Behind AI industry’s call for independent reviews, a fight over what gets made public

Anthropic and OpenAI are seeking control over what independent evaluators can publish, fearing revelations of trade secrets, according to sources familiar with the talks. Anthropic CEO Dario Amodei has called for evaluators to have the right to publish key findings without editorial control. OpenAI agreed to this stance shortly after.

The companies are concerned that evaluators working inside their labs might leak proprietary information to competitors. To prevent leaks, companies are considering measures like having lab employees observe embedded teams and preventing evaluators from working simultaneously at different labs. However, what is considered critical intellectual property can overlap with information used to assess a model's risks, leading to arguments and redactions.

The discussions about independence and publication rights between the evaluators and AI labs have evolved, with Anthropic now stating that it wants reviewers to control their methods and conclusions, while still reserving the right to comment on reports. OpenAI, meanwhile, says it only redacts confidential information from evaluator reports and allows them to disclose if major redactions were made.

The outcome of these talks could determine how powerful AI systems are scrutinized, as the developers warn they could advance beyond human control.