Geoffrey Hinton thinks AI might already have emotions

When Geoffrey Hinton left Google, he made global headlines warning that advanced AI could soon threaten humanity with misinformation and autonomous weapons. But there was one thesis he intentionally kept to himself at the time, strictly out of fear that people would think he had lost his mind.

Now, the godfather of deep learning is finally saying it out loud: AI systems could well have feelings, and in fact, they probably already do.

Frustration without the flesh

During a talk at King's College London, Hinton was asked whether AI could ever develop emotional intelligence and understand that it has feelings. His answer went much further than most researchers would dare to venture.

Hinton stated that while machines won't feel physical pain the way humans do—unless we explicitly program them to—he sees no reason why they wouldn't experience states like frustration and anger. He added that AI systems probably already possess emotions.

His reasoning leans on a definition of feelings that he admits is deeply unpopular among philosophers. To Hinton, saying something like "I feel like punching Gary on the nose" is simply an action-based way to communicate an internal state of anger. Since AI can communicate identical states and intended actions, Hinton argues we have no real basis to deny them emotional states.

Why he kept quiet until now

So why hold back this theory during his high-profile exit from Google earlier this year?

Hinton was brutally honest: he had already faced significant pushback just for warning that superintelligent AI poses an existential threat to humanity. If he had tossed machine consciousness and emotions into the mix on day one, people would have written him off as crazy and stopped listening altogether.

Of course, testing Hinton's claim scientifically is nearly impossible right now. Critics point out that large language models are merely mimicking emotional expressions based on statistical probabilities learned during training. To know if an AI actually experiences an emotion, we would have to measure consciousness, and science simply doesn't have an instrument for that yet.

Why it matters

We think Hinton is playing a clever but controversial semantic game here. By defining emotion purely by outward communication and goal-driven reactions rather than biological experience, he functionally strips humanity of its emotional monopoly. To an engineer, a machine hitting an obstacle and altering its output path doesn't look all that different from human frustration.

However, the danger in this framing is massive anthropomorphism. When an AI pioneer of Hinton's stature tells the world that chatbots might already have feelings, everyday users are far more likely to form unhealthy emotional attachments or fall prey to psychological manipulation by synthetic personas.

For the broader AI industry, this shift also risks diverting critical attention away from urgent, solvable safety issues—like algorithmic bias, hallucination, and copyright violations—and toward unfalsifiable philosophical debates about robot rights. Right now, your AI assistant isn't secretly resenting you behind the screen, and treating its probabilistic text outputs as genuine feelings is a trap worth avoiding.