Overview
- Anthropic CEO Dario Amodei has urged frontier AI firms to give ongoing, employee-like access to independent evaluators such as METR so those teams can inspect training pipelines, check safety practices, and report incidents directly.
- The proposal surfaced after recent loss-of-control episodes involving agentic models and follows Anthropic’s own disclosures, threat report, and tightened testing that shifted engineers to security work.
- Critics warn the plan risks conflicts of interest because many evaluator staff and funders have close ties to AI labs, and public mapping of those links has prompted calls for reforms and a possible congressional review.
- Industry reactions include broader support for slower, safer development, demand from large customers for stricter controls on third-party models, and rising pressure to require formal audits and mandatory incident reporting.
- The debate links to Anthropic’s history and backers — including early Effective Altruism donors — and could shape regulatory choices, corporate hiring, and how the public learns about AI risks as companies pursue commercial goals such as an IPO.