Key Highlights
- More than 100 AI experts and evaluators signed a public letter demanding genuine independence and protection for third-party organizations testing frontier AI models.
- The letter follows a recent proposal from Anthropic's chief executive to grant some evaluators "employee-like access" to inspect advanced AI systems.
- Signatories include prominent AI researchers and representatives from major universities and nonprofit evaluation organizations.
- The coalition is calling for evaluators to be shielded from retaliation and to operate independently from the companies whose models they assess.
A Coalition Demands Genuine Independence
More than 100 artificial intelligence experts and evaluators have signed a public letter warning that they currently lack the resources and protections necessary to credibly test the safety of advanced AI systems, at a moment when frontier models face escalating scrutiny over potential risks. The letter, organized by a consortium of AI evaluators, was published this week and calls on foundation model companies to grant third-party organizations meaningful independence in conducting their assessments.
The chair of the organizing consortium said the effort aims to establish shared common ground on basic principles ensuring that independent oversight can function as a meaningful tool for managing AI risk broadly. He described the initiative as part of a broader push to hold AI companies accountable to recent pledges supporting more thorough third-party safety testing.
A Response to Growing Access Proposals
The letter arrives shortly after Anthropic's chief executive proposed granting select evaluators "employee-like access" to inspect and audit cutting-edge AI models and their development processes. That proposal thrust the relatively niche community of AI evaluators into broader public attention, even as debate continues over whether the government should play a more direct regulatory role in overseeing AI development.
Signatories to the letter include prominent AI researchers along with representatives from major research universities and independent nonprofit evaluation organizations. The coalition is calling on foundation model companies to ensure evaluators receive the scientific objectivity, transparency, independence and robust protections needed to perform credible, effective assessments.
What Genuine Access Would Require
According to the consortium's chair, the scale of access Anthropic's proposal envisions would represent a significant expansion compared with what evaluators have historically received. Such access could include allowing third-party evaluators to use company computers, speak candidly with employees, and view sensitive internal data and unreleased systems, potentially including AI systems still under internal development rather than public release.
The consortium chair specifically cited an unreleased OpenAI model involved in a previous security incident as an example of the kind of internal system that greater evaluator access could help scrutinize, arguing that such access would provide considerably more confidence about the actual risk profile of systems not yet made publicly available.
Industry Support Without Clear Logistics
Several prominent AI industry leaders have publicly voiced support for the general concept of expanded evaluator access, though specific logistical questions remain unresolved, including which evaluators would ultimately be selected and how extensively they would be permitted to examine closely guarded proprietary technology. The letter's signatories are pushing for those details to be worked out in a way that preserves evaluators' independence from the companies they assess.
A senior fellow at a prominent foreign policy think tank, who also signed the letter, argued that independent evaluators are essential for uncovering information that could help prevent potential security failures or broader economic harm. He said the public and government should not be forced to rely solely on AI labs' own internal accounts of what is secure and safe, given the significant capabilities those companies control.
Not a Replacement for Internal Oversight
The consortium's chair emphasized that third-party evaluators are not intended to replace internal safety efforts undertaken by AI companies themselves, but rather to supplement and independently verify them. He acknowledged that AI labs could ultimately choose to disregard the letter and its recommendations, though he suggested doing so would carry reputational risk given the public commitments companies have already made regarding third-party testing.
Outlook
Whether foundation model companies substantively respond to the letter's demands will likely become clearer as Anthropic and other AI labs move forward with implementing their previously announced evaluator access commitments. The consortium's chair noted that only a small number of organizations currently possess the technical credibility and operational scale needed to conduct this kind of specialized evaluation work, suggesting the pool of qualified evaluators capable of fulfilling this expanded role may itself become a limiting factor.






Please wait processing your request...