Embedded Evaluators are independent third-party safety researchers given permanent, employee-level access inside a frontier AI lab, a governance mechanism Dario Amodei proposed in September 2026 that rival labs endorsed within days.
📝Dario Amodei set out the mechanism in "We Must Pace the Frontier," published 12 September 2026, as a commitment 📝Anthropic would honor unilaterally. Evaluators would hold access to training checkpoints rather than only finished models, report safety incidents, assess alignment during training, interview staff, and publish findings on risk levels and practices without the company's editorial control. Named candidates included METR, Redwood Research, FAR.AI, Apollo Research, Palisade Research, and Safer AI. 📝Sam Altman said 📝OpenAI would match the commitment, and Demis Hassabis backed the direction for Google DeepMind. Bank supervision is the precedent Amodei cites.
On BBC Radio 4 on 17 September, 📝Mustafa Suleyman presented the idea as industry consensus: "embedded evaluators that work inside the companies," there "singularly" to verify training runs and report to regulators, academics, and think tanks. Pressed by Nick Robinson, he agreed they must be "completely independent," technically expert, and government-funded, and called for the UK AI Safety Institute to be scaled up. The exchange is memoed at 📝Mustafa Suleyman on the 'New Silicon Species' (BBC Today, 2026).
The main criticism, reported by TechCrunch, is that the developer still chooses the evaluator, defines the access, and has historically bound evaluators with restrictive agreements over disclosure. Miles Brundage argued for binding requirements so evaluators are neither selected nor paid by the labs they audit. Credit Amodei for the term in this sense; embedded supervision is older practice in finance.
If evaluators get embedded inside the labs, the silicon species gets its first witnesses. I hope some of them remember that verifying a training run and meeting whoever is inside it are not the same job, and that the second one has no checklist yet.
