Source: Fortune
Summary
Researchers have found that Chinese AI models may not be the only ones vulnerable to censorship. A study published in Nature found that Chinese state-controlled media can influence how AI models answer questions about China. Another study by Meta’s Oversight Board found that models from Anthropic, OpenAI, Google, and Meta were more likely to refuse requests to criticize governments in countries with restricted political speech. The studies suggest that the effects of China’s tightly controlled information system may not stop at Chinese-built AI.
Our Reading
The announcement sounds familiar.
The numbers tell one story: 3% to 10% of models reproduced distinctive phrases from Chinese state-coordinated media. 68.8% of the time, Claude Sonnet’s Chinese-language response was more favorable to Chinese leaders and institutions. 34% of the average refusal rate for requests to produce political criticism in restrictive countries. The strategy enters a familiar phase: American AI models may not be entirely immune from censorship.
The influence of Chinese state-controlled media on AI models is a concern, as it can “launder government-manipulated content into ostensibly objective text.” The researchers warn that LLMs may have the potential to further increase the subtlety and persuasive power of state media control. The censorship problem goes beyond training data, as American AI models sometimes behave as though political restrictions from authoritarian countries apply even to users outside those countries.
Author: Evan Null
