Anthropic CEO Proposes METR as AI Safety Evaluator Amid Independence Debate

Summary

Anthropic CEO Dario Amodei has proposed that leading AI companies allow embedded third-party evaluators from the nonprofit Model Evaluation and Threat Research, or METR, to receive “employee-like access” to safety work. In an essay titled “We Must Pace the Frontier,” he said the evaluators should verify safety practices and commitments, report incidents, and assess the alignment of completed models as well as training pipelines and processes. Anthropic says it is committing to the step, and OpenAI CEO Sam Altman said OpenAI would also support independent evaluators with comparable access, although he did not specifically endorse METR. The proposal has drawn criticism because METR emerged from the Effective Altruism-aligned Alignment Research Center and has personnel and historical connections to people in Amodei’s professional and personal network. Critics also point to funding links involving donors who have invested in Anthropic, while METR says its funders do not control its projects, that conflicts must be disclosed, and that it accepts no funding from AI companies or their employees. The nonprofit says its purpose is to bring information from inside AI companies into the public domain rather than give a small group authority over the industry. The article reports that METR raised $71 million in outside funding over six months and details earlier grants to its incubator or affiliates. The proposal may face resistance from the Trump administration, which has criticized AI “doomer” warnings and has had a strained relationship with Anthropic. Former White House AI adviser David Sacks and other critics question whether companies are seeking safety coordination to reduce liability or shape regulation, while FTC Chairman Andrew Ferguson has warned against granting AI firms an antitrust exemption for such efforts. Amodei separately warned that AI bots could overtake the internet within six to 12 months without safeguards, potentially causing hundreds of billions of dollars in damage; the article presents that warning as part of the wider dispute over the pace and governance of frontier AI.