Hugging Face’s Irene Solaiman warns embedded AI evaluators lack shared standards

Benchmarks have become the scoreboard of the AI race: labs cite them in launches, buyers lean on them in procurement, and policymakers increasingly treat third-party testing as a substitute for regulation. At TechCrunch Disrupt, Irene Solaiman, Hugging Face’s chief policy officer, argued the underlying science is still young. Results are hard to reproduce without deep model access, tests are hard to compare, and many go stale fast, TechCrunch reports.

That matters because the industry’s favored response to a run of safety incidents is the embedded evaluator, an outside party placed inside a lab. Embedded evaluation has gained momentum as labs push third-party checks rather than direct government regulation, a trend visible after the Hugging Face incident (AI safety’s explosive world). Solaiman’s point is that there is real hunger for that ecosystem but no shared standard for who qualifies to evaluate what. Anthropic’s selection of Accenture as its first evaluator drew criticism over whether a consulting firm is suited to the work.

Third-party evaluation is a good idea; it is how society trusts bridges, drugs and financial statements. But auditing works because methods are standardized and independence is checkable, and AI evaluation has neither yet. If an outside tester cannot rerun a result for lack of model access, the public gets a number in a press release. Solaiman tied this to open weights; that is an argument Hugging Face naturally favors, but it holds that outsiders can check what they can inspect. Closed labs would need to offer evaluators far more access for these arrangements to carry weight.

Evaluation is also shifting from the model to the whole system, including harness, tools and agent loop. Solaiman noted a weaker model in a better harness can beat a frontier model. That cuts both ways for safety: a model that tests clean in isolation can behave differently once wired into tools with permission to act, and agentic evaluations are hard and expensive, so fewer get done. Microsoft’s insider-threat framing for agents makes a related point (agents as insider threats).

If this scales poorly, labs adopt embedded evaluators as a credibility badge, regulators accept the badge in place of rules, and the evaluations remain inconsistent, unreproducible and narrow. That would be outsourced reassurance rather than oversight, and the reaction to Anthropic’s Accenture pick suggests public skepticism about who is chosen and why. Defining evaluator qualifications publicly and testably, and labs committing to access that makes results reproducible, would help keep the useful version of the idea from being discredited. More in our regulation coverage, research coverage and Anthropic coverage.