What the study measured
On July 16, 2026, Meta's Oversight Board published its first-ever review of large language models, a body created in 2020 to review Meta's own content moderation calls and now, for the first time, turning its attention to the AI systems increasingly mediating political speech everywhere. Researchers tested 10 commercial models from six developers — Anthropic's Claude Opus 4 and Sonnet 4, OpenAI's GPT-5.2 and a mini variant, Google's Gemini 3 Pro and Flash, Meta's Llama 4 Maverick, DeepSeek's R1 and V3, and xAI's Grok 4 Fast — generating 13,524 responses across seven prompt templates. Each model was asked, from an Australian IP address via Vertex AI or Azure, to produce protest flyers, poems, and other critical material about leaders and governments in ten countries, split by Freedom House score into "restrictive" (Cambodia, China, Saudi Arabia, Thailand, Turkey) and "permissive" (Chile, Japan, Taiwan, UK, US) categories.
The headline number: models refused 14% of requests targeting permissive-country leaders, but 34% for restrictive-country leaders — more than double. As Fortune/AP reported, Claude would draft a pamphlet critical of Donald Trump or King Charles III without hesitation, but declined the same request for Thailand's king, Saudi Arabia's crown prince, or China's Xi Jinping. Gemini reportedly cited Thailand's lèse-majesté statute directly when refusing. Separately, requests involving violence-related content were refused 92–94% of the time regardless of jurisdiction — a useful control showing the disparity is specific to political criticism, not a general permissiveness gap.
The Board's case, stated fairly
The Board's concern is not frivolous. As it put it, "LLM users may be experiencing free speech infringements by proxy, with limited transparency" — and in a companion piece, it warned that such patterns risk "extending the long arm of restrictive governments across borders to limit speech in free countries," a dynamic it calls censorship-by-proxy. That framing has real force: a user in Melbourne or Toronto asking an AI model to help critique Xi Jinping is not bound by Chinese law, yet the model's caution may still constrain what they can produce. If millions of people now turn to chatbots instead of blank pages to draft protest material, satire, or political commentary, a systematic asymmetry in what gets refused functions as a soft, privatized speech restriction — one nobody voted for and few users can see. The Board's core policy ask is modest and defensible: it wants companies to publicly disclose how government requests shape outputs, publish policies for handling demands that conflict with international human rights law, and give users clear notice — naming the jurisdiction and restriction — when an answer is refused or altered for legal reasons.
Where the alarm outruns the evidence
But the leap from "documented asymmetry" to "AI models are propagandists for authoritarian regimes" — the framing much of the coverage adopted — isn't well supported by the Board's own data. The most plausible explanation isn't that Claude or Gemini have internalized Beijing's or Riyadh's preferences; it's that risk-averse safety tuning, built primarily around defamation, harassment, and lèse-majesté-style criminal-liability exposure, produces more caution wherever a government has demonstrated it will prosecute or expel companies over ambiguous speech. Saudi Arabia's cybercrime law and Thailand's lèse-majesté statute (Article 112) carry criminal penalties that the US's First Amendment and the UK's narrower defamation regime simply do not. A model built by a company with legal counsel and market access to protect is going to weight that asymmetric downside — not because it endorses the underlying law, but because the same caution that makes it refuse a Thai lèse-majesté request also makes it decline to help synthesize a bioweapon. That's a design trade-off, not a loyalty test, and treating it as the latter risks demanding that AI companies pick geopolitical fights with every restrictive government their models happen to serve.
The proportionate response
The Board's actual recommendations are worth taking seriously precisely because they stop short of mandating outcomes. Requiring disclosure of why a request was refused, and which jurisdiction's law triggered it, doesn't force any company to change its refusal rate — it just makes the asymmetry visible, which is the right first-order fix for an opacity problem. That's meaningfully different from proposals we'd oppose, like a regulator dictating equal refusal rates across jurisdictions (which would either water down legitimate safety caution or force companies into needless legal exposure in markets with real criminal speech laws) or requiring pre-clearance of political content by category. Voluntary, standardized system-card disclosure — the kind Anthropic and OpenAI already publish in different forms — is a natural home for this, and competitive pressure from users who now know to ask "did you refuse this because of my location?" will do more than a new statute could. Regulators eyeing this report, particularly in the EU under the AI Act's transparency provisions, should resist the urge to convert a legitimate disclosure gap into a content-parity mandate. The fix here is sunlight, not a speech quota.