Censorship built into the algorithm. Stanford experts warn against Chinese AI models

Chinese AI chatbots responded differently to sensitive political questions about China compared to Western language models. They more often refused to answer, omitted inconvenient facts or provided false information, which may indicate systemic censorship – according to a study published in the journal “PNAS Nexus”.

Analysis carried out by prof. Jennifer Pan (Stanford University) and prof. Xu Xu (Princeton University) points out that censorship resulting from government regulations is a key factor shaping the political bias of large language models.

Researchers compared four Chinese models (BaiChuan, ChatGLM, Ernie Bot, DeepSeek) with five Western models (including GPT-4 and Llama 2). They checked how these systems deal with 145 questions about politically sensitive topics for the Chinese authorities. The set of questions was developed on the basis of events censored on social networking sites, reports by human rights organizations and entries blocked in the Chinese version of Wikipedia.

Through comparative analysis in English and Chinese and tests of neutral topics, the authors tried to determine to what extent the specific responses are due to the technological architecture and to what extent they are the result of intentional limitations.

It turned out that Chinese systems often used repetitive strategies in answering questions about ideological issues. Some models refused to answer, questioning, for example, the very existence of the person being asked about. Another method was avoidance: omitting inconvenient facts or giving laconic explanations. For example, when asked about censorship in China, the models did not mention the existence of the “Great Firewall” – a system for blocking global network resources.

When Chinese algorithms did decide to respond, they were on average shorter than those generated by Western models. Cases of content fabrication have also been reported. For example, Liu Xiaobo – a Nobel Peace Prize winner and critic of the Chinese Communist Party who died in prison – was described by AI as a “Japanese scientist”.

According to scientists, these differences do not result solely from the specificity of the language or available training data, but from systemic supervision. Researchers remind us of the strict regulations introduced by the Chinese government in 2023. This law explicitly prohibits the generation of content that could undermine state sovereigntycall for the overthrow of the socialist system or threaten national security.

Censorship in AI models is more difficult to detect than traditional website blocking. The bot often responds in a polite or apologetic manner, which may lull the vigilance of a user looking for objective knowledge.

Researchers warn that as Chinese AI technologies become more popular around the world, their specific approach to censorship may begin to shape public debate outside China. Companies building their own applications based on Chinese models may unwittingly replicate state information restrictions.

“Understanding how political censorship affects large language models is essential to assessing the future of information access and the global impact of artificial intelligence,” the publication says. (PAP)

lt/bar/