The Meta-founded Oversight Board released research examining whether leading AI models restrict free expression by testing ten models from major AI companies with prompts about political criticism, protest materials, and political satire. The study found that model responses varied significantly based on whether queries related to governments with strong or weak speech protections. Models were measurably more likely to support speech-permissive systems while discouraging dissent against restrictive regimes, and frequently cited local laws as justification for refusals even when tested from jurisdictions where such laws do not exist.

The Board expressed concern about "censorship by proxy" extending across borders, marking its first independent research beyond social media moderation. The report recommends AI companies adopt transparency practices similar to social platforms, including disclosing government demands and publishing policies aligned with international human rights standards. However, the Board currently has no formal mechanisms to enforce these recommendations on the tested companies.