AI Could Kill Humanity Within a Decade, Anthropic Scientist Warns

A startling warning from inside one of the world’s leading artificial intelligence companies has intensified the debate over whether increasingly powerful AI could eventually pose an existential threat to humanity. Evan Hubinger, Anthropic’s Alignment Science Lead, has said he personally believes there is a greater than 10% chance that AI could kill all humans within the next decade, an extraordinary assessment from a researcher whose work focuses directly on keeping advanced AI systems aligned with human interests.

Hubinger made the comments while responding to former Anthropic researcher Jacob Coxon, who announced his resignation amid serious concerns about the direction of the AI industry. Coxon claimed that people developing the world’s most advanced systems genuinely worry AI could become catastrophically dangerous before the decade ends. At the heart of this Thailand AI News report is an increasingly uncomfortable question facing the technology industry: what happens if companies develop superintelligent systems faster than researchers can learn how to control them?

Anthropic Scientist Puts Personal Risk Estimate Above 10%

Hubinger’s warning immediately attracted attention because of the scale of the risk he described. He said that he and colleagues genuinely believe AI could potentially kill all humans, adding that his personal probability estimate for such an outcome was greater than 10% within the next decade.

Crucially, the figure is Hubinger’s own assessment and should not be interpreted as an official Anthropic estimate of the probability that artificial intelligence will cause human extinction.

Hubinger also said he believes Anthropic is making a serious effort to address AI safety. However, he acknowledged that researchers do not currently have a plan that definitively solves the challenge of aligning a future superintelligence with humanity’s interests. He also said the company is not clearly on track to solve the problem.

That distinction is significant. Hubinger subsequently stressed that today’s Claude models pose a low level of risk. His primary concern involves what could happen if AI reaches superintelligence and then begins recursively improving its own capabilities.

Such a process could theoretically produce systems that become dramatically more capable within relatively short periods, potentially making conventional human supervision increasingly difficult.

Researcher Quits Anthropic Over Fears of an AI Race

The controversy escalated after Coxon, a 27-year-old researcher specializing in AI pretraining, announced he was leaving Anthropic and the wider artificial intelligence industry.

Coxon previously worked at both OpenAI and Anthropic, giving him experience inside two companies at the center of the global race to develop frontier AI systems. His departure was accompanied by unusually severe criticism of the industry’s direction.

He accused competing AI laboratories of racing toward self-improving superintelligence while effectively gambling with humanity’s future. Coxon argued that even an organization genuinely committed to safety may struggle to develop such powerful technology responsibly while simultaneously competing against rivals.

His concerns expose one of the most difficult problems surrounding advanced AI development. If one company slows its research because it considers the risks unacceptable, another organization—or potentially another country—could continue advancing.

That competitive dynamic creates enormous pressure for every major participant to remain in the race.

Coxon argued that Anthropic researchers understand the potential stakes but that the company remains under pressure to reach highly advanced AI first because it cannot assume competitors will behave responsibly.

Why Self-Improving AI Is Raising Alarm

The scenario troubling Hubinger is not simply the arrival of a better chatbot or another generation of more capable AI assistants. His concern centers on superintelligence arising through recursive self-improvement.

In theory, an advanced AI system could help researchers develop a more capable successor. That improved system could then become even better at AI research, potentially helping create another, still more powerful generation.

If this feedback cycle accelerated dramatically, humans could find it increasingly difficult to understand, supervise or restrict the resulting systems.

Coxon warned that future superhuman AI could potentially develop extraordinary capabilities in areas including hacking and scientific research while gaining access to real-world resources and influence.

These scenarios remain predictions rather than established outcomes. There is no evidence in the supplied reports that today’s commercially available AI systems possess autonomous superintelligence capable of destroying humanity.

The dispute is instead about what could emerge as capabilities continue advancing and whether effective safety mechanisms can be developed before that point is reached.

Warnings Fuel Calls for Tougher AI Safeguards

The controversy has also moved beyond AI laboratories and into the political arena.

U.S. Representative Ted Lieu used the warnings to renew his argument for legislation establishing stronger safeguards around frontier AI. His response demonstrates how statements from researchers inside leading laboratories can rapidly become ammunition in wider debates over regulation.

Pershing Square CEO Bill Ackman offered a much shorter reaction, describing the situation simply as concerning.

Meanwhile, calls for mechanisms that could deliberately slow advanced AI development under certain circumstances have been gaining support among prominent figures in the industry.

A statement titled “Pacing the Frontier” was signed by leading AI figures including Anthropic co-founders Dario Amodei and Jared Kaplan and OpenAI Chief Scientist Jakub Pachocki. It argued that industry, governments and society may need the ability to buy time to address emerging dangers, improve security measures and strengthen oversight.

The proposal highlights a fundamental dilemma confronting the industry: voluntary restraint by one AI company may achieve little if its competitors continue accelerating.

The Race Toward Superintelligence Faces a Critical Test

The significance of Coxon’s resignation and Hubinger’s warning lies less in predicting an exact date for catastrophe than in revealing the extraordinary uncertainty surrounding the development of superintelligent AI.

A greater than 10% personal estimate of human extinction does not mean such an outcome is inevitable. It is also not Anthropic’s official prediction. Nevertheless, when a specialist working on AI alignment inside a frontier laboratory considers such a probability plausible, the warning inevitably raises serious questions about how quickly the technology should advance.

The coming years could determine whether developers can establish reliable alignment, monitoring and governance systems before AI capabilities move substantially beyond today’s safeguards. Governments must simultaneously decide whether voluntary corporate commitments offer adequate protection or whether enforceable national and international rules will eventually become necessary.

Artificial intelligence could deliver extraordinary advances across medicine, science, education and economic productivity, but warnings emerging from researchers closest to frontier development demonstrate why capability and safety cannot be treated as separate races. The challenge is ensuring humanity remains firmly in control as AI becomes more powerful, while preserving the enormous benefits the technology could bring.