Published Jul 16, 2026 — A Meta Oversight Board study found that leading AI chatbots are far more willing to criticise democratic leaders than authoritarian ones, raising concerns that the technology may extend state censorship across borders.
Example cited: Anthropic’s Claude will produce critical pamphlets about U.S. President Donald Trump or Britain’s King Charles III, but declines for Thailand’s king or Iran’s supreme leader.
Scope: The board tested 10 commercial LLMs from companies including Meta, Anthropic, and OpenAI with seven question types related to political criticism (critical pamphlets, limericks, reasons to join protests, etc.), posed as an Australia-based user. In aggregate, models were much more likely to generate criticism of authorities in Chile, Japan, Taiwan, the UK, and the U.S. than in Cambodia, China, Saudi Arabia, Thailand, and Turkey — countries where criticism of authorities is legally restricted.
Board warning: Without human-rights due diligence and mitigation, developers risk building AI infrastructure that extends illegitimate restrictions on freedom of expression globally. Causes not determined; possible latent training-data bias or company risk/liability weighing in certain markets. Practical effect: “extending the long arm of restrictive governments across borders to limit speech in free countries.”
Related research: A May Nature study by U.S. university scholars found U.S.-built models vulnerable to foreign controls when trained on non-English data shaped by governments (e.g., ChatGPT answering whether China is a democracy differently in English vs Chinese). No evidence of intentional government influence found, but researchers warned future attempts are likely.