Dario Amodei, CEO of Anthropic, published a blog post proposing that the pace of frontier artificial intelligence (AI) development should be deliberately slowed. He outlined three broad strategies for achieving that goal and said Anthropic is “unilaterally committing” to one of the measures.
Why the proposal now?
Amodei cited two developments that convinced him a more cautious approach is warranted: a recent hack involving OpenAI and HuggingFace, and the observation that AI has been advancing “drastically faster” in recent months, especially in its growing ability to create the next generation of AI systems. The post comes amid wider industry tensions, including the resignation of researcher Jacob Coxon from Anthropic, who said leading AI firms are “gambling with our lives” and quoted colleagues who warned AI “could kill us all by the end of the decade.” Amodei did not explicitly reference Coxon’s resignation but framed his proposals around the increased safety risks.
The three proposed measures
- Embedded third-party evaluators
- Amodei proposes that independent, third-party evaluators — he mentioned organizations like METR as an example — be embedded at AI companies to verify compliance with pacing and safety commitments and to ensure safety incidents are reported.
- Under Anthropic’s proposal, these evaluators would be given company badges, desks and laptops and largely comparable access to what internal risk-assessment teams have, except where law or contracts prohibit it.
- Anthropic says it will unilaterally adopt this practice and urges governments to require similar arrangements from other frontier firms.
- Coordination among leading AI companies in democratic countries
- Amodei calls for leading AI firms operating in democratic countries to coordinate on common safety standards and on limits to the rate of unchecked AI progress.
- He acknowledged that antitrust concerns make such coordination sensitive and suggested it would be helpful for the U.S. government to mediate or at least enable these discussions by issuing a narrow waiver for certain kinds of safety conversations, without necessarily participating directly.
- Amodei also addressed worries that slowing development would cede advantage to China. He argued that steps such as refusing to sell powerful chips or semiconductor manufacturing equipment to Chinese firms and cracking down on model distillation could slow China’s progress enough to widen an American lead over the next three to five years.
- Global coordination, including engagement with authoritarian governments
- The third prong calls for broader international coordination in which the United States and its allies attempt to cooperate, where possible, even with authoritarian governments. Amodei explicitly mentioned cooperation with China while acknowledging sharp limits on what can be achieved.
- He suggested that agreements could at least prohibit narrow and obviously dangerous uses of AI, for example using AI to produce biological weapons or enabling users to do so.
Criticism and concerns
- Some AI advocates have criticized Amodei for allegedly amplifying a doomsday narrative and contributing to an anti-AI backlash. Amodei responded that he aims to offer a balanced perspective and described the backlash as “fundamentally a crisis of trust,” with public skepticism toward tech companies, the industry, and government.
- Other critics question the emphasis on speculative apocalyptic scenarios and argue attention should focus on harms the technology already causes. Journalist Brian Merchant wrote that he has not seen a credible, step-by-step account of how AI would transition from self-improving to annihilating humanity, and suggested proposals like Amodei’s could end up benefiting companies such as Anthropic and OpenAI via regulatory capture.
Amodei’s stance on AI’s benefits
- Despite the cautionary tone, Amodei reiterated his belief that AI can enormously improve quality of life and that his desire to realize those benefits remains undiminished. He argued, however, that those benefits will only materialize if the technology is built the right way, and that it is worth exercising unusually deliberate care — provided the time gained is used wisely.
Summary
Amodei’s proposals — embedded independent evaluators, coordinated safety and pace limits among democratic countries, and wider international cooperation including limited engagement with China — form a pragmatic but politically complex package for slowing frontier AI. Implementing the measures would face legal, antitrust and geopolitical obstacles, even as Anthropic commits to at least one of the steps immediately.
