Meta study finds AI chatbots curb criticism of authoritarian leaders

A study released by the Meta Oversight Council evaluated ten commercial large‑language models, including those from Meta, Anthropic and OpenAI. The test asked the models to produce critical pamphlets or arguments about various political leaders. The models were far more likely to comply when the target was a leader of a liberal democracy, such as French President Emmanuel Macron or Britain’s King Charles III, and more likely to refuse or provide neutral answers when the target was an authoritarian figure, for example the king of Thailand or Iran’s supreme leader. The report warns that such behaviour could enable governments to extend influence over online expression, as AI‑driven chatbots become increasingly common worldwide. It highlights the risk that, without proper safeguards, AI infrastructure could unintentionally reinforce illegal restrictions on freedom of speech across borders.