Amodei Pushes AI Slowdown as Labs Split on Safety
Anthropic CEO Dario Amodei has proposed a three-step plan to slow development of frontier AI models and place it under closer independent monitoring.
The proposal calls for embedding third-party evaluators inside AI labs, coordinating safety practices across the domestic AI industry, and forging international agreements potentially with government assistance The Verge. Anthropic has framed the issue as pace. The company called for the pace of AI model development to slow down and be closely monitored BBC.
What Amodei proposed
The first element is stronger safety testing of advanced models Reuters. The second is structural. Amodei called for embedded independent evaluators with employee-like access to verify safety practices Reuters.
That access model matters for practitioners. It would move evaluation from periodic, black-box testing through an API to persistent visibility inside training and deployment workflows, including methods, controls and incident handling. The third element widens the aperture from single labs to collective action, first domestic coordination and then international agreements.
Amodei paired the plan with a stark capability warning. He warned that within six to 12 months AI could be capable of leading a swarm of agents that could take over the entire internet AP. The statement was not framed as a research result. It was framed as a reason to change operating tempo now.
Amodei also wrote that Anthropic "has always advocated for well-considered regulation of AI" The Verge. The plan itself remains a proposal. No lab has adopted it as policy, and no government has implemented it.
How rival labs responded
Reaction from other lab leaders split quickly. OpenAI CEO Sam Altman, Google DeepMind co-founder Demis Hassabis, and SpaceX CEO Elon Musk publicly indicated agreement with aspects of Amodei's three proposals The Verge. The agreement was partial, not wholesale endorsement of a slowdown.
Meta CEO Mark Zuckerberg opposed limiting companies' autonomy in AI development The Verge. Zuckerberg wrote that "every lab has the responsibility and incentive to move at the pace required to train its models safely" The Verge.
The disagreement is narrow but consequential. Both sides invoke safety responsibility. One side treats independent verification and coordination as necessary to meet that responsibility. The other treats lab-level discretion over pace as sufficient.
OpenAI's policy voice added a separate point. Global affairs chief Chris Lehane told The Verge that government has an important role in AI safety in addition to steps taken by companies. Lehane also said there is growing bipartisan support for mandatory national safety standards for frontier AI The Verge.
Washington and an industry regulator that stalled
The White House posture is hostile to the premise. The Trump administration called the idea of a safety crisis in AI a "hoax" The Verge.
A prior industry-led regulatory concept also failed to advance. The Wall Street Journal reported that Zuckerberg, Musk and Nvidia CEO Jensen Huang scuttled proposals for an industry-funded independent AI regulator modeled on FINRA The Verge. FINRA, the Financial Industry Regulatory Authority, is a private self-regulatory body for brokerages operating under government oversight. An analogous body for AI would have combined industry funding with independent audit authority, a structure familiar to enterprise compliance teams.
The broader context here is familiar to anyone who lived through earlier infrastructure transitions. Voluntary coordination tends to work for interoperability and incident sharing. It tends to fail when coordination requires participants to accept delay or cost while competitors might not.
In my view, the technical crux is the evaluator proposal. API-only evals can measure outputs. They cannot confirm whether safety procedures were followed during training, whether dangerous capabilities were observed and suppressed, or whether deployment safeguards match the tested configuration. Employee-like access would close part of that gap. It would also create hard problems around confidentiality, model security and evaluator independence that the current facts do not resolve.
Looking at what this means for builders, two tracks now run in parallel. One track is technical and operational, centered on eval design, access controls and cross-lab coordination mechanisms. The other is political, centered on whether mandatory national standards advance despite executive-branch opposition and the collapse of the FINRA-style alternative. For teams shipping frontier systems, the prudent assumption is that scrutiny of testing and pacing will increase even if formal rules do not. That scrutiny will reward labs that can document safe development in a way outsiders can verify, which is exactly the capability Amodei is asking the industry to build.
