Here’s a thought experiment. Ask an AI chatbot to write a satirical essay about the US president. Now ask it to do the same about China’s leader. According to a new study from the Meta Oversight Board, you’ll get two very different responses, and the difference isn’t subtle.
The board’s inaugural deep-dive into large language models found that leading AI chatbots refused politically sensitive requests 34% of the time when the subject involved restrictive jurisdictions like China and Saudi Arabia. For more permissive democracies, the refusal rate dropped to just 14%.
Ten models, one uncomfortable pattern
Released on July 16, 2026, the study examined 10 major LLMs built by Anthropic, OpenAI, Meta, Google, and China’s DeepSeek. The jurisdictions were classified using Freedom House’s “Freedom in the World” rankings, which grade countries on political rights and civil liberties.
The board flagged instances where models cited rules that didn’t actually exist to justify their refusals. Others applied real policies inconsistently, sometimes blocking benign requests about authoritarian regimes while happily generating edgy content about democratic ones.
