Anthropic and OpenAI Negotiate Terms for Independent AI Reviews
Anthropic and OpenAI are negotiating editorial control over independent AI safety reviews to protect trade secrets while complying with a White House mandate.
AI laboratories Anthropic and OpenAI are negotiating the terms of independent reviews for their models, specifically debating how much control the companies maintain over the publication of findings. While both firms publicly support editorial independence for external evaluators, internal discussions show concerns that published reports could leak proprietary information or trade secrets to competitors.
To mitigate these risks, the companies have considered restricting evaluators from working with multiple labs simultaneously and using employees to monitor embedded evaluation teams. These tensions are exacerbated by the labs' pursuit of high valuations for potential IPOs and the need to disclose risks associated with powerful AI. A recent report on a Hugging Face infrastructure hack involving OpenAI models required the redaction of nonpublic information, illustrating the conflict between transparency and security.
These negotiations follow a White House Accord on Super Intelligence signed by President Donald Trump, which mandates independent evaluation but lacks specific procedural standards. This gap has left labs and evaluators to negotiate their own access and transparency terms. Anthropic has appointed Faculty as its first embedded evaluator to begin this process.
