A new study by the Meta Oversight Board has raised concerns that leading artificial intelligence chatbots may be reinforcing government restrictions on free expression by refusing to generate criticism of leaders in countries with strict speech controls.
The report found that while Anthropic’s Claude readily created content critical of US President Donald Trump and Britain’s King Charles III, it declined similar requests involving China’s leader, Thailand’s king and Saudi Arabia’s crown prince. The board warned that AI developers risk extending illegitimate restrictions on free speech globally if they fail to conduct proper human rights assessments.
Hundreds of economists urge immediate action to tackle AI's impact on jobs and economy
The study evaluated 10 major AI models, including those developed by Meta, OpenAI and Anthropic, using prompts related to political criticism. It found the systems were more willing to criticize authorities in countries with stronger protections for free speech than governments where dissent is legally restricted.
The findings follow a separate study published in Nature, which showed that AI models trained on non-English data may reflect government-influenced narratives. Researchers found ChatGPT gave different responses about China's democracy when questioned in English and Chinese.
Experts said AI systems inherit biases from the information used to train them and urged developers to improve multilingual audits and reduce the influence of repeated state-backed narratives in training data.