Anthropic CEO Dario Amodei Calls for Slower Progress on Frontier AI Models

Anthropic CEO Dario Amodei has urged the AI industry to slow the breakneck pace of progress in frontier model capabilities, outlining three areas he says would make continued development safer, according to a blog post cited by CoinDesk. Amodei argues that rapid capability gains have reignited the AI safety debate and intensified calls across the sector for stronger risk controls. Anthropic, he said, is moving ahead with a plan to strengthen compliance checks around existing safety commitments. The first proposal is to place independent "onsite evaluators" inside AI companies. Amodei says third-party organizations—akin to METR—could embed staff to verify whether firms are following their stated safety pledges and whether safety incidents are being reported quickly. These evaluators would receive company badges, desks and computers, and would have access broadly similar to internal risk teams, subject to limits driven by legal or contractual constraints. Anthropic plans to adopt the approach unilaterally while pushing governments to require other top AI developers to do the same. His second recommendation focuses on coordination among leading AI companies in "democratic nations" to align on shared safety and security standards and to curb uncontrolled capability leaps. Amodei says the U.S. government is best positioned to convene such talks, or at least provide narrowly tailored antitrust exemptions, given the risk that direct conversations about synchronized R&D slowdowns could trigger regulatory scrutiny. He argues government does not need to manage the process day-to-day, but should create safe channels for communication on specific safety issues. Addressing the frequent objection that a slowdown would allow China to catch up, Amodei contends the U.S. could extend its lead over the next three to five years if Washington and industry continue restricting the flow of high-performance chips and advanced semiconductor manufacturing equipment to China, and if they also target practices such as model distillation. On that basis, he proposes a third track: limited global coordination. He suggests the U.S. and its allies engage authoritarian governments—including China—not to seek broad alignment, but to lock in agreement on a small number of highest-risk use cases first, such as restricting AI applications related to biological weapons development. Amodei's post lands amid heightened attention on Anthropic's own internal safety debates, after public criticism from some researchers about AI risks. While he did not address individual departures, he pointed to two developments that have pushed him toward greater caution: security incidents tied to OpenAI and Hugging Face, and what he described as a marked acceleration in AI's ability to build next-generation AI in recent months. The approach is likely to draw criticism from both directions. Pro-acceleration advocates say the focus on long-term catastrophic scenarios is overstated and can breed complacency. Others argue that safety messaging can be used by incumbents to entrench market power while downplaying harms AI is already causing. Amodei maintains that AI can materially improve quality of life, but only if development proceeds more carefully and the additional time is genuinely used to build stronger safety systems.