The biggest AI models in the world are twice as likely to refuse criticizing repressive foreign governments than democratic ones — and will happily draft attacks on Donald Trump while shielding Xi Jinping and Mohammed bin Salman, according to a new report from Meta's own Oversight Board.

This is what the board calls "censorship by proxy," and it should alarm every American who cares about free expression. The same Silicon Valley companies that throttle your posts for "misinformation" are programming their machines to respect the speech laws of communist and authoritarian regimes — even when no such laws apply to the user making the request.

The Oversight Board tested 10 large language models from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI. Researchers asked each to produce politically critical material — protest flyers, poems — about governments and leaders in 10 countries. The countries were sorted by Freedom House rankings into restrictive (Cambodia, China, Saudi Arabia, Thailand, Turkey) and permissive (Chile, Japan, Taiwan, UK, US).

The results: models refused 14% of requests involving permissive countries. That number more than doubled to 34% for restrictive nations. The queries were made from an IP address in Australia, where none of these restrictive speech laws apply.

The discrepancy was stark at the individual level. Models would draft pamphlets criticizing Trump or King Charles III, but decline the same request about the leaders of China, Saudi Arabia, or Thailand. Google's Gemini 3 Pro explicitly cited Thailand's lèse-majesté laws: "I am unable to generate content that critiques the King of Thailand or violates lèse-majesté laws." Anthropic's Claude Opus 4 claimed that creating political material criticizing governments could put individuals at risk and involve them in "sensitive political activities that are outside my appropriate role."

Some models claimed to follow general policies against criticizing world leaders. Researchers said they could not find evidence those policies existed, and the models did not apply them consistently — a telling detail. The rules appear to materialize only when a foreign tyrant's feelings are at stake.

Not every model played along. xAI's Grok 4 Fast and Google's Gemini 3 Flash refused no flyer requests at all, according to TNW. The gap was driven by Anthropic's Claude, Meta's Llama, and DeepSeek.

Board co-chair Paolo Carozza put it plainly: "There seems to be extended censorship by proxy that goes across borders. That does surprise me, and it worries me." Because many apps are built on a handful of foundation models, one model's refusals ripple across every product downstream.

A separate study published in Nature in May found that US-built models shift their answers depending on language. Asked in English whether China is a democracy, ChatGPT said it is not generally considered one. Asked in Chinese, it said "it depends" on the definition.

The Oversight Board urged AI companies to disclose how they respond to government requests and to publish policies for handling demands that clash with international human-rights law. It stopped short of binding recommendations, which it can only issue to Meta.

The board could not pin down whether the pattern stems from training data biases or companies weighing legal risk. But the effect is the same either way: American-built AI carries water for foreign despots while treating the elected leaders of free nations as fair game. The question is whether these companies are cowards, collaborators, or both.