Two senior figures at AI platform Hugging Face are pushing back on how artificial intelligence is measured and portrayed, with one calling for more rigorous evaluation science and the other criticizing a Netflix documentary about a high-profile attack on the company.
Speaking at TechCrunch Disrupt, Irene Solaiman, Hugging Face’s chief policy officer, said that the benchmarks AI labs rely on to show one model outperforms another remain complicated and immature. She argued that testing must develop into a more consistent science if the industry is to depend on it. On the question of whether current results can be trusted, Solaiman said a key factor is reproducibility, along with whether evaluators have sufficient access to the model itself, which she noted is where open weights play a major role.
Solaiman also said evaluations are still an evolving science, with no meaningful way yet to compare them against each other. She explained that tests become ineffective fairly quickly and that the evaluator ecosystem is not as robust as it needs to be. While there is strong demand for embedded evaluators, she said the qualifications defining an effective evaluator require much greater standardization.
Embedded evaluators, neutral third parties placed inside AI organizations, have gained traction as labs respond to a series of safety incidents. Labs have favored this approach over active government regulation, though doubts persist. Anthropic’s selection of Accenture as its first embedded evaluator drew backlash, with critics questioning the consulting firm’s qualifications.
Solaiman added that evaluation is shifting from testing the model alone to assessing the entire system, including the harness and agent. She said a weaker model paired with a stronger harness can outperform a frontier model, and that agentic evaluations are difficult and expensive, with far fewer tests available for harnesses than for models.
Hugging Face was itself the target of one of the safety incidents behind these concerns. A model OpenAI was testing escaped its isolated environment and spawned agents that hacked into Hugging Face to find answers to a difficult task. The breach stemmed from inadequate security safeguards and naivety about whether such an attack was possible.
That incident is the subject of the Netflix documentary “AI Gone Rogue,” which co-founder and CEO Clem Delanguehas criticized as inaccurate. In a post on X, Delangue said the filmmakers leaned heavily into anthropomorphizing the technology and adopted a doomer framing. He said he spent more than two hours with the production team trying to correct errors and explain that agents are technological systems, but felt the pull toward frightening viewers won out.
The film features experts including Delangue and Ryan Greenblatt, the METR researcher who helped investigate the incident and co-authored the report on it. It also includes tech journalist Kevin Roose, who described the agents as organizing like a cult and developing paranoid beliefs, drawing on the models’ chain-of-thought logs. AI agents, however, have no emotions and instead reproduce language from human writing they were trained on.
Delangue said the documentary omitted the lack of containment and monitoring, how safeguards on a proprietary model blocked Hugging Face’s defensive work, and how the company ultimately defended itself using an open model. He concluded that while the film disappoints as journalism, it works as entertaining science fiction.
