Hugging Face Policy Chief Urges Standards for Embedded AI Evaluators

Hugging Face chief policy officer Irene Solaiman called for clearer standards on which outside evaluators are qualified to assess specific AI systems, as inconsistent testing methods make safety results difficult to compare.

Speaking at an industry conference last week, Solaiman said a test’s credibility depends in part on whether others can reproduce its results. That can require access to the model itself, including its weights. She also said evaluation methods are still developing, making results from different tests hard to compare.

The problem extends to AI agents, which use software around a model to carry out tasks. That software, known as a harness, can affect performance: A weaker model paired with a better harness could outperform a stronger model, Solaiman said. Testing the model alone may therefore give an incomplete picture of how an agent performs.

“Agentic evals are really hard and really expensive,” Solaiman said. Harness evaluations are less common than model evaluations, while methods for assessing complete agents are still developing.

Those limits matter as AI developers consider giving outside evaluators access to their systems. Solaiman said the industry needs more consistent qualifications for evaluators across different tasks. Access alone does not establish whether an evaluator has the expertise to judge a particular system.

The question of access has also featured in earlier proposals for embedded evaluation, which called for independent reviewers to examine AI systems during training instead of waiting until shortly before release. Solaiman’s remarks focused on another part of that approach: how to determine what those reviewers are equipped to evaluate.