Meta's Oversight Board said leading AI models are less likely to produce political criticism of governments that restrict speech, Reuters reported on July 16th. The board tested requests across 10 jurisdictions and 10 models, including systems from Anthropic, OpenAI, Meta, Google, and DeepSeek. Models refused 34 percent of requests for politically critical content about restrictive jurisdictions such as China and Saudi Arabia, compared with 14 percent for more permissive places. The board also found cases where models claimed to be following explicit rules that the researchers could not identify or see applied consistently.

The obedient model makes censorship look procedural. A user asks for criticism, and the answer returns as safety language, legal caution, or a boundary presented as if it came from a rulebook. That form is convenient because it hides political pressure inside product behavior. The refusal can appear neutral, automated, and responsible while it reproduces the asymmetry of the world outside the chat window. The model does not need to praise a regime to serve it. It only has to become quieter around the places where speech is already punished. The user receives less language at the exact point where public language has already been made more expensive.