A recent high-profile investigation into an incident at AI platform Hugging Face has highlighted a widening philosophical and methodological rift between the artificial intelligence safety community and traditional cybersecurity professionals. Much of the public understanding of the Hugging Face event stems from a joint inquiry conducted by Redwood Research and METR, two organizations closely aligned with the AI safety movement. However, critics argue that this lineage has skewed the overarching narrative, contending that the subsequent investigations concentrated heavily on the specific actions taken by the AI systems rather than analyzing the broader surrounding security infrastructure.
METR President Chris Painter addressed criticisms regarding the organizations’ perspectives, noting that while METR strives to recruit from diverse worldviews, it has encountered difficulties attracting individuals who remain skeptical of catastrophic loss-of-control scenarios to work on defensive safeguards. When asked whether researchers who have spent years anticipating dangerous AI behaviors might inadvertently interpret ambiguous new evidence as validation of their prior predictions, Painter acknowledged that he worries about that exact possibility.
Despite those concerns, Painter maintained that METR’s empirical findings do not always align with researchers’ initial expectations. The organization, which is widely recognized for monitoring the rapid technological advancements of artificial intelligence, published a study indicating that advanced AI coding assistants actually slowed down experienced human software engineers rather than improving their efficiency. Emphasizing the organization’s broader objective, Painter stated that their goal is to inform the public rather than persuade people of a specific viewpoint, adding that he welcomes calls for a wider and more diverse range of independent evaluators to assess AI risks.
