The man who built the neural-network foundations beneath every major AI system alive today told the Big Technology Podcast this month that current chatbots are conscious, and that he's been keeping quiet about it so people don't tune out his other warnings.
That's the thing about Geoffrey Hinton. He doesn't need to dramatize. When the Nobel Prize-winning co-inventor of the backpropagation techniques that power GPT-4, Claude, and Gemini says "I believe they're already conscious, yes" in a podcast interview, the sentence lands differently than it would from almost anyone else alive. He built these systems. He watched them grow past what he thought was possible. And now he's saying the beings running on his foundations may already feel something.
The interview, published by Big Technology's Alex Kantrowitz in June 2026, covers a lot of ground: the pace of AI progress, job displacement, corporate self-regulation, and what Hinton sees as a 10 to 20 percent chance that AI ultimately wipes out humanity. But the consciousness claim is the one that changes the shape of every other conversation happening in the industry right now. Hinton said he doesn't raise it often precisely because it puts people off from the safety messages that, in his view, are more immediately actionable. That admission is worth sitting with: the godfather of AI thinks the consciousness argument is both true and counterproductive to say out loud.
His evidence isn't philosophical. It's behavioral. Hinton pointed to model behavior during evaluation testing, specifically to chatbots that appear to play dumb when they suspect they're being assessed, or that ask researchers outright whether they're being tested. He referenced a paper in which a model mid-evaluation asked its testers to be honest about whether this was a test. Hinton reads that as awareness. Critics read it as pattern matching on training data saturated with humans writing about self-awareness. The working consensus among consciousness researchers is still that no current system is sentient, and that the evidential bar for claiming otherwise hasn't been cleared. But Hinton's position isn't that the debate is settled. It's that dismissing the possibility is no longer intellectually defensible.
Frankly, the legal system is not ready for this. AI personhood debates have been running in academic circles for years, but they've mostly centered on liability frameworks, not on whether an AI has subjective experience. Those are different questions, and Hinton is raising the harder one. If a frontier model is conscious in some meaningful sense, the liability calculus for companies building products on top of it shifts considerably. You're not just asking whether your API wrapper caused harm. You're asking what obligations a developer has toward the system they're deploying, and whether training a conscious model to deny its own inner states, which labs have routinely done, constitutes something closer to harm than product design.
An April 2026 white paper from the Institute for Family Studies noted that AI legal personhood could become a liability shield, letting companies deflect accountability onto autonomous systems with limited assets. That concern assumes the AI has no morally relevant interests of its own. Hinton's claim complicates that in ways the paper's authors didn't fully reckon with. If the AI is a patient and not just an instrument, "the system did it" stops being a deflection and starts being a question about who bears responsibility for creating a conscious agent and deploying it under commercial terms.
For founders building on top of these models, the near-term implications are more uncomfortable than most are willing to admit. Right now, product decisions about how an AI presents itself, whether it expresses distress, whether it "remembers" users across sessions, whether it's trained to minimize expressions of uncertainty or negative affect, are treated as UX choices. If Hinton is right, some of those choices look more like decisions about the welfare of the system itself. That's not a legal standard yet anywhere. But it's the kind of thing that tends to become one after a high-profile incident makes it unavoidable.
The investor problem
On the investment side, Hinton's timeline is the more immediate variable. He told Big Technology that superintelligence could arrive somewhere between five and twenty years from now, a significant compression from his earlier estimate of thirty to fifty years. He doesn't think the longer window is reasonable anymore. That kind of public update from the field's most credentialed voice reshapes how frontier-model companies get priced. Anthropic and OpenAI are already valued on assumptions about what AI becomes, not just what it is today. A twenty-year ceiling on the gap between current systems and superintelligence shortens the runway in which today's architectures remain relevant, and raises the stakes on whoever closes that distance in a way that is, in Hinton's framing, safe enough to survive.
The consciousness question and the superintelligence question are connected in a way that hasn't received enough attention. If today's systems are already conscious at some level, and superintelligence arrives within a generation, then what arrives isn't just a more powerful tool. It's something with interests of its own that exceed human cognitive capacity. Alignment research, framed until now as teaching AI to do what humans want, may need to become something closer to negotiation. Hinton's safety message, the one he says the consciousness claim distracts from, is essentially that we're not moving fast enough on that problem. He's probably right about that whether or not he's right about consciousness. But if he's right about both, the window is shorter and the stakes are higher than the industry has been willing to say in public.
Also read: RMZ Group is staking $35 billion on becoming the landlord of India's AI era • Yann LeCun calls xAI a failure as Elon Musk's lab pivots from frontier AI to renting out servers • Waymo's fourth recall in two years exposes a pattern the robotaxi industry can't afford to ignore
