AI labs are embedding evaluators, but the rules for paying them are still unsettled

The next phase of AI safety is not just about model benchmarks or red-teaming reports. It is about who gets to sit inside the lab, what they can see, and whether they can speak freely when they find a problem.

In its September 18 announcement, Anthropic says it is partnering with Accenture, with the work led by Faculty. The company describes evaluators working inside the lab with access comparable to employees. It says it will directly fund Accenture’s work, while access, reporting standards and longer-term funding remain unsettled. This is Anthropic’s stated plan, not independent verification of its implementation.

CNBC’s October 11 reporting examines the growing role of third-party evaluators and the unresolved questions about their funding, access and reporting. Its interviews offer a broader view of the emerging field. They do not establish that an embedded arrangement has already delivered independent oversight or measurable safety improvements.

Our analysis concerns how a buyer might inspect an evaluation arrangement. Access to internal work could permit different questions from a test limited to model outputs. That difference does not make either approach inherently superior. A proposed review would ask what evidence a named assessment can actually obtain, and what conclusions its scope can support.

We propose a bounded document review before relying on a vendor’s evaluation claim. Compare the stated remit with the evaluator’s access, reporting terms and funding. These are suggested questions for a purchasing decision, not an established standard or evidence that a particular lab lacks safeguards. Missing documentation should remain unknown rather than become proof of either independence or failure.

Analysis: the procurement questions that decide whether embedded evaluation is real oversight

The following hypothetical criteria are our editorial proposal for inspecting a contract. They have not been tested as predictors of evaluator behavior. A buyer could use them to identify questions requiring a documented answer.

• Who pays, and for how long? Direct lab funding may be practical for launch, but it also creates a dependency risk. A proposed review could ask whether losing a client would materially affect the evaluator’s ability to continue its work. A more resilient design would spread funding across multiple customers or pooled sources, but Anthropic says that system does not yet exist.
• What access is required? “Employee-like” access sounds strong, but it can mean very different things in practice. Does the evaluator see training runs, internal incident logs, safety review meetings and deployment gates? Or only selected summaries? The proposed review could map each access permission to a specific test and record which claims remain outside its scope.
• What must be reported, and to whom? A proposed review could distinguish a confidential management report from a public statement. If they can publish directly, the lab may resist broad access. A realistic procurement document would need to specify whether reports go to the safety team, the board, outside customers, regulators or the public, and under what trigger conditions.
• What happens after a bad finding? Independence is not just the right to look. It is the right to disagree without losing the engagement. A serious contract would define escalation steps, remediation timelines and whether the evaluator can remain on the contract while documenting unresolved risk.
• How is conflict handled when the evaluator is also a vendor? The evaluator may want future work from the same lab or from the same industry. That does not automatically disqualify it, but it does mean the contract needs disclosure rules, rotation policies and a clear record of who reviewed the report before publication.
• What is the fallback if the company and evaluator disagree? In a field without settled standards, the answer may be that there is no neutral referee yet. That makes the governance layer as important as the technical test itself. A proposed review could document who considers a disputed finding and whether the unresolved disagreement appears in the final report.

The important limit here is that none of these arrangements proves safety on its own. They only make safety claims more inspectable. Anthropic’s own announcement is explicit that standards for access and reporting are still being worked out, and that long-term funding ideally comes from pooled or government sources rather than a single company. CNBC’s reporting suggests the broader market agrees that the field is moving fast, but not yet in a stable structure.

Our proposed acceptance record would list the model version, assessment date, scope, evidence available and findings still disputed. It could then state which purchasing decision the report informs. A favorable result for one configuration should not be assumed to cover another configuration without evidence. This is a proposed way to keep the decision traceable, not a claim that any specific evaluation has passed these checks.

Before relying on an embedded evaluation claim, request its documented access scope, funding arrangement and reporting terms, and record any unanswered questions.