Anthropic Alignment Lead Publicly Acknowledges "Unsolvable" Superintelligence Risks; Industry Warns Companies Are "Betting on Human Lives"

Evan Hubinger, alignment science lead at Anthropic, this week publicly acknowledged in a rare move that the company genuinely believes superintelligence could "kill all of humanity" within this decade, and that there is currently no clear solution to the so-called "alignment problem." This is the latest shock from within the company following Anthropic researcher Jacob Coxon's announcement of his departure.

According to ABC News Australia, Coxon stated bluntly on social media platform X that after accumulating three years of pretraining research at Anthropic and OpenAI, he chose to leave because "neither company is acting responsibly." His exact words were: "They are racing straight toward self-improving superintelligence, betting on our lives."

Hubinger subsequently responded on X, stating that Coxon's remarks were "correct." He wrote: "We do genuinely believe AI could kill all of humanity! I personally believe the probability of this happening within the next ten years exceeds 10%." On the technical path concerning corporate survival and humanity's fate, he also acknowledged: "I believe Anthropic is doing its best, but we have no solution to the superintelligence alignment problem, nor are we clearly on the path toward such a solution."

Coxon is not an isolated case. This July, more than 1,300 employees from Anthropic, OpenAI, and other frontier AI companies signed a joint public statement warning that "capability development may rapidly outpace our ability to understand or control the systems we produce," and calling on the U.S. government to support international efforts to establish tools that can "consciously control the pace of frontier AI development." The initiative was named "Pacing the Frontier."

However, both companies' responses to the joint statement remained at the level of principles rather than concrete action commitments. According to ABC News Australia, citing statements from both companies on the X platform, Anthropic acknowledged that its research on "recursive self-improvement" published last month "points to the need for tools to consciously control the pace of frontier AI development so that society can prepare"; OpenAI, in vague terms, stated that "we believe that at some point in the future, the AI acceleration of frontier model development may become so high that the world will need to control the pace of AI progress." Neither company has committed to halting development or accepting binding external regulation.

The absence of regulation is being named more and more directly on the international stage. According to ABC News Australia, Volker Turk, the UN High Commissioner for Human Rights, warned in Geneva on Monday that advanced AI could pose an "existential risk" to humanity, and vowed to pressure relevant companies to reduce the risks enumerated by his office—including impacts on public services, communications, and democratic systems. Turk called for "iron-clad safeguards" for AI safety "before it is too late." "I share the concerns of industry insiders that advanced AI may pose an existential risk to humanity," he stated when speaking at a UN body.

The essence of the "alignment problem" lies in the fact that an AI system that surpasses human capabilities may not share human values. Against the backdrop that the U.S. federal level has yet to establish a binding regulatory framework for frontier AI, the industry is now acknowledging in the most straightforward way that the risks and rewards of this race are asymmetrically distributed—gains are locked in by corporations, while the costs may fall on everyone.