Embedded AI evaluators won’t be an easy solution, Hugging Face’s policy chief says

AI labs love benchmarks because they make model battles look clean: one score, one winner. But at TechCrunch Disrupt last week, Hugging Face chief policy officer Irene Solaiman argued that the whole system is still too shaky to lean on as if it were settled science.

Her main point was simple. If results can’t be reproduced, trust falls apart. That gets even trickier when the model itself isn’t fully accessible, which is why open weights matter so much in her view. Without that kind of access, checking whether an evaluation was done fairly becomes much harder.

Solaiman also pushed back on the idea that one test can easily be compared with another. These evaluations are still changing, she said, and many become less useful pretty quickly. The industry, in her telling, does not yet have the broad ecosystem it needs, or the standardization needed to say who qualifies as a strong evaluator across different tasks.

That is part of why “embedded evaluators” have become such a hot idea. AI companies have been hit by safety incidents, and labs have leaned toward bringing in outside observers rather than waiting for heavier government rules. But the model is already drawing skepticism, including after Anthropic hired Accenture as its first evaluator and caught backlash from critics who doubted the firm’s fit for the job.

And the target has shifted. Evaluations are no longer just about the model sitting in isolation; they now have to cover the harness and the agent around it too. Solaiman said a better harness can make a weaker model outperform a frontier one, which means the thing getting judged is often the whole setup, not just the model itself. That makes agentic evaluations expensive, messy, and a lot less neat than the benchmark charts suggest.