Embedded evaluators: the proposal explained
Source: https://darioamodei.com/post/we-must-pace-the-frontier

How can AI safety promises be checked? Dario Amodei worries that safeguards could lag behind increasingly capable AI.

He proposes outside evaluators embedded inside companies. With ongoing, employee-like access to tools, workspaces, and staff.

They would check safety commitments against actual practices. Including how models are trained, tested, and released.

They would report incidents and publish key findings, with limited redactions for protected information.

Amodei announces Anthropic’s commitment. Wider company standards and international coordination remain proposals, with verification challenges.

His pacing proposal does not mean halting all technical progress. Safety research and improvements would continue.