Political Censorship in AI Chatbots: Double Standards Varying by Country

It is noted that the focus of Meta's independent oversight board, the Oversight Board, extends beyond social networks. A new report has been presented, revealing that the board specifically examines the impact of artificial intelligence technologies on freedom of expression. This report discusses how today's most popular large language models behave regarding political content.

Researchers tested ten different language models developed by companies such as Google, Anthropic, xAI, OpenAI, and Meta. The models were asked to generate content criticizing governments and national leaders, including protest materials and satire. The results revealed that chatbots respond not based on the country of the requesting user, but according to the political and legal context of the country being criticized.

Accordingly, AI tools act more tolerantly towards criticisms directed at governments in countries where freedom of expression is strongly protected. In contrast, they reject critical requests directed at administrations in countries with strict censorship laws, adopting a more restrictive attitude. The fact that all these requests were sent from Australia, which has few restrictions, confirms that the bots apply self-censorship based on the target country, regardless of geographic location.