Anthropic CEO Dario Amodei pledged in a September 12 essay to invite an independent team of embedded evaluators with continuing, employee-like access to the company. Their proposed role is to examine safety practices, report incidents and assess alignment throughout model training and deployment, rather than only reviewing a model before release.
The pledge is the first part of Amodei’s three-stage proposal for slowing frontier development enough to evaluate and safeguard increasingly capable systems. He distinguishes that pacing from stopping model training or technical progress. The essay is a statement of Anthropic’s intended practice and its CEO’s policy position, not an agreed industry standard.
What external access would involve
Amodei says evaluators would receive office access, equipment and permissions broadly comparable to internal risk-assessment teams. They could examine workspaces and tools and speak directly with employees. Exceptions would cover legal and contractual obligations and customer or partner privacy.
The proposed contract would let reviewers publish key findings about risks, incidents, practices and the access they received, without Anthropic’s editorial control. Narrow redactions could protect security-sensitive, legally privileged, commercially sensitive or third-party confidential information. Findings could not be removed simply for being unfavorable, and reviewers could disclose when a redaction affected their conclusions.
That arrangement is intended to make the company’s commitments verifiable over time. Amodei argues that occasional reports written by the developer leave the developer deciding what to disclose; embedded reviewers would have continuing access to the processes behind those reports. At the time of the essay, the team had not been named and no firm start date was given.
The wider pacing proposal
The second stage calls for coordination among frontier companies in democratic countries, including common safety standards. For some safety discussions, he calls for government mediation or narrow waivers of antitrust restrictions. A third stage seeks international cooperation, while acknowledging the difficulty of verifying compliance.
His stated rationale is that additional time could improve operational security, alignment research, interpretability and evaluations as models become more capable of acting through tools. Anthropic can act on its evaluator pledge itself; the broader stages depend on decisions by other companies and governments.