Anthropic and OpenAI have both put forward plans in recent weeks to give external safety evaluators a more formal place inside their organizations. Anthropic's proposal includes desks, access badges, company laptops, and permissions roughly equivalent to those of its internal risk teams, along with contracts that support publishing key results except in narrowly defined security or confidential cases. The company said the arrangement is meant to test the viability of embedded external reviewers.
OpenAI responded days later with a framework that focuses on defined, mutually agreed assessment claims, explicit methodology, relevant technical expertise, and conflict-of-interest disclosure. A separate coalition, the AI Evaluator Forum, which includes METR and AVERI, published a letter setting minimum conditions such as transparency, protection from retaliation, and access matching the AI companies' own privileged employees. The letter also cautioned that embedded evaluations are a supplement to broader oversight efforts, not a substitute.
The proposals agree on the core need for independent visibility into advanced AI development, but they differ in emphasis, with Anthropic stressing physical access and publication rights and OpenAI stressing scoping and disclosure discipline. Observers quoted in the CNBC report see the shift as a response to the realization that advanced systems cannot be rolled out without public confidence. At the same time, the two companies remain competitors aiming at enormous valuations, and one expert said there is notable personal distrust between them even where they share goals on evaluation.