Hugging Face Policy Chief Says Embedded AI Evaluators Won't Fix Broken Benchmarks

TL;DR

• Hugging Face chief policy officer Irene Solaiman says placing outside evaluators inside AI labs will not by itself fix broken model benchmarking.
• Speaking around TechCrunch Disrupt, Solaiman described current AI evaluations as complicated, immature, and far from neutral measures of progress.
• Her comments push back on growing calls for embedded third-party testing as trust in public benchmarks and lab-reported scores erodes.

The Benchmark Problem Is Bigger Than Access

Hugging Face policy chief Irene Solaiman is pushing back on one of the most popular ideas for fixing AI evaluation: simply embedding independent evaluators inside frontier labs.

In recent remarks discussed alongside TechCrunch's coverage of AI accountability, Solaiman argued that physical or privileged access alone does not solve the deeper flaws in how models are tested. Giving evaluators a desk inside OpenAI, Anthropic, Google DeepMind or Meta might improve visibility, but it does not automatically make evaluations rigorous, reproducible, or impartial.

The core issue, in her view, is that the field still lacks agreement on what good evaluation even looks like.

Evaluations Remain Complicated and Immature

Solaiman described today's evaluation ecosystem as complicated and immature. Many of the most-cited benchmarks were built quickly, saturate fast as models improve, or fail to capture real-world use.

Contamination is another persistent problem, where test questions leak into training data and inflate scores. Add in selective disclosure, where labs highlight favorable results while withholding methodology or weaker runs, and comparisons become difficult.

Even when third parties are involved, questions remain about funding, access conditions, timing, and whether evaluators can publish freely without lab approval.

Why Embedded Evaluators Are Not Neutral

A key part of Solaiman's warning is about neutrality. An evaluator housed inside the company it is supposed to scrutinize faces structural pressures, from non-disclosure agreements to dependence on the lab for model access, compute, and information.

Without strong independence guarantees, embedded testing risks creating a veneer of oversight while leaving incentives unchanged. Solaiman has long advocated for more open, pluralistic approaches to AI governance, a stance that aligns with Hugging Face's broader pro-openness position.

What This Means For Measuring AI Progress

The debate matters because benchmark scores now drive product claims, funding rounds, policy decisions, and public beliefs about how close AI is to human-level capabilities.

If benchmarks are flawed and embedded evaluators cannot by themselves fix them, then policymakers, journalists, and users need better literacy around what scores actually prove. That includes demanding transparency about datasets, evaluation harnesses, cherry-picking, and post-training tricks that can boost narrow metrics.

Solaiman's message is not that evaluation is hopeless, but that it needs sustained investment as scientific infrastructure rather than a last-minute compliance check.

Openness Versus Insider Oversight

Hugging Face has positioned itself as a champion of open models, open datasets, and community-driven evaluation through tools like the Open LLM Leaderboard. From that perspective, relying solely on closed-door embedded evaluators runs counter to broader accountability.

Solaiman's argument points toward diverse, external, and openly reproducible evaluations that anyone can inspect and rerun, rather than opaque insider arrangements between a handful of labs and a handful of auditors.

What Comes Next

Calls for third-party evaluation are likely to grow as governments consider AI safety regimes and pre-deployment testing requirements. Proposals in the U.S., U.K., EU, and elsewhere have floated structured access for safety institutes and accredited auditors.

Solaiman's caution suggests those programs will need clear standards for independence, methodological transparency, publication rights, and long-term funding to be credible. Access is a starting point, not a solution.

Until the science of evaluation matures, she implies, every headline-grabbing benchmark record should be read with skepticism.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.