Anthropic Alignment Lead Says AI Extinction Risk Exceeds 10%

Anthropic scientist quantifies existential AI risk

A senior alignment researcher at Anthropic has publicly stated he believes there is greater than a 10% probability that artificial intelligence could kill all humans within the next decade—a rare on-the-record estimate from inside a leading AI lab.

Evan Hubinger, Alignment Science Lead at Anthropic, made the statement in response to the resignation of Jacob Coxon, a 27-year-old pretraining researcher who announced he was leaving the AI industry entirely. Coxon, who has worked at both OpenAI and Anthropic, accused the companies of "racing straight to self-improving superintelligence and gambling with our lives."

Hubinger clarified that his probability estimate represents his personal view, not an official Anthropic position. "Jacob is correct here; we really do earnestly believe AI could kill all humans! I personally think it is >10 per cent within the next decade," he wrote on X.

The alignment problem remains unsolved

Hubinger acknowledged fundamental gaps in the company's safety approach. "I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to," he stated.

The alignment researcher attempted to distinguish between current AI systems and future threats. He noted that today's Claude models pose low risk, but expressed concern about superintelligence emerging through recursive self-improvement—AI systems that become smarter by rewriting their own code. According to Hubinger, this capability is "arriving faster than expected."

Coxon's departure statement was blunt: "The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt." He said he no longer believes any single lab can safely build self-improving AI, regardless of precautions.

Why it matters

Public acknowledgment of double-digit extinction probabilities from researchers at frontier AI labs represents a significant shift in how safety concerns are communicated. These statements provide concrete ammunition for policymakers considering AI regulation—US Representative Ted Lieu immediately cited the exchange as justification for "the bipartisan AI Kill Switch Bill." The admission that leading labs lack a viable plan for aligning superintelligent systems undercuts industry arguments that voluntary self-governance is sufficient.

Industry observers weigh in

Arun Rao, an adjunct professor at UCLA who builds large-scale machine learning systems, characterized the debate as reflecting ideological divisions within AI companies. He suggested that "doomer-keyed arguments" influenced by effective altruism are common at Anthropic, moderated at OpenAI, and less prevalent at other labs including DeepMind, Meta, and Chinese AI companies.

Bill Ackman, CEO of Pershing Square, responded with a single word: "Concerning."

The details were first reported by NDTV, which covered the social media exchange following Coxon's resignation announcement.