Hugging Face policy chief questions feasibility of embedded AI evaluators

Hugging Face’s chief policy officer, Irene Solaiman, argued at TechCrunch Disrupt that embedded evaluators—proposed as neutral third-party observers within AI organizations—are unlikely to provide a straightforward solution to safety concerns. She emphasized that evaluations are still an evolving science, lacking consistent standards and reproducibility, which complicates trust in their results.

Solaiman pointed to the need for open model access as a critical factor in enabling reproducible evaluations. Without such access, she suggested, it becomes difficult to verify claims about model performance or safety. She also noted that evaluations are not yet mature enough to meaningfully compare across different tests or tasks.

The discussion comes as AI labs face increased scrutiny over safety incidents, with some advocating for third-party evaluators as an alternative to government regulation. However, Solaiman questioned the neutrality and qualifications of such evaluators, citing criticism of Anthropic’s decision to hire Accenture as its first evaluator as an example of the skepticism surrounding these proposals.

Solaiman also highlighted gaps in evaluating agentic systems and harnesses—the tools users employ to interact with models. She described agentic evaluations as both difficult and expensive, and noted that evaluations for harnesses and holistic agent systems lag behind those for standalone models. This imbalance, she argued, risks creating blind spots in safety assessments.

Overall, Solaiman called for a more uniform, industry-wide approach to evaluations, suggesting that without standardization and better ecosystem support, benchmarking will struggle to serve as a reliable tool for assessing AI systems.