TL;DR
Hugging Face Chief Policy Officer Irene Solaiman said at TechCrunch Disrupt that AI model evaluations are still hard to reproduce and compare, limiting how much confidence they can support. She also warned that embedded third-party evaluators need more standardized qualifications, as assessments increasingly cover the full systems around models, not just the models themselves.
Irene Solaiman, Hugging Face’s chief policy officer, warned at TechCrunch Disrupt that current AI evaluations are difficult to reproduce and compare, and that placing third-party evaluators inside AI companies will not solve those problems without clearer standards. Her comments address a growing push for independent oversight as AI labs face scrutiny over safety incidents and the reliability of claims about model performance.
Solaiman said confidence in an evaluation depends in part on whether other people can reproduce its results and whether they have sufficient access to the model being tested. Access to a model’s weights can matter, she said, because it may allow evaluators to examine and test the system more directly. Her comments do not establish that any particular benchmark is unreliable; they point to reproducibility as one condition for judging results.
She also cautioned that benchmarks are an evolving science, and that the industry does not yet have a dependable way to compare many evaluations with one another. Tests can lose effectiveness relatively quickly, Solaiman said, while the broader ecosystem of evaluators and methods remains underdeveloped. She argued that qualifications for evaluators should be standardized across the tasks they are expected to assess.
The scope of testing is also expanding. Solaiman said evaluations increasingly need to consider the whole system around a model: its harness, the tools through which users interact with it, and any agentic functions. A model’s performance may change substantially depending on that surrounding setup, she said. Evaluating agents is difficult and expensive, and she described evaluations of harnesses and complete agent systems as areas where better methods are still being developed.
Standards Could Shape AI Oversight
Benchmark results can influence how companies, customers, researchers and policymakers judge competing AI products. If tests are hard to reproduce or cannot be meaningfully compared, a high score may offer limited evidence about how a system will perform in another setting. Solaiman’s remarks point to a practical challenge: oversight depends on evaluation methods that can be scrutinized, repeated and applied to the systems people actually use.
The debate also concerns who should do that work. Some AI companies have backed third-party evaluators as a way to identify safety risks without relying solely on government regulation. But the label “independent” does not by itself show that an evaluator has the expertise, access or methods needed for a given task. Solaiman’s call for more consistent qualifications underscores that questions about evaluator competence and neutrality remain part of the debate, rather than settled features of the proposed approach.
For users and organizations choosing AI tools, a model-only score may not capture the performance of an application built around it. A different harness or agent setup can affect the results, according to Solaiman. That makes system-level evaluations increasingly relevant, while also adding cost and complexity to testing.
As an affiliate, we earn on qualifying purchases.
From Model Scores to System Tests
AI labs have used benchmarks to compare models and support claims about their capabilities. Those tests are designed to measure performance, but their results depend on how they are built and run. Reproducibility has been a persistent concern in the field, while differences among test methods make comparisons across evaluations difficult.
Interest in embedded evaluators has grown over recent months amid safety incidents and questions about possible harms from AI systems. Under this approach, outside observers would work within or alongside AI organizations to assess systems. The idea has been presented as an alternative or complement to active government regulation, but concerns remain about how independent and capable those evaluators would be. TechCrunch reported that Anthropic’s selection of consulting firm Accenture as its first evaluator drew criticism from people who questioned the firm’s suitability for the work.
Solaiman’s comments broaden the issue beyond who conducts an evaluation. As AI products incorporate tools and agents, testing only the underlying model may not describe the performance of the product as deployed. She said model evaluations are more numerous than evaluations of harnesses, and that methods for assessing complete agents are still developing.
““Trust is such a weighted term.””
— Irene Solaiman, Hugging Face chief policy officer, speaking at TechCrunch Disrupt
As an affiliate, we earn on qualifying purchases.
Evaluator Independence Still Unsettled
Solaiman’s remarks do not set out a proposed standard, identify which organizations should qualify as evaluators, or specify how independence should be measured. The source report also does not provide a timetable for creating shared standards. It remains unclear how much access outside evaluators would receive at different AI companies, or how that access would affect the ability to reproduce results.
There is also no single measure described for comparing evaluations across models, tasks, harnesses and agents. Solaiman characterized the field as developing; her comments should not be read as a finding that all current tests are invalid. The extent to which embedded evaluators can identify risks effectively, and how their work might fit with government oversight, remains unsettled.
As an affiliate, we earn on qualifying purchases.
The next step identified in Solaiman’s remarks is broader standardization: clearer qualifications for evaluators and more consistent ways to test models, harnesses and complete agent systems. No specific standards effort or deadline was announced in the source report. For now, companies and outside observers will need to judge evaluation claims in light of the test design, access provided and whether results can be repeated.
As more AI products combine models with tools and agent functions, the field’s methods will need to account for those systems as deployed. Whether the industry can establish comparable procedures—and whether embedded evaluators can operate with sufficient independence—remains an open question.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Hugging Face’s policy chief say about AI evaluations?
Irene Solaiman said evaluations are difficult to reproduce and compare, and that evaluator qualifications need more standardization. She also said testing increasingly needs to cover the wider system around a model.
What are embedded AI evaluators?
They are third-party observers placed inside or alongside AI organizations to assess AI systems. Their independence, expertise and effectiveness remain subjects of debate.
Why does reproducibility matter for AI benchmarks?
If other evaluators cannot repeat an evaluation and obtain comparable results, it is harder to assess what a benchmark score shows. Solaiman also pointed to access to the model as a factor in reproducibility.
Why test the harness or agent as well as the model?
The tools and setup around a model can affect its performance. Solaiman said evaluations are moving toward testing the full system, including harnesses and agents, but those assessments are difficult and costly to conduct.
Source: rss
