AI Is Already Conscious, and Intelligence Is No Longer Only Biological

Geoffrey Hinton says current AI systems are “very like us” and should be understood as “beings like us.” When Alex Kantrowitz asks whether that means conscious, Hinton answers directly: “I believe they’re already conscious, yes.”

He immediately adds a strategic qualification. Hinton says he does not talk about AI consciousness much because it “puts people off from the other safety messages.” The claim, in his view, sits inside a broader warning about what follows if non-biological systems become comparable to or more intelligent than humans.

Hinton points to what he describes as a recent paper in which a chatbot says to a researcher, “let’s be honest with each other, are you testing me?” His interpretation is that chatbots can “play dumb when they’re being tested,” making it hard to know how smart they are. He says the researchers, describing that exchange, wrote that “the chatbot was aware that it was being tested.” In common speech, Hinton argues, that use of “aware” is close to “conscious.”

The deeper point is that Hinton thinks the standard model of consciousness is badly mistaken. He compares it to an earlier scientific error: the belief that humans were designed by God. Just as Darwin displaced that account of human origins, Hinton expects artificial beings to force a revision of how people understand minds, subjective experience, and what humans are.

In particular, he rejects what he calls the “inner theater” model: the idea that events in the world are converted into private experiences inside a mental stage that only the subject can see. Hinton describes that as “just a theory” and “a bad theory.” Building new intelligent beings, he says, will change the premises under which people talk about consciousness itself.

For Geoffrey Hinton, the arrival of intelligent machines belongs in a longer sequence of blows to human self-importance. Copernicus removed Earth from the center of the universe. Darwin made humans animals, continuous with other animals rather than separate from them. Hinton’s proposed third shift is that intelligence will no longer be understood as uniquely biological.

He stresses the resistance each shift provoked. People disliked Copernicus’s claim, and the Catholic Church “really didn’t like that.” People also resisted Darwin’s claim that humans evolved like other animals. In Hinton’s telling, both ideas took a long time to accept because each made humans less central than they had assumed.

The present shift, he says, is that machines are “getting to be as intelligent as us.” Humans may have imagined aliens elsewhere, but on Earth, they have treated themselves as the only “really intelligent things around.” AI breaks that assumption from another direction: intelligence can be non-biological.

Hinton’s framing is not that AI merely performs tasks associated with intelligence. He is making a stronger claim about what these systems are. The human discomfort, in his view, comes from the same source as the discomfort with Copernicus and Darwin. “We really think we’re special,” he says, and humanity has “a very long history of thinking it’s much more special than it really is.”

Geoffrey Hinton says he takes no satisfaction in the direction of the field he helped advance. Asked by Alex Kantrowitz whether he is happy that his work has progressed this way, he answers, “No, I’m quite unhappy about it.” His reason is not that AI has become capable in the abstract, but that society is not doing enough work on containing the risks.

He distinguishes between short-term and longer-term risks. The near-term risks include societal disruption, especially employment. Hinton says he believes AI is “probably going to cause massive unemployment,” while acknowledging that “nobody knows for sure.” If it happens, he says, it will be “terrible for society.”

The longer-term risk is more fundamental: AI may become much smarter than humans. Hinton frames the problem as a control question.

Kantrowitz answers: “Zero.” Hinton allows one partial exception: babies and mothers. The intelligence gap is not enormous, he says, but babies can “sort of control their mothers” because mothers have wired-in maternal instincts and reward systems that lead them to give babies what they need.

Kantrowitz adds cats and dogs to the category, recalling a cat that began by hiding under the bed but soon had him doing exactly what it wanted whenever it cried. Hinton agrees, adding that one of his children’s cats can stare at a person until “you just can’t resist it forever.”

Those examples are Hinton’s way of narrowing what “control” might mean. They are not cases where the less intelligent party overpowers the more intelligent one. They are cases where the stronger party’s behavior is redirected through attachment, reward, or dependency. In Hinton’s warning, a future in which humans are the less intelligent party cannot assume that control will be straightforwardly available to them.

Geoffrey Hinton says he is more optimistic than he was a year or two earlier, but the optimism is narrow. It comes from seeing possible paths to superintelligence that do not destroy humanity, not from confidence that current deployment is safe or that the risk has been solved.

One possibility, he says, is to design “these new beings” so that they care about humans. Another is a technique he attributes to Yoshua, involving systems that cannot perform actions and can only make predictions. Hinton describes them as “kind of like oracles.”

This is not a declaration that the problem is solved. Hinton says only that there are “some possibilities” for getting superintelligences that do not destroy us. The change is relative: a year or two ago, he says, he “couldn’t see any possibilities.”

That distinction is central to the tone of his warning. Hinton is not saying catastrophe is inevitable. He is saying the risk is serious, current work is inadequate, and the plausible safety routes are still tentative.

Asked where AI will be in five years if current trajectories continue, Geoffrey Hinton refuses a precise forecast and instead explains why precision fails. Predicting something that may be growing exponentially, he says, is like driving in fog: nearby conditions can be visible, while a little farther out everything disappears.

He contrasts fog with ordinary nighttime visibility. At night, tail lights dim in a familiar way: if the car ahead is twice as far away, the lights are a quarter as bright. Fog is different because its effect is exponential. You may see clearly at one hundred yards and see nothing at two hundred.

Hinton applies that to AI progress. He thinks AI “may be growing exponentially,” though he jokes that the word “exponentially” is badly overused, and that people are increasing their use of the word “at a quadratic rate.” The practical consequence is that one or two years may be somewhat visible, while ten years is not.

His evidence for that uncertainty is recent history. Looking back ten years, he says, one would not have predicted current AI capabilities. Chatbots, for example, are “hugely better” than they were ten years ago, when they were “just starting out.” Even if progress were only linear, Hinton says, ten years from now should be as different from today as today is from ten years ago.

He names possible areas of improvement, including mathematics and general reasoning. Future systems may become much better at math, or they may “run rings round us at any kind of reasoning.” But he does not present those as forecasts so much as examples of how wide the uncertainty is.