100+ Researchers Set Minimum Conditions for Independent AI Evaluator Access
Miles_Brundage · x · 2026-09-18
More than 100 AI researchers and experts, including AVERI, have endorsed a set of minimum requirements for embedded external AI evaluation, responding to labs' recent calls for independent oversight.
- Meaningful independence: no conflicts of interest between labs and evaluators
- Strong access: employee-level system access guaranteed
- Transparency and editorial control: evaluators decide what to publish
- Protections against retaliation for unfavorable findings
- Diversity of perspectives and competencies
The post argues voluntary arrangements alone won't produce credible evaluation — these principles need to be embedded in policy rather than ad hoc one-off agreements, and labs must be accountable for meeting them so the public can trust the process.
More from Companies & People
- John Lewis's worst-reviewed Christmas ad: why it wants to stop cash gift cards — aakashgupta · 2026-11-04
- Technical founders: as CEO, storytelling is the core job you can't skip — paigeinsf · 2026-11-04
- Andreessen: Agents Are Entering Enterprises as Chaos, Just Like PCs and Spreadsheets Did — Dan_Jeffries1 · 2026-11-04
- DeepMind, Cognition host bouldering social ahead of AI Engineer Code Summit — osanseviero · 2026-11-04
- Google says it removed post-employment non-disparagement clauses from new-hire terms — HaydnBelfield · 2026-11-04
- Accenture Accused of Publishing AI Paper Littered with AI Hallucinations — nordicinst · 2026-11-04