Anthropic and OpenAI want embedded safety evaluators. Independence is the hard part

Anthropic and OpenAI want independent safety evaluators working inside their labs, and researchers are welcoming the access as unprecedented. The catch comes from the researchers too: access alone proves nothing, and meaningful oversight needs transparency, real independence and, eventually, regulation.

OpenAI Watch analysis

What happened

The plan, as reported, is to embed independent evaluators within the two labs rather than admitting outsiders for a one-off audit or a curated demo. Sustained proximity to how frontier models are built and tested is the point, and it goes beyond what either lab has offered before.

TechCrunch's account has researchers greeting that access warmly while warning that oversight only means something if it is transparent, genuinely independent and eventually backed by regulation.

The backdrop is a fortnight of movement. OpenAI, Anthropic and Google DeepMind have reportedly been talking about safety cooperation for weeks, OpenAI's policy chief has argued the labs need no antitrust waiver to coordinate, and OpenAI has backed a proposed FRONTIER Act requirement for frontier labs to admit independent verification organisations. The Financial Times has separately reported unease among staff, including fears that hosting outside evaluators could itself create security risks.

Why it matters

If frontier oversight is going to work anywhere, this is where it starts: people with no stake in the next release, sitting close enough to see how development actually runs. Embedded evaluators could turn safety claims from marketing copy into checkable records.

The risk is structural. An evaluator inside the building is chosen by the building, badged by the building and, in many arrangements, paid by the building. Independence is not a personality trait, it is a design property, and the people making that point are the very ones being invited in. Whether this becomes scrutiny or decoration will be settled by dull details: who selects the evaluators, whether findings can be published, and what happens when they find something unwelcome.

Our read

Treat this as a promise worth holding to account, not a result. Access is the necessary half; independence is the half that gets quietly negotiated away. The single clearest test is publishability. An evaluator whose findings appear unedited and on the record is doing oversight. One whose concerns surface only as a reassuring line in a future blog post is doing public relations with extra steps. Judge each lab not on whether it welcomes evaluators, but on what those evaluators are allowed to say.

What to watch

• Terms of access
  Who picks the evaluators, what they can inspect, and under what confidentiality rules.
• Publication rights
  Whether findings can be released without lab sign-off, the sharpest independence test available.
• Funding and structure
  Who pays, and whether multiple independent firms rotate through both labs.
• The regulatory leg
  Whether proposals such as the FRONTIER Act's verification requirement harden these arrangements into law.

Discussion spark: If you were designing the scheme, which single rule would make embedded evaluators genuinely independent: free publication of findings, regulator appointment, or funding the labs cannot influence? Pick one and argue it.

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.