Dario Amodei is putting skin in the game. The Anthropic CEO announced his company will unilaterally allow third-party evaluators inside its offices, granting them badges, desks, and near-full access to internal risk data. This move follows a wave of internal dissent, including researcher Jacob Coxon’s resignation over fears that leading labs are “gambling with our lives.” While Amodei didn’t name Coxon directly, he cited the recent OpenAI-HuggingFace hack and the accelerating speed of AI self-improvement as catalysts for this new caution. He argues we must slow capability gains to buy time for safety protocols to catch up.
The proposal goes beyond voluntary guidelines. Amodei wants these embedded evaluators, potentially from organizations like METR, to verify compliance and ensure incidents aren’t swept under the rug. Think of it as regulatory oversight baked into the corporate structure, similar to how bank regulators operate. But he knows going it alone isn’t enough. He’s calling for coordinated safety standards among democratic nations’ tech giants, acknowledging that antitrust laws currently block such collusion. His solution? A narrow government waiver to allow these critical safety conversations without triggering federal prosecution.
Critics aren’t buying the altruism. Journalist Brian Merchant dismisses the apocalyptic warnings as a distraction from current harms, labeling the proposal as classic regulatory capture designed to cement Anthropic and OpenAI’s dominance. There’s also the geopolitical elephant in the room: China. Amodei argues that restricting chip exports and cracking down on model distillation can widen the US lead by three to five years, creating space for global coordination on banning dangerous uses like biological weapon development. It’s a bold claim that assumes technical barriers can effectively stall a rival nation’s progress.
This isn’t just about safety; it’s about trust. Amodei admits the industry faces a credibility crisis, with public skepticism toward tech firms and government at an all-time high. By inviting outsiders in, he hopes to rebuild that fractured trust. Yet, the tension remains palpable. Can competitors truly coordinate without breaking antitrust laws? Will embedded auditors have the teeth to stop a runaway model? Or is this just a sophisticated PR maneuver to delay regulation while maintaining market lead? The clock is ticking, and the stakes have never been higher.
