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

What happened

At a TechCrunch Disrupt event, Irene Solaiman spoke about how AI models are tested. She said the methods are still new and need to become more consistent. She also questioned whether AI companies can use outside checkers, called embedded evaluators, to judge their models fairly.

Why tests are hard to trust

AI labs use benchmarks, which are tests that claim to show whether one AI product is better than another. To trust a test, people need to be able to repeat it and get the same result. That is hard to do unless they can access the model itself. Solaiman said open weights, which means the model's inner settings are shared publicly, help a lot with this. Tests also get less useful quickly, and there is no common way to compare one test to another.

Outside checkers and the harness

Labs have turned to outside 'neutral' observers after a series of AI safety problems, instead of relying on government rules. Anthropic hired the consulting firm Accenture as its first evaluator, and many critics doubted that choice. Testing is also shifting from the model alone to the whole system, including the harness, which is the tool people use to reach the model. A weaker model with a better harness can beat a stronger one.

Key facts

• Solaiman spoke at TechCrunch Disrupt last week.
• Anthropic hired Accenture as its first embedded evaluator.
• A weaker model with a better harness could outperform a frontier model, according to Solaiman.
• Solaiman said there are more evaluations for models than for harnesses.
• Agentic evaluations, which test AI agents that act on their own, are 'really hard and really expensive.'

Why it matters

AI companies use test scores to show their products are good, and buyers and leaders make choices based on them. If the tests are not repeatable or fair, those choices may be wrong. Solaiman's view suggests the industry needs shared rules for how tests are built and who is qualified to run them.