OpenAI has brought in four research groups to assess its models and safety measures, but their access will vary and the company retains a say over parts of the work. That makes the independence of the new checks as important as the names on the list.
OpenAI Watch analysis
What happened
The assessors are METR, Apollo Research, Transluce and Goodfire. Mother Jones reports that OpenAI’s terms allow it to withhold information for reasons including security or intellectual-property protection, and to select staff members assessors may interview. The arrangements differ by group: Apollo is set to work with OpenAI for four months, while Transluce’s access is limited to models, safety documents and researchers approved by the company. Goodfire describes its access as “deep”, without specifying its scope in the material reported.
Mother Jones also says METR’s earlier work following the Hugging Face incident involved six days on site in total. The article’s account of the new arrangements is reporting on OpenAI’s announcement and the assessors’ statements, not a completed independent evaluation of the terms in practice. Mother Jones disclosed that its parent organisation, the Center for Investigative Reporting, has sued OpenAI over copyright; OpenAI denies those allegations.
Why it matters
An assessment can only tell the public so much if the company being assessed controls what researchers can inspect and whom they can question. Different access arrangements may also make the groups’ findings difficult to compare. For people trying to judge whether safeguards work, the eventual conclusions will matter, but so will the boundaries around the work that produced them.
Our read
Naming outside assessors is a start, not a certificate of independence. The useful test is whether each group can explain what it examined, what it could not examine and how those limits affected its conclusions. “Deep access” sounds promising; a clear scope would be more useful.
What to watch
- What each assessor is permitted to examine, and whether those limits are published.
- Whether the groups can choose whom they interview and report findings without company approval.
- How the assessors describe the limits of their methods when their conclusions are released.
Discussion spark: Should companies being assessed retain a say over researchers’ access and interviewees, or does meaningful independent scrutiny require the company to give up that control?
Sources and evidence
- Source update (17 November 2026, 20:07 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.