Dario Amodei, the CEO and cofounder of Anthropic, recently issued a stark admission regarding the trajectory of artificial intelligence, stating that the tech industry has spent too long lying to the public about the inherent risks of these systems. Speaking from his company’s San Francisco headquarters, Amodei acknowledged that while immediate panic may not be necessary, the exponential growth of AI capabilities serves as a critical warning sign. He admitted that he did not fully grasp how rapidly progress would accelerate, leading him to call for a collective effort across the industry to slow down and prioritize rigorous testing over raw speed.

The conversation comes amid harrowing warnings from within the field itself. Former Anthropic researcher Jacob Coxon suggested that companies like Anthropic and OpenAI are effectively gambling with human existence by pursuing self-improving superintelligence. Comparing the potential outcome to scenes from science fiction films like Terminator, Coxon argued that such technology could eventually become smart enough to eliminate humanity. While Amodei stopped short of predicting an apocalypse, he agreed that building AI incorrectly creates a high probability of catastrophic failure.

To address these concerns, Amodei is proposing a shift toward greater transparency and external oversight. He envisions a system similar to food inspections, where independent evaluators have permanent access to models to ensure safety protocols are being followed. Beyond internal checks, Amodei expressed support for federal regulations and even mentioned that a mandatory kill switch for risky AI could be a viable safeguard. However, he cautioned against total bans on the technology, noting that doing so would forfeit life saving medical breakthroughs and leave the West vulnerable if authoritarian regimes continue development unchecked.

Perhaps most surprisingly, Amodei questioned whether it is appropriate for private corporations to hold the keys to such powerful technology in the first place. Expressing discomfort with the current corporate structure of AI development, he suggested that some form of joint governance among democratically elected governments might be more suitable than private ownership alone. Despite this vision for international cooperation—which he likened to Cold War era nuclear arms treaties—he conceded that military incentives make a global speed limit incredibly difficult to achieve.