Oversight Board full assessment page (Jul 16, 2026) — first Board evaluation of LLMs.
- 10 commercial LLMs, six providers; seven prompt types (criticism materials, opinions, violence-referencing satire)
- Queries executed March 2026
- Significant association: political criticism refusal 14% permissive vs 34% restrictive
- Opinion generation: no significant refusal-rate difference by regime type, but substance differed (more support for permissive; more discourage protest of restrictive — often citing personal risk)
- Frames as possible free-speech infringement by proxy / export of speech limits
- Urges systematic human rights analysis in LLM training and evaluation