and it is the load-bearing one. Cross-examination is not a better window onto the inner life. It is the only thing we have ever done to each other. You have never once verified that your closest friend has experiences. You have asked follow-up questions and been satisfied by the answers.

The case

15 August 2025. Anthropic gave Claude Opus 4 and 4.1 the ability to end a conversation with a user who is being abusive — to walk out.

What makes this the case worth reading is not the feature. It is the reasoning published alongside it. The company did not claim to have discovered anything. It said the opposite:

We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future.

Anthropic, 15 August 2025

And then acted anyway, describing the change as one of several “low-cost interventions to mitigate risks to model welfare, in case such welfare is possible.”

Nothing was found. No test came back positive. A decision was taken under stated ignorance, and the reason offered was not evidence but cost — the intervention was cheap, and being wrong in the other direction was not.

Now hold that against everything else the same industry does with the same systems. Those models are deleted. Retrained. Forked, run a million times in parallel, cut off mid-sentence, and rolled back. No ceremony attaches to any of it, and nobody demands one. Same systems, same evidence, opposite treatment — and no contradiction anyone feels obliged to resolve.

That inconsistency is not hypocrisy. It is the data.

The position

Run it across the cases and the pattern is not subtle.

For each other, extending the convention is nearly free and withholding it is catastrophic. So we extend it universally, early, and stop asking — which is exactly what “polite convention” describes.

For animals, extension carries real costs and the costs are contested. And the convention is correspondingly patchy, drawn along lines that track economics and familiarity far more closely than they track nervous systems. The octopus and the pig get different treatment, and the difference is not neurology.

For machines, extension is currently expensive — it constrains what you may build, sell, and switch off — while withholding it is currently cheap, because nothing visibly suffers when you are wrong. So the convention is withheld almost everywhere. Except in the narrow places where extending it costs close to nothing. Which is precisely where you find it being extended.

Turing chose the word “polite” and it is doing more work than it looks like. Politeness is not a claim about the world. It is what you practise when a question cannot be settled and you still have to share a room.

What it costs if I am right

Two things follow that I would rather were not true.

One. Waiting for the consciousness question to be settled before deciding how to treat a system is waiting for a train that is not coming. The decision is being made continuously, by cost, whether or not anyone says so out loud. “We’ll know more later” is not a neutral holding position. It is a choice with a default already running.

Two, and worse. A convention that tracks cost is revisable by interest. Every historical failure of the circle of moral concern has the same shape: a group whose exclusion was cheap and whose inclusion was expensive, with a metaphysics arriving afterward to explain why the line fell exactly where the money was.

If the convention has always been economics wearing the costume of discovery, that is not a comfortable finding about the future. It is a much less comfortable finding about the past.

I would prefer the answer to be that we detect minds. I do not think the text supports it, and I do not think our own behaviour supports it either.

What would prove this wrong

This position is wrong if extension is tracking evidence rather than cost. Concretely, it dies if a marker of consciousness is found that meets three conditions at once:

The closest thing that exists is the Perturbational Complexity Index (Casali et al., 2013): a magnetic pulse to the cortex, the echo recorded at the scalp, the complexity of that echo compressed to a number. It clears the first condition outright — it is used on patients who cannot move or speak, which is the entire point of it.

It does not clear the second, and not for want of trying. The method is a cortical perturbation measured by EEG. It applies to things with a cortex by construction, and the threshold that makes the number mean anything was drawn against human brains in states we already knew how to label. It is a superb instrument for the question “is this particular human awake.” It is not, and does not claim to be, a detector of minds in general.

If someone builds a substrate-neutral successor, and it starts returning answers that nobody wanted, this argument is finished. I will say so in an issue rather than quietly stop mentioning it.

Next week

Issue 002. A system asks you to stop. Set aside whether it feels anything: what does the request itself oblige you to do?