Anthropic CEO Calls for AI Safety Slowdown, Criticizes Industry Deception on Risks

Anthropic CEO and co-founder Dario Amodei has publicly stated that the artificial intelligence sector has misled the public about the dangers posed by its technology. Speaking on CBS News’ “Sunday Morning,” Amodei acknowledged the genuine perils associated with AI development, marking a shift from the previously optimistic narratives often pushed by tech companies.

The urgency of his comments follows warnings issued by former Anthropic researcher Jacob Coxon, who recently suggested on social media that companies like Anthropic and OpenAI are “gambling with our lives” by pursuing self-improving superintelligence. Coxon told CBS News that the trajectory of AI development resembles the dystopian scenarios depicted in science fiction films like “The Terminator,” warning that sufficiently intelligent machines could pose an existential threat to humanity.

Amodei admitted that the rapid pace of progress has been underestimated even by industry leaders. Describing AI development as following an exponential curve that is becoming increasingly steep, he clarified that this trend does not necessitate immediate panic or a complete shutdown of research. However, he characterized the acceleration as a significant warning sign that demands a deliberate slowdown in the rate of capability advancements.

In an essay published over the weekend, Amodei argued that the entire technology industry must collaborate to reduce the speed of AI capability growth. He emphasized that slowing down does not mean halting the release of new models, but rather ensuring rigorous testing for every generation. To support this, Anthropic plans to grant independent evaluators permanent, employee-like access to its systems to verify compliance with safety protocols, comparing the process to food safety inspections.

Despite these measures, experts argue the steps are insufficient. Amodei concurred, stating his support for federal regulation. Current legislative proposals range from outright bans on superintelligence to mandatory “kill switches” for dangerous systems. Amodei expressed openness to the kill switch concept but rejected a total ban, citing the potential for medical breakthroughs and the likelihood that such technology would still be developed by other nations, including authoritarian regimes.

Addressing the geopolitical dilemma of whether the U.S. should decelerate if China does not, Amodei called it the most difficult aspect of the situation. He pointed to past nuclear arms control agreements with the Soviet Union as a potential model for establishing speed limits on AI progress, though he admitted that the military incentives to outpace rivals make international cooperation extremely challenging.

Amodei also expressed discomfort with the current structure of AI development being controlled by private entities rather than governments. While wary of any single government abusing the technology, he suggested that joint governance or oversight among democratic nations might be a viable path forward.

He concluded by noting that the critical question is no longer if AI will become powerful enough to require constraints, but who will implement them. “If we build in the right way, I think the probability of something bad happening is very low,” Amodei said. “If we build in the wrong way, the probability of something bad happening is very high.”