Anthropic researcher says more than 10% chance AI "could kill all humans"

A senior safety researcher at Anthropic has said he places the chance that artificial intelligence could cause humanity’s extinction at above one in 10 during the coming decade. Evan Hubinger, who leads alignment science at the San Francisco AI company, made the assessment in a Wednesday post on X, presenting it as his own view of the danger posed by increasingly capable systems and the challenge of keeping them under human control.

The warning is notable because it came from within Anthropic, a prominent developer in the fast-moving AI industry. Hubinger did not describe the risk as remote or confined to a far-off future. Instead, he attached a numerical estimate to a 10-year horizon, saying he believed the possibility that AI might “kill all humans” exceeded 10%.

Hubinger’s role centers on alignment, the field concerned with making sure AI systems act in ways consistent with intended goals and human interests. His post suggested that the problem becomes especially consequential if AI reaches capabilities beyond those of people. He said Anthropic was working seriously on the issue, while also arguing that the company does not yet have a solution for aligning a superintelligent system.

That distinction matters because superintelligence remains a hypothetical concept rather than an established technology. The term refers to an AI agent whose intellectual abilities surpass those of the most capable human thinkers. Hubinger’s concern was therefore not a claim that such a system already exists, but a warning about what could follow if one is developed before reliable methods of control are in place.

His comments place a stark estimate into a wider argument about how quickly AI capabilities may advance and how much confidence the public should have in safety work. The central question is not simply whether systems can become more useful or more powerful. It is whether their behavior can be made dependable when their reasoning, planning, or technical capacity exceeds the ability of people to supervise them directly or to correct mistakes after they occur.

Anthropic has become one of the companies most closely associated with research into AI safety, making Hubinger’s statement particularly consequential for the industry’s broader credibility. He credited the company with trying to address the problem, but said the effort was not visibly progressing toward a plan sufficient for superintelligence. That leaves a gap between recognizing a severe potential threat and demonstrating a way to prevent it before more advanced systems arrive.

The estimate should also be understood as a personal risk judgment, rather than proof that such an outcome is expected or imminent. Hubinger presented it as his own assessment, while the technology he highlighted remains theoretical. Still, assigning a probability and a relatively near-term deadline makes the warning more concrete than broad discussions of hypothetical AI dangers, and raises questions about what level of uncertainty society should accept.

For policymakers, companies and the public, the development to watch is whether AI labs can show concrete progress on alignment as capabilities improve. Hubinger’s remarks suggest that safety claims will face closer scrutiny not only from outside critics but from researchers working inside leading developers. The debate is likely to focus on whether technical safeguards can keep pace with the systems they are meant to govern, and whether a credible plan emerges before superintelligence becomes more than a theoretical possibility.

Source: CBS World