The proposal TechCrunch described on September 16, 2026 is not a small thing. Frontier labs have historically treated external evaluators as either vendors, conference speakers, or adversaries. Embedding them inside the lab — with access to pre-deployment models and internal testing pipelines — is a structural change. But structure is not the same as independence, and the difference is the whole ballgame.

# What exactly are Anthropic and OpenAI proposing to embed?

According to TechCrunch, both labs want independent safety evaluators working inside their organizations, with access that outside researchers have generally not had. The framing is deliberate: "embedded" evaluators, not external auditors. That distinction matters because embedded reviewers sit inside the org chart they are supposed to scrutinize. TechCrunch reported that researchers welcome the access while warning that meaningful oversight requires transparency, independence, and eventually regulation.

Read the sequence of those three conditions carefully. Access is the easy part — labs control it and can revoke it. Transparency is harder — it requires deciding what gets published. Regulation is hardest — it removes the lab's veto. The proposal as described solves the first problem and defers the other two.

# Why does "embedded" cut against independence?

Independence in audit has a long institutional history, and it rarely survives when the audited party controls the auditor's funding, scope, and publication rights. That is not a claim about bad faith. It is a claim about incentives. An evaluator whose next contract depends on the lab's satisfaction faces pressure that no amount of personal integrity fully neutralizes.

Anthropic said the goal is to give outside experts meaningful visibility into model behavior before deployment. OpenAI has made similar arguments about external testing. Both positions are defensible on their face. But neither lab has committed to a mechanism that lets an evaluator publish a finding the lab disagrees with. Until that mechanism exists, "independent" is a description of the evaluator's résumé, not of the arrangement.

# What does the evidence actually support about effectiveness?

The honest answer is: not much yet. There is no published dataset showing that embedded evaluators catch more serious failures than external red-teamers, because the program is new and the findings are not public. TechCrunch reported that researchers welcome the access but warn about the conditions — which is the standard posture when evidence is thin but the direction is right.

What we can say with confidence is narrower. Access to pre-deployment models is a precondition for useful evaluation. Without it, external reviewers test last-generation systems and call it safety research. So the access component is real progress. The question is whether access without publication rights produces accountability or just better-informed insiders who cannot say what they found.

# How do the two labs' approaches compare?

# What would make this oversight rather than optics?

Three things, in order of difficulty. First, evaluators must be able to publish findings — including negative ones — without lab approval. Second, evaluators must have stable, non-revocable funding, ideally from a neutral body rather than the lab being evaluated. Third, there must be a regulator with statutory authority to compel access if labs stop volunteering it.

TechCrunch reported that researchers see regulation as the eventual endpoint. That is not cynicism; it is the standard trajectory of every voluntary safety regime, from financial auditing to aviation certification. Voluntary programs work until they become inconvenient, and then they become mandatory or they become theater.

# What are the falsifiable predictions?

1. Anthropic and OpenAI will not commit to unvetoed publication rights before 2028. If either lab does so before Q4 2027, this thesis is wrong.
2. The EU AI Office will cite embedded evaluator programs as insufficient for Article 15-style third-party conformity by mid-2027, pushing toward statutory auditor access.
3. At least one named evaluator will resign or publicly criticize the arrangement by Q3 2027, citing publication or funding constraints.

# Article summary

• Embedded evaluators solve the access problem and defer the independence problem — and the deferral is deliberate.
• Independence requires three conditions: publication rights, neutral funding, and a regulatory backstop. None are in the current proposal.
• The comparison between Anthropic and OpenAI is less interesting than their convergence: both offer access without authority.
• The most likely near-term accountability event is an evaluator departure, not a lab disclosure.
• Voluntary safety regimes historically convert to mandatory ones only after a visible failure. Plan for that failure, not against it.