Anthropic and OpenAI face questions over what outside AI evaluators can publish

Anthropic and OpenAI have sought control over what outside AI evaluators can publish, Reuters reported on October 13. Publication terms can affect how reviewers explain their findings to the public. The companies gave Reuters differing accounts of the limits they seek.

What the companies say about publication rights

Two people familiar with the discussions told Reuters that the companies were seeking strict editorial control as recently as October 5. One said Anthropic’s position appeared to have changed by October 8. Anthropic disputed that the characterization applied to all its discussions: it said established evaluators controlled their methods, criteria, conclusions and publication, while it reserved a right to comment on and approve reports from new evaluator firms as their relationships and model access developed.

OpenAI told Reuters it did not seek approval rights over full reports. It described redacting confidential information as standard practice and said evaluators could disclose whether major redactions had been made. Reuters said a contract it reviewed barred evaluator firms from publishing confidential information without company consent.

How access and feedback affected one evaluation

In an investigation published in August into an incident involving OpenAI and Hugging Face, METR said two of its employees and one Redwood Research employee spent six days working on-site at OpenAI. METR described limits on the access and data available to its investigation. Reuters reported that METR and Redwood changed their report’s structure, emphasis, clarity and tone after OpenAI feedback. The groups said no other important material affecting their conclusions was removed. Reuters said concern about deterring developers from inviting outside evaluators shaped their drafting and redaction decisions.