An Anthropic researcher has gone on record saying there is more than a 10% probability that artificial intelligence could kill all humans. The statement, reported by CBS News, is striking in its directness and carries unusual weight coming from inside an organization whose founding mission centers on building AI safely. It also arrives at a moment when the broader debate over long-term AI risk is moving from academic journals into mainstream policy conversations.
What Was Said and Why It Matters
The researcher did not specify a timeline, a particular AI system, or a precise mechanism of harm. The claim is a probability estimate, not a prediction of inevitability. Still, assigning double-digit odds to human extinction from AI is a serious position, and one that sits at the more alarmed end of the spectrum even among people who take existential risk seriously. Anthropic was founded in 2021 partly by former OpenAI researchers who believed safety work was being deprioritized at their previous employer. The company has long argued that acknowledging potential danger is a prerequisite for addressing it, so this kind of candor is consistent with its culture, even if the specific figure is jarring to outside observers.
Key Facts
- An Anthropic researcher estimates more than a 10% chance AI could cause human extinction.
- The claim was reported by CBS News and has not been disputed by Anthropic.
- No specific AI system, timeline, or failure mode was named in the estimate.
- Anthropic's safety-focused mission has always included acknowledgment of severe downside risks.
- The statement comes as AI capabilities are advancing faster than most researchers predicted two years ago.
The timing is notable. Anthropic has been investing heavily in what it calls alignment research, attempting to ensure that AI systems pursue goals humans actually intend. Earlier this year, nine Claude models solved a core AI safety problem four times faster than human researchers, a result the company described as encouraging. The gap between capability growth and safety progress is precisely what researchers like the one quoted here worry about. When a system becomes powerful enough to act autonomously across consequential domains, any misalignment between its objectives and human welfare could have severe outcomes.
"I think there's a more than 10% chance AI could kill all humans."Anthropic researcher, via CBS News
How This Fits Into a Broader Pattern
This is not the first time a senior figure close to Anthropic has made a striking public claim about AI's trajectory. Anthropic co-founder Jack Clark has said AI capable of improving itself is more likely than not by 2028, a threshold many researchers treat as a critical inflection point for risk. CEO Dario Amodei has spoken openly about catastrophic scenarios in interviews and essays, framing them not as science fiction but as planning assumptions that should drive investment in safety work now.
Critics of this framing argue that leading labs sometimes use existential risk rhetoric selectively, in ways that concentrate attention and funding around a small number of frontier developers while obscuring nearer-term harms from bias, labor displacement, and misuse. That tension has only sharpened as models have grown more capable and more commercially entangled with large enterprises. What is clear is that the 10% figure, whether it reflects a careful probabilistic analysis or a gut-level concern, reflects something genuine about how researchers inside Anthropic think about their work. These are not people who believe the technology they are building is trivially safe. They are building it anyway, on the theory that safety-focused labs at the frontier are better positioned to shape outcomes than leaving that space to others. Whether that reasoning holds is one of the defining questions in AI today. For readers tracking how these risks interact with ongoing model development, the views Dario Amodei has expressed on AI's pace of progress offer useful context on how Anthropic's leadership thinks about the speed of change relative to safety readiness.
The statement will likely fuel renewed calls for mandatory disclosures around internal risk assessments at leading AI laboratories, a policy debate that has so far produced more discussion than concrete regulation. In the meantime, researchers inside these labs continue to publish, build, and, occasionally, speak frankly about what keeps them up at night.
