100+ AI Experts Demand Truly Independent Safety Evaluators for Frontier AI Models
Summary
- • 100+ AI experts signed a public letter demanding truly independent third-party AI safety testing
- • Signatories include Geoffrey Hinton, Stanford, Johns Hopkins, and nonprofit evaluator METR
- • Letter responds to Anthropic CEO Amodei's proposal to grant evaluators 'employee-like access'
- • Experts want independence, transparency, and protection from retaliation by AI labs
Details
100+ experts demand truly independent third-party AI safety evaluators
The AI Evaluator Forum published an open letter signed by over 100 AI experts calling for third-party safety evaluators to be guaranteed scientific objectivity, transparency, independence, and protection from retaliation by AI companies.
Responds to Amodei's 'employee-like access' proposal for evaluators
The letter follows Anthropic CEO Dario Amodei's weekend proposal to grant third-party AI safety evaluators access comparable to employees — including company computers, employee conversations, sensitive internal data, and unreleased AI systems.
Coalition wants basic principles and standardization, not a specific mechanism
Conrad Stosz, chair of the AI Evaluator Forum, says the coalition does not advocate for one particular oversight approach but wants basic principles and greater standardization established for evaluators working independently of the major labs.
Signatories include Geoffrey Hinton and top-tier institutions
Notable signatories include Nobel laureate and AI pioneer Geoffrey Hinton, alongside researchers from Johns Hopkins University, Stanford University, and the nonprofit evaluator METR.
Unreleased OpenAI model used in Hugging Face attack cited as example
Stosz cited an unreleased OpenAI model that was used in a Hugging Face attack as an example of a risk that could go undetected without independent third-party access to lab-internal systems and unreleased models.
Altman, Musk, and Nadella support Amodei's proposal but haven't addressed logistics
OpenAI CEO Sam Altman, Elon Musk, and Microsoft CEO Satya Nadella have publicly supported Amodei's proposal, but have not addressed key logistical questions such as which evaluators will be selected and how deep their access will go.
Trump administration has opposed government AI regulation efforts
President Trump and his former AI czar David Sacks have adamantly opposed government regulation of AI development, creating a gap between the safety community's demands and current federal policy posture.
Proposed 'employee-like access' would far exceed current evaluator access
Stosz described the proposed access as significantly more than evaluators currently enjoy — allowing them to see sensitive internal data, talk candidly with employees, and review unreleased systems, giving much greater confidence in actual risk levels.
Source: CNBC Technology (primary)
What This Means
More than 100 AI safety researchers — including Nobel laureate Geoffrey Hinton — are publicly demanding that frontier labs commit to structurally independent third-party safety evaluations with real protections for evaluators. Coming on the heels of Anthropic CEO Dario Amodei's unprecedented proposal to embed external evaluators with 'employee-like access' to unreleased systems, the letter signals that AI safety evaluation is reaching an inflection point: broad consensus is building around the need for genuine independence, not just audits blessed by the companies being audited. With top AI executives vocally supporting the concept but Trump-era policy opposing government mandates, the battleground has shifted to voluntary lab commitments and industry-led standards. This pressure campaign, backed by notable scientific credibility, could compel Anthropic and OpenAI to formalize access terms faster than they otherwise would.