AI Labs Face Internal Warnings on Existential Risks Amid Political Pushback

Recent warnings from current and former employees at Anthropic and OpenAI have renewed attention on claims that advanced artificial intelligence could pose an existential threat to humanity. Jacob Coxon, who left Anthropic after a short time there, posted that people building AI believe it could kill everyone by the end of the decade. Similar statements have come from other researchers at the labs, including concerns over recursive self-improvement where systems might enhance their own capabilities at accelerating rates.

These alerts follow a pattern that dates back years. In 2015, Sam Altman described superintelligence as the greatest threat to human existence. More recently, Anthropic cofounder Jack Clark referred to a non-zero chance of killing everyone on the planet. Critics note that such predictions have circulated in Silicon Valley without clear evidence of imminent catastrophe, often resurfacing amid public interest or internal lab debates.

Y Combinator CEO Garry Tan has pushed back against framing the issue around Hollywood scenarios. He argued that attention should stay on science fact rather than fiction, citing films like Terminator 2 as entertaining but irrelevant to policy. Tan called for focus on immediate challenges such as cybersecurity threats, infrastructure protection, and the behavior of autonomous AI agents. Questions like how to locate and shut down a rogue agent operating across data centers, he said, demand practical answers over speculation about human extinction.

President Trump has echoed this emphasis on tangible priorities. When asked about extinction risks, he stated he had none, instead stressing the need for the United States to maintain its lead over China in AI development. He described the current advantage as roughly a year, which he viewed as significant in a competitive global landscape. Trump's position aligns with concerns that pausing progress could hand advantages to adversaries without addressing real vulnerabilities in systems already in use.

Professor Melanie Mitchell has observed that identical claims about AI dangers have persisted for as long as she can recall, leaving her puzzled by the current surge in attention. The pattern suggests that insider warnings can amplify existing public unease, particularly when tied to events like security incidents at platforms hosting AI models. Yet data on actual AI performance shows incremental advances rather than sudden leaps toward uncontrollable intelligence.

Discussions around slowing development have gained traction among some lab staff, but responses from figures like Tan and Trump highlight the tradeoffs. Regulatory efforts aimed at hypothetical long-term risks could impose costs on innovation, diverting resources from problems such as securing networks against coordinated attacks or clarifying accountability for AI actions today. Historical examples of technological progress, from computing to biotechnology, indicate that competition and empirical testing have managed risks more effectively than preemptive restrictions driven by worst-case forecasts.

The divide underscores differing views on where attention belongs. While some employees advocate pausing to evaluate recursive capabilities, others stress that frontier labs already incorporate safety measures and that broader economic and security dynamics will shape outcomes. Public discourse may benefit from separating longstanding alarmist rhetoric from evidence-based assessments of current capabilities and competitive realities.