Anthropic and OpenAI push for tighter control over external AI evaluator publications

Anthropic and OpenAI have been pressing for extensive control over the publications of outside evaluators, citing concerns that embedded reviews might expose trade secrets. Initially, both firms sought strict editorial oversight, but recent talks indicate a move toward letting evaluators release core findings independently, while still permitting the labs to comment on and redact sensitive material. Proposed measures to prevent leaks include having lab employees observe evaluator work and prohibiting evaluators from serving multiple labs at once.

Critics note that the line between proprietary data and risk-relevant information is blurry, risking disputes and redactions. The negotiations occur amid broader pressure to make powerful AI systems transparent, highlighted by a recent White House accord on super-intelligence and a public letter from over 200 researchers demanding one outlet's review. Without agreed standards, the industry risks allowing AI developers to shape the scope of safety assessments.