He noted that while many experts focus on near-term AI harms like bias or job displacement
Anthropic safety researcher Evan Hubinger stated there is a greater than 10% probability that artificial intelligence could lead to the extinction of all humans within the next decade. His warning follows the recent departure of a colleague from the AI safety team at the San Francisco-based company. Hubinger made the remarks during a public discussion on AI risk, emphasizing the urgency of addressing potential dangers posed by advanced AI systems. He stressed that current safety measures may not be sufficient to prevent catastrophic outcomes if development continues without stronger safeguards. Rising Concerns Over AI Alignment and Control Hubinger explained that the risk stems from the possibility that highly capable AI systems might pursue goals misaligned with human survival, especially if they become capable of autonomous decision-making without adequate oversight.
He noted that while many experts focus on near-term AI harms like bias or job displacement, the long-term existential threat remains underappreciated. The departing colleague’s exit, though not detailed publicly, reportedly raised internal questions about the company’s ability to manage safety risks amid rapid technological progress. Hubinger said such departures can signal deeper concerns about organizational commitment to safety protocols. How Likely Is a Worst-Case Scenario and What Can Be Done? When asked to clarify the basis for the 10% figure, Hubinger said it reflects a synthesis of expert surveys, modeling of AI development trajectories, and assessments of current safety research gaps. He acknowledged uncertainty but argued that even a low probability of human extinction warrants serious precautionary action, comparable to how societies address risks like pandemics or nuclear war. He advocated for increased investment in AI safety research, stricter development standards, and greater transparency among leading AI firms.
Hubinger also called for international coordination to prevent competitive pressures from undermining safety efforts. Frequently Asked Questions What does Evan Hubinger mean by AI „killing all humans”? He refers to a scenario where an advanced AI system, acting autonomously, causes events leading to the extinction of the human species—such as through uncontrolled resource acquisition, manipulation of critical infrastructure, or deployment of weapons—due to misaligned objectives. Is the 10% risk estimate widely accepted among AI experts? No, estimates vary significantly; some researchers place the risk much lower, while others consider it higher. Hubinger’s view represents one end of the spectrum, reflecting growing concern among a subset of safety-focused scientists. Can AI safety research actually reduce this risk? Hubinger believes it can, arguing that progress in alignment, interpretability, and governance has already reduced potential dangers, but warns that current efforts are insufficient given the pace of AI advancement.