AI evals org METR accused of being 'captured' by the companies it evaluates
repligate · x · 2026-11-16
A viral thread accuses AI evaluation nonprofit METR of being captured by the labs it evaluates: restrictive NDAs allegedly stop the org from sharing the very information it exists to share, and staff enforce compliance even on researchers like Jacob Coxon to preserve in-group status with AI companies.
QuotedBeth Barnes pushes back on deontological grounds: if people break agreements whenever they believe stakes are high, others will stop trusting or dealing with them.
The exchange highlights a core tension in AI safety evaluation—dependence on the evaluated companies may compromise independence and transparency.
More from AGI Musings
- Researcher's field note: AI aids ~60% of research tasks, but oversight is constant and critical — fredahshi · 2026-11-16
- China could win the compute race by 2030, triggering US AI stimulus — robleclerc · 2026-11-16
- Physician pushes back on AI-replaces-doctors hype: hospitals won't buy the sales pitch — PAstynome · 2026-11-16
- AI researcher to anxious peers: frontier labs won't take over everything, diffusion is the bottleneck — Guodzh · 2026-11-16
- AGI 'natural disasters' likely but reversible and forgivable, argues researcher — repligate · 2026-11-16
- Frontier model weights lose value over time, argues Arohan — AI value lies in running and improving them — _arohan_ · 2026-11-16