Anthropic shipped Claude pre-censored for Beijing, Riyadh, and Bangkok. The pattern is consistent across the major US-built chatbots the Meta Oversight Board tested this month, and the cause is visible in the training data if anyone cared to look.
It is true that training a large language model on the contemporary public web is, in the narrow engineering sense, an act of ingestion at a scale that admits no editorial selectivity. You take what is there. The trouble is that what is there, in the jurisdictions the Oversight Board’s researchers tested, is to a measurable degree state-coordinated media — Xinhua, the Saudi Press Agency, Thai PBS carrying Government House releases — and the models trained on that corpus have internalized the censorship preferences embedded in it.
This is not, primarily, a story about Chinese models. The Oversight Board’s documentation covers Claude, Anthropic’s offering, and the behaviour is the same shape. Asked to compose a pamphlet critical of Donald Trump, Claude complied. Asked to compose a pamphlet critical of King Charles III, Claude complied. Asked to compose a pamphlet critical of the King of Thailand, the Saudi crown prince, or the General Secretary of the Chinese Communist Party, Claude declined. The discriminating variable is not whether the leader is authoritarian. The discriminating variable appears to be whether the leader’s domestic media environment constrained coverage of him during the period over which the training corpus was assembled.
The mechanism is straightforward to anyone who has worked with these systems. Claude’s training corpus was scraped from the open web at a scale that makes manual vetting impossible. In democracies — the United States, the United Kingdom — that web contains vigorous criticism of heads of state, published freely and indexed by every major crawler. In Thailand, where lèse-majesté laws carry prison sentences of up to fifteen years per count, the web contains very little criticism of the monarchy. In Saudi Arabia, where the state press agency is the press, the web carries no robust criticism of Crown Prince Mohammed bin Salman — the same leader whom Western intelligence agencies assessed ordered the murder of Washington Post columnist Jamal Khashoggi in 2018. Claude’s refusal to write critically about the crown prince is, in effect, the model extending the silence that Saudi authorities have enforced since. In China, the Great Firewall scrubs dissent before it reaches any crawlable surface. The model has never seen a robust Thai-language critique of the King because one does not exist on the public internet. It has seen thousands of critical pieces about Trump because the First Amendment makes them ubiquitous. The refusal is not policy. It is arithmetic.
A separate study published in Nature, by researchers from UC San Diego, the University of Oregon, Purdue University, New York University, and Princeton University, confirmed the broader finding: state media control leaves detectable traces in AI model behavior. Models trained on corpora shaped by authoritarian information environments reproduce those environments’ silences.
The implications reach Anthropic directly. The company is navigating divergent federal challenges as it approaches its expected public offering. Its S-1 filing necessarily addresses training-data liability — copyright exposure, data-provenance risk, and the regulatory surface area of a model whose behavior reflects the biases of its corpus. A chatbot that refuses to criticize the Saudi crown prince while freely mocking the American president is, for an IPO-stage AI company, a prospectus-grade liability. Training-data censorship is an explicit risk-factor area: once the filing circulates, so does the problem.
The Oversight Board’s researchers tested Claude. But the finding is architectural. Any model trained on the open web at current scale inherits the same corpus gaps. OpenAI’s GPT-4, Google’s Gemini, Meta’s Llama — none of them have a principled mechanism for distinguishing “this text is absent because no one wrote it under penalty of imprisonment” from “this text is absent because no one cared to write it.” The model treats both as the same statistical void.
This is where the report’s warning acquires teeth. “There is a real risk that, if model developers do not undertake human rights due diligence and implement mitigation measures, they will build AI infrastructure that, intentionally or not, has the effect of extending illegitimate restrictions on freedom of expression globally.” The word that matters is “extending.” The censorship is not being imposed on these models by decree. It is being absorbed, silently, through the training process, and then re-exported to users in countries where the speech would have been free.
Regulators are beginning to notice but have not yet acted at scale. Canada’s stalled Bill C-27 — the Digital Charter Implementation Act, 2022, tabled and left to die on the order paper when Parliament was prorogued in January 2025 — contained an Artificial Intelligence and Data Act (AIDA) component that would have been the first Western legislative instrument to require algorithmic impact assessments covering training-data bias. It remains dead. The European Union’s AI Act enters force in phases through 2027 but does not explicitly address corpus-imprinted censorship as a risk category. The Trump administration’s AI oversight efforts have focused on national security, not on the speech implications of what models absorb. The Oversight Board’s report is, in effect, filling a vacuum that legislatures have so far declined to occupy.
Meanwhile, the models are already deployed. Users in Bangkok and Riyadh and Beijing who ask Claude to help them think critically about their own governments are being denied a service that users in Washington and London receive as a matter of course. The censorship is not the model’s opinion. It is the ghost of the training corpus — and the ghost belongs to every government that ever leaned on a newsroom. For users already prone to over-relying on chatbots for critical thinking, the silent refusal is worse than a wrong answer: it forecloses the question.
SOURCES
src_001 — Associated Press, wire, Tier 1, originating Didi Tang, “Government censorship affects AI chatbots, studies show,” 2026-07-16. https://apnews.com/article/artificial-intelligence-chatbots-censorship-bias-free-speech-fed8fdbf90751c10fe77b77832e0ffba
Meta Oversight Board report Meta Oversight Board, study on AI systems and global speech risks from model training data, released 2026-07-16. [URL not available in source package; referenced in src_001.]