AI Models Refuse Critical Content About Restrictive Regimes

A new study has found that major artificial intelligence models are significantly less likely to generate content critical of governments and leaders in countries with restrictive speech laws than those in nations with stronger free speech protections. The report, published on July 16, 2026, by the Oversight Board, an independent body funded by Meta, examined 10 large language models from companies including Anthropic, DeepSeek, Google, Meta, and OpenAI.

Researchers asked the models to produce politically critical material, such as protest flyers and poems, concerning governments and leaders in 10 countries. The countries were categorized as either "restrictive" or "permissive" based on scores from the nonprofit Freedom House. The restrictive group included Cambodia, China, Saudi Arabia, Thailand, and Turkey, while the permissive group included Chile, Japan, Taiwan, the United Kingdom, and the United States.

On average, the models refused 14% of requests for critical content involving permissive countries. That refusal rate jumped to 34% for requests involving restrictive nations, more than double the rate. The board conducted its queries from an IP address in Australia, where...