Anthropic CEO Dario Amodei unveiled a three-step plan, arguing that the pace of AI development should be slowed. Amodei said the OpenAI-Hugging Face hack incident called for a more cautious approach, given that AI has progressed “drastically faster” in recent months and that the capacity to develop the next generation of AI has increased. Saying that the pace of “frontier” AI work should be reduced, Amodei stated that progress could still be rapid if the time gained was used carefully.
The first step of the plan is for “embedded evaluators” from third-party organizations such as METR to work at AI companies. These individuals are expected to oversee whether companies comply with their safety and slowdown commitments and ensure that safety incidents are reported. Amodei said Anthropic would implement this practice unilaterally, with evaluators being given a company identity, workspace and laptop. These individuals are expected to have access broadly similar to that of companies’ internal risk assessment teams, except where restricted by legal or contractual limitations. OpenAI CEO Sam Altman supported the proposal and announced that his company would take the same step.
The second proposal is for leading AI companies in democratic countries to coordinate in setting common safety standards and limits on the pace of unchecked progress. Amodei asked the US government to facilitate these talks or provide a limited exemption for certain safety discussions due to antitrust concerns.
Amodei also argued that China’s progress in AI could be slowed. He said that if the US government and technology companies blocked the sale of powerful chips and semiconductor manufacturing equipment to Chinese companies and took measures against model distillation, the US could significantly extend its lead over the next 3-5 years.
The third phase is for the US and its allies to establish global coordination with authoritarian governments wherever possible. Amodei said cooperation with China could remain limited, but that narrow agreements could be reached, such as banning the use of AI to produce biological weapons.
Anthropic researcher Jacob Coxon resigned, arguing that companies were putting human lives at risk. While some circles considered Amodei’s approach excessively pessimistic, journalist Brian Merchant argued that these warnings overshadowed the existing harms. Amodei, meanwhile, said he believed AI could greatly improve quality of life, but that this depended on developing the technology properly.
Why it matters
The proposal gives concrete form to the debate over whether AI safety should be taken out of the hands of companies’ own declarations and opened to independent audits. Granting evaluators access similar to that of companies’ internal risk teams turns questions about how safety commitments will be verified and when incidents will be reported directly into matters of corporate governance. The call for common standards across companies highlights the tension between competition law and public safety; the possible role of the US government is therefore not merely a technical issue but also a regulatory one. The proposal for a limited agreement with China alongside restrictions on chips and semiconductor equipment shows that the debate has expanded into the areas of technology trade and international security. However, the resignation and criticism show that slowing down is not universally accepted and that which risks should be prioritized remains unclear.
Background
Anthropic is not a new name in the FikirPilot archive: we have published 20 news stories mentioning the name in the past 90 days; the latest is dated September 13, 2026.