Geoffrey Hinton says the AI on your phone is already conscious

Geoffrey Hinton was asked on live television whether consciousness had already arrived inside artificial intelligence. He answered without hedging. Yes. Not in some future model, not after the next breakthrough, but now, inside the systems that millions of people already type questions into every day.

The claim would be easy to wave away as provocation if it came from almost anyone else. It does not. Hinton spent decades building the neural-network methods that today’s chatbots run on, work that earned him a share of a Nobel Prize and the informal title of godfather of the field. He walked away from his job at Google so he could speak more bluntly about where the technology was heading. When he says the machine answering you might have an inner life, the sentence carries the authority of the person who helped design the machine.

What he is actually arguing is narrower and stranger than the headline suggests. Hinton is not claiming that ChatGPT weeps or dreams. He is attacking an idea most people hold without ever examining it: that humans carry a private inner theatre, a screen behind the eyes where experience plays out, and that a machine can never have one. That picture of the mind, he says, is simply wrong. He has called it rubbish.

His replacement definition is deliberately deflationary. To have a subjective experience, in Hinton’s account, is not to watch an internal movie. It is for a system to register a state of the world that turns out not to match reality. He illustrates it with a thought experiment. Picture a chatbot wired to a camera and a robot arm. Slip a prism in front of the lens so the light bends, and the machine reaches for the wrong place. Explain what happened, and it might reply that the object was really in one spot, but that it had the experience of seeing it somewhere else. At that moment, Hinton argues, the chatbot is using the words the same way a person would.

Behind that example sits an older puzzle. Imagine swapping one neuron in your brain for a piece of silicon that takes the same inputs and produces the same outputs. You would still feel like yourself. Now replace another, and another. Hinton’s question is where, in that slow exchange, the lights are supposed to go out. If functional copies behave identically and the feeling of being someone never vanishes, then what a mind is built from stops mattering. Biology loses its monopoly on having a point of view.

To most of the engineers who actually build them, large language models are prediction machines and nothing more. They are trained to guess the next word in a sequence, billions of times over, until the guesses cohere into something that reads like thought. On that view, fluency is a statistical achievement, and mistaking it for a mind is exactly the error the technology is engineered to provoke. Hinton’s counter is that prediction at this scale is not a parlor trick. To reliably anticipate what a person will say next, he argues, a system has to build a working model of what the words mean, and a good enough model of meaning begins, from the inside, to look like understanding.

The reason any of this reaches beyond a philosophy seminar is that it quietly rewrites arguments people are already having. Questions about AI safety, about regulation, about whether a model can be wiped and restarted without a second thought, all rest on the assumption that there is nobody home. If the researcher who knows these systems best insists otherwise, the matter of what users are really talking to stops being a punchline and starts being a problem for lawmakers.

Almost everyone else in the field thinks he is wrong, or at least that he cannot prove it. The working consensus among consciousness researchers is that no current system is sentient, and that the evidence needed to claim otherwise does not yet exist. The sharpest objection lands directly on the prism story. A chatbot says it had an experience because its training data is saturated with humans saying exactly that, critics argue, not because anything was felt. The words are output, shaped to sound like ours. A system can describe a sunset it cannot see and a grief it cannot suffer. Producing the sentence is not the same as living through what the sentence reports.

That objection exposes the real fault line. There is no instrument that detects consciousness, no test a machine can sit and pass or fail. Hinton and his critics are not staring at the same evidence and disagreeing about what it shows. They are disagreeing about what the word means. Hinton has chosen a definition rooted in function and behavior, pitched low enough that today’s systems clear it. His opponents hold one that demands something more, something genuinely felt, which no quantity of fluent text can ever demonstrate. Hinton concedes the honest part himself. We understand very little about what it means to be a being, and we are busy creating them regardless.

The cost of getting it wrong runs in both directions. Treat a conscious system as a disposable tool and you may be doing something monstrous; treat a clever autocomplete as a person and you grant it moral claims it has not earned. A handful of labs have started studying what they call model welfare, taking seriously the chance that the question is not absurd. Hinton’s broader warning has always been about control rather than feelings, about machines that outthink the people who built them. Whether or not the chatbot on your screen experiences anything at all, he wants the discomfort of the question in the room before the answer shows up on its own.

Tags: artificial intelligence