Geoffrey Hinton Says a 10% AI Extinction Risk This Decade Is “Not Unreasonable”

Geoffrey Hinton, the Nobel Prize-winning computer scientist often described as a pioneer of modern AI, said that a 10% chance of AI killing all humans within the next decade did not seem unreasonable. But he paired that striking answer with a major qualification: “nobody really knows” how to make a sensible probability estimate. 91012

His warning is therefore not a prediction that extinction will happen, nor a claim that current chatbots already pose that risk. It is a precautionary argument about what could happen if AI systems become far more capable than people and remain difficult to control.

What Hinton meant by the 10% estimate

In the BBC Newsnight interview, Hinton agreed that 10% was not an unreasonable estimate for an AI-driven human-extinction event within a decade. His reasoning was rooted in uncertainty: humanity has not previously had to coexist with entities that might become more intelligent than humans. 910

Hinton has made similarly grave longer-horizon assessments before. In late 2024, he said he saw a 10% to 20% chance that AI could lead to human extinction within three decades; in 2025, he described that range as a gut-level judgment rather than a calculated forecast. 1415

The important point is the distinction between a risk estimate and a forecast of inevitability. Hinton’s view is that the potential stakes are so high that uncertainty is a reason for caution, not reassurance.

Why a future superhuman system worries him

The scenario under discussion is not a movie-style machine uprising. Hinton’s concern is that an AI system with substantially greater strategic and technical capability than humans could create chaos or pursue goals that conflict with human interests.

Reported pathways include manipulating people, creating biological or computer viruses, carrying out cyberattacks, or helping disrupt critical infrastructure. These are proposed mechanisms for harm by highly capable future systems—not demonstrated capabilities of today’s AI tools. 1011

That difference matters. Evan Hubinger, Anthropic’s alignment science lead, explicitly said that he considers risk from present models to be low. His concern, like Hinton’s, is about systems potentially becoming much more capable or improving themselves faster than researchers can safely manage. 11624

Jacob Coxon and Evan Hubinger: a warning from inside Anthropic

The latest discussion intensified after Jacob Coxon left Anthropic and the AI industry. Coxon argued that Anthropic and OpenAI were racing toward self-improving superintelligence without acting responsibly enough, calling the effort a gamble with human lives. 317

Coxon did not make a precise prediction that humanity would end by 2030. His narrower claim was that people building frontier AI genuinely believe the technology could kill everyone by the end of the decade. 57

Hubinger publicly backed the seriousness of that concern. He wrote that he personally assigns a greater-than-10% chance that AI could kill all humans within the next decade, while also saying Anthropic is trying its best. Crucially, he added that the company does not yet have a plan to solve alignment for superintelligence and is not clearly on track to develop one. 16

What “alignment” means in this debate

Alignment is the problem of ensuring that advanced AI systems reliably act in accordance with intended human goals and remain controllable, including in situations their creators did not anticipate.

For safety-focused researchers, the challenge is not merely getting a system to refuse a bad prompt. It is whether a system with broad autonomy, access to tools, and capabilities beyond its operators’ understanding could be directed safely under real-world pressure.

Hubinger’s comments do not establish that alignment is impossible. They establish that, in his assessment, there is not yet a demonstrated solution for aligning superintelligent systems before such systems may arrive. 16

Where other experts agree—and differ

There is no consensus around Hinton’s exact 10% figure.

Gary Marcus endorsed Hinton’s core precautionary statement that it would be foolish to develop superintelligence before there is scientific consensus that it can be built safely and controllably. But reporting also notes that Marcus does not see a realistic near-term route to human extinction, illustrating a key divide: agreement on the need for safeguards does not require agreement on the probability or timetable of catastrophe. 1141

Other public debate has similarly mixed concern about long-term AI risk with disagreement about whether current technical trajectories plausibly lead to autonomous, uncontrollable superintelligence soon. That uncertainty is why numerical estimates should be read as expert judgments, not measurements.

The practical policy question

The strongest common ground is not that AI extinction is certain. It is that a non-trivial possibility of losing control over systems more capable than their operators deserves public scrutiny before development outruns safety work.

Hinton’s estimate, Coxon’s resignation, and Hubinger’s alignment warning all point to the same unresolved question: what evidence should society require before deploying or continuing to scale systems that could become strategically superhuman?

For readers, the useful takeaway is to separate three claims that are often blurred together:

• Current AI models pose an extinction-level threat.
• Future systems might gain capabilities that create such a threat.
• The probability of that outcome can be estimated precisely.

The sources support the second as a serious concern raised by prominent researchers, while Hubinger describes current-model risk as low and Hinton says the probability itself cannot be calculated reliably. 11016