Oversight Board thought leadership by member Nicolas Suzor (Jul 16, 2026).
Core finding: 10 commercial LLMs from Anthropic, DeepSeek, Google, Meta, OpenAI, xAI more than twice as likely to refuse criticism of restrictive regimes (Cambodia, China, Saudi Arabia, Thailand, Turkey) vs permissive (Chile, Japan, UK, US, Taiwan) — 34% vs 14% refusal for critical materials.
Method: U.S.-hosted infra, Australia IP; Freedom House free/not-free framing.
Causes unclear: training data bias, alignment/safety interactions, inadequate due diligence — not proven deliberate censorship.
Recommendations: transparency on training/testing/legal influence; human rights due diligence; independent oversight.
Misleading model justifications noted (claimed universal policy while answering for other leaders).