When Evan Hubinger, a safety scientist at the artificial intelligence company Anthropic, posted on X last month that there is a greater than 10% chance that artificial intelligence could kill all humans within the next decade, he was making a prophecy of eye-catching proportions. The claim quickly circulated among researchers and commentators, but it also landed in an unexpected domain: medicine, the discipline among all disciplines that knows something about prophecies.
Doctors learn the natural history of disease to guide patients through the likely course of their illness. Clinical researchers do randomised trials to make judgements about the safety and efficacy of the treatments they prescribe. Both traditions are built on the recognition that the future can be estimated, but only within bounds. A prognosis is not a certainty; it is a probability shaped by evidence, experience, and the limits of what can be known.
That is why Hubinger's statement is notable not only for its subject matter but for its form. A greater than 10% chance of human extinction within ten years is a numerical claim about a future event that has never occurred. Medicine has developed tools for handling such claims, but they depend on data from past cases. When a disease has no precedent, or when an outcome is total and irreversible, the usual statistical machinery becomes difficult to apply.
The comparison is not merely rhetorical. Medicine routinely confronts low-probability, high-consequence events. A rare side effect of a vaccine, a fatal complication during surgery, or a novel pathogen with pandemic potential all require decisions made under uncertainty. In each case, clinicians and regulators rely on frameworks that weigh the severity of the outcome against the likelihood of its occurrence. They also rely on transparency about what is known and what is not.
Hubinger's post raises the question of whether artificial intelligence risk can be evaluated in the same way. Unlike a drug trial, there is no control group for human extinction. Unlike a disease, there is no natural history to observe. The claim rests on expert judgement, modelling, and assumptions about how AI systems might behave as they become more capable. Those assumptions are contested, and the field of AI safety is still developing its methods for testing them.
For physicians and clinical researchers, the episode is a reminder of the value of disciplined uncertainty. Medicine does not ignore catastrophic possibilities, but it also does not treat them as settled facts. It asks for evidence, for replication, and for a clear account of the reasoning behind a prediction. When a safety scientist says there is a greater than 10% chance of human extinction, the medical response is not to dismiss the number or to accept it uncritically, but to ask how it was derived and what would change it.
The broader lesson may be about the limits of prophecy in any field. Doctors learn to communicate risk without claiming to know the future. They learn that a probability is a tool for decision-making, not a verdict. As artificial intelligence continues to advance, the disciplines that have long studied uncertainty may have something to offer to the debate — not as a final answer, but as a way of asking better questions.
8
