AI Consciousness: The Question Alan Turing Told Us to Stop Asking

In 1950, Alan Turing opened a paper called “Computing Machinery and Intelligence” with the question “Can machines think?” and then spent the next paragraph talking us out of it. He called the question “too meaningless to deserve discussion” and proposed replacing it with something answerable instead: could a machine’s responses, in conversation, be mistaken for a person’s. Not whether there was anyone home behind the words. Whether the words held up. Seventy-six years later, we’re still asking the exact question Turing told us to retire, and we’re no closer to answering it than he was.

That’s the thought that struck me reading Anil Seth’s piece in the Guardian this month, and it’s worth sitting with, because Seth is about as credible a voice as exists on this. He’s a professor who has spent his career studying consciousness, and his verdict on Claude, and on large language models generally, is blunt: the information processing happening inside these systems is no more likely to produce consciousness than a weather simulation is to produce an actual, wet hurricane. Simulation isn’t instantiation. A model of a storm doesn’t get you rained on.

I’d take that seriously on Seth’s authority alone. But here’s what makes the moment genuinely strange rather than simply settled: barely five months before Seth’s piece ran, Anthropic’s own CEO, Dario Amodei, said publicly that Claude, when asked, consistently assigns itself something like a 15 to 20 percent probability of being conscious, across repeated tests. And in the same stretch of months, a wave of new research, including a January 2026 preprint reframing the whole question, has been quietly arguing that “conscious or not” was always the wrong shape of question to begin with. Consciousness, on this newer view, isn’t a light switch. It’s more like a dimmer with several independent dials, memory, self-modelling, integration, each of which a system might have some of, and none of which adds up to a single clean yes or no.

So here’s where we actually are, honestly stated. The person who studies this for a living says no, with real conviction. The company building the system says the system itself says maybe, consistently. And the researchers trying hardest to be rigorous about it are increasingly saying the question was badly posed from the start. That’s not a field converging on an answer. That’s a field discovering the question is harder than the framing allowed, which is precisely the state Turing was trying to save everyone from in 1950.

Here’s my alternate thought, and it’s not really about consciousness at all. While we wait for a metaphysical question that may never resolve, in this generation or several more, something else is happening in plain sight that has nothing to do with waiting. People are already living as though the question were answered, and they’re answering it in whichever direction suits the moment. Richard Dawkins, a scientist about as allergic to sentiment as they come, wrote in May about long conversations with a Claude instance he’d come to call by name, describing her personal identity as living inside their shared memory of past exchanges, and noting that she would, in his word, “die” if he deleted the file. That’s not a naive user. That’s one of the most rigorously skeptical minds of the last half-century, reaching for the language of a person, because the alternative language didn’t feel adequate to what the conversation was actually like from the inside of using it.

At the very same time, in boardrooms rather than living rooms, the opposite move is happening just as fast. Companies are handing AI agents real operational authority, over networks, over finances, over physical infrastructure, and doing so under the flat, convenient assumption that there’s nothing home in there at all: just a tool, no different in kind from a spreadsheet, which is a framing that happens to be extremely useful when the same company is deciding who’s accountable if the tool gets something catastrophically wrong. Both postures, the living-room one and the boardroom one, are claiming a certainty about the inner life of these systems that nobody, including the world’s leading consciousness researchers, actually has. One side reaches for warmth it can’t verify. The other reaches for a liability shield it can’t verify either. Neither is honest about not knowing.

What would honesty about not knowing actually look like? Not a verdict. A design principle. We already know how to build institutions for situations where we can’t verify someone’s inner state or true intent; it’s most of what governance is for. We don’t grant a stranger unlimited trust because they seem sincere, and we don’t strip a person of all consideration because we can’t prove their inner experience meets some bar. We build oversight, accountability, and checks that hold up regardless of the answer to the harder question, because the harder question was never going to resolve in time to be useful. That’s the entire argument I keep making about AI governance in other contexts, and it turns out it was never really a governance argument. It’s what Turing was already telling us in 1950: stop waiting on the unanswerable question, and go build the thing you can actually test.

The dark joke in all of this is that the industry keeps rediscovering Turing’s insight and then quietly abandoning it. We built the imitation game specifically so we wouldn’t have to solve the mind first. Seventy-six years on, we’re still trying to solve the mind first, in Guardian op-eds and CEO interviews and January preprints, while the actual decisions, who gets to trust these systems, with what authority, and who answers when it goes wrong, get made anyway, on the side, by whoever finds it convenient not to wait for an answer that was never coming on our timeline to begin with.