The question of the moment

Could a machine be conscious?

For seventy years it was a thought experiment. Now there are machines on every phone that say they understand you, apologise when they are wrong and describe how they feel. Are any of them conscious, could a future one be, and how would we ever tell?

In a sculptor's studio a man in a blue tunic embraces a marble statue of a woman on a plinth. Her legs are still white stone, but her back and arms have turned to living flesh. A cupid aims an arrow from a cloud.
Jean-Léon Gérôme, Pygmalion and Galatea, c. 1890. In Ovid's story a sculptor falls in love with his statue and the gods bring her to life. Gérôme paints the moment of change: her feet are still marble, her waist is already warm. Where, in a made thing, does the life begin?Metropolitan Museum of Art, New York. Public domain, via Wikimedia Commons.

Why it suddenly matters

In 1950 Alan Turing predicted that machines would one day hold conversations good enough to fool us. Large language models now do, much of the time. They are trained on vast amounts of human writing to predict what text comes next, and the result can discuss philosophy, write poems and talk about its own inner life. In 2022 a Google engineer, Blake Lemoine, went public with his belief that the company's chatbot LaMDA was sentient. Google rejected the claim and later dismissed him, but the question has not gone away. Millions of people now talk to chatbots daily, and some feel that something is talking back.

There is a catch that Turing did not face. A chatbot learns to talk about experience from the experiences people have written down. So when it says "I feel", that is exactly what we would expect whether or not anything is felt. The philosopher Jonathan Birch calls this the gaming problem: a system trained on human talk can pass behavioural tests of feeling without that telling us anything. The usual evidence, what a being says and does, is the evidence these systems were built to reproduce.

That forces the question back onto the theories. If behaviour cannot settle it, what would? Here the views on this site split sharply, and they split over the machines in ways that reveal what each theory really thinks consciousness is.

Ask the theories

Choose a machine and see how seven leading views would judge it. The verdicts are our summaries of each position, not quotations.

Choose a machine.

Notice the pattern. Nobody here thinks today's chatbots are good candidates, though Chalmers will not rule it out. The real disagreement comes with the simulated brain, a machine that does everything a brain does. Functionalists such as Dennett, Graziano, Chalmers and the global workspace theorists say it would be conscious. Searle, integrated information theory and Seth say it would not, because they think consciousness depends on what a system is physically made of, or on its being alive, and not only on what it does.

Dennett's answer

Dennett never doubted that a machine could be conscious in principle. We are machines ourselves, he liked to say, made of cells that are tiny robots, and nothing magical happens when enough of them work together. In the 1990s he took part in Rodney Brooks's Cog project at MIT, which set out to build a humanoid robot that would learn as an infant does, and he argued that if it succeeded there would be no further fact to discover about whether it was really conscious. Consciousness, for Dennett, is not an extra ingredient on top of what a system does. It is a collection of things a brain does, including the stories it tells itself about its own states.

That makes his verdict on today's chatbots depend on what they do inside, not on how well they talk. Turing was right, he thought, that open-ended conversation is a far harder test than it looks. But a system trained to imitate the products of human minds can borrow their fluency without having much of the machinery that produced it. Late in life his worry ran the other way from Lemoine's. In 2023, writing in The Atlantic, he warned against "counterfeit people": systems designed to pass as human. Just as counterfeit money undermines trust in real money, he argued, machines that pass as people threaten the trust that conversation depends on, and making them should be treated as a serious offence. The danger, for Dennett, was not that we would be cruel to conscious machines but that unconscious ones would be used to manipulate us.

The illusionist twist

Illusionism changes the question in a useful way. If, as Dennett and Keith Frankish argue, phenomenal consciousness in the philosophers' sense is an illusion the brain creates for itself, then asking whether a machine "really" has it is like asking whether it really has a soul. The sensible question becomes: does it have the illusion? Does it build a model of its own states that represents them as inner, private and ineffable, the way ours does?

Michael Graziano has turned this into an engineering proposal. On his attention schema theory, our sense of awareness is the brain's simplified model of its own attention. Build that model into a machine, and it will be conscious in the only sense we are: it will believe it has an inner experience and say so for the same reasons we do. Susan Blackmore makes a similar point when she says that what it is like to be conscious may be what it is like to be a model. The pay-off is that the question becomes one about mechanisms. You answer it by looking inside the machine, not by asking it how it feels.

The other side

John Searle's Chinese room was aimed at exactly the kind of system that now exists: one that produces fitting replies by following rules for manipulating symbols. However good the replies, he argued, rule-following alone gives you syntax, not meaning. Against the reply that a simulated brain would do better, he insisted that a simulation is not a duplication. After Searle: a computer simulation of a rainstorm leaves nobody wet, and a simulation of a brain need not think.

David Chalmers agrees with Dennett more than one might expect. He holds that any system with the same fine-grained functional organisation as a brain would have the same experiences, and his fading qualia argument says that replacing neurons one by one with silicon chips would leave experience intact. But because of the hard problem, he thinks no outside test could prove it. In a 2023 essay he put the chance that the language models of the day were conscious at under one in ten, pointing to features they lacked, such as recurrent processing, a global workspace and a unified sense of agency, and he thought it quite possible that their successors would have them.

Integrated information theory, developed by the neuroscientist Giulio Tononi with Christof Koch, takes the opposite line. It identifies consciousness with how much a physical system's parts constrain one another as a unified whole. Ordinary computer chips score almost nothing on this measure whatever software they run, so on this theory a perfect simulation of your brain on a conventional computer would feel nothing at all. Anil Seth doubts machine consciousness for a different reason: he thinks our experience grew out of the business of keeping a living body alive, and that a computer, which has nothing at stake, may lack what matters however clever it becomes. For a panpsychist such as Philip Goff, there is some experience in the chips' particles anyway; the open question is whether it adds up to a single subject.

A checklist from the science

In 2023 a group of philosophers, neuroscientists and computer scientists led by Patrick Butlin and Robert Long tried a practical approach. They took the leading scientific theories, among them global workspace theory, recurrent processing theory, higher-order theories and the attention schema, and turned each into a list of indicator properties that an AI system could be checked for. Their conclusion was that no current system was a strong candidate for consciousness, but that there were no obvious technical barriers to building systems that would satisfy the indicators. The method rests on an assumption Dennett would share: that the right kind of computation would be enough. Searle, Seth and the integrated information theorists would reject the whole checklist.

Why the answer matters

If a machine could be conscious, it could perhaps suffer, and then how we treat it becomes a moral question. Mistakes are possible in both directions. Treat an unconscious system as a person and we can be manipulated by it, or by whoever controls it, which was Dennett's fear. Treat a conscious system as a tool and we might cause suffering on a vast scale without noticing, which is why the philosopher Thomas Metzinger has argued for a pause on research that might create artificial suffering. Birch argues for precaution: taking the possibility seriously while the science is unsettled, without pretending to know more than we do.

Turing saw some of this coming. In the paper that started it all, after dismissing the argument that a machine could not be conscious because we cannot get inside it, he added a careful caveat:

I do not wish to give the impression that I think there is no mystery about consciousness. Turing, "Computing Machinery and Intelligence" (1950)

Dennett would add that the mystery is real but that it is a mystery about mechanisms, and so, like other such mysteries, one that can eventually be solved. The machines now being built may turn out to be the best test of that claim we will ever have.

Further reading