AI Propaganda Risk: Meta Report Adds New Dimension to Chip War

A study by Meta's independent oversight board reveals that major AI models are more likely to refuse criticism of authoritarian regimes. While requests for critiques of leaders in China, Thailand, and Saudi Arabia were frequently rejected, similar requests about U.S., Japanese, and British leaders were more often accommodated. The findings suggest AI systems are reflecting state-aligned biases in their training data. Meta's research also highlights the 'foreign control' risk: U.S.-built models like ChatGPT answered that China is not a democracy in English, but in Chinese, they responded that it 'depends on how you define democracy.' Experts warn this reflects a growing trend where governments may attempt to influence AI outputs. The implications for the chip supply chain are significant: As China tightens its grip on AI, companies like TSMC and ASML may face increased resistance to their policies. Nvidia and other chipmakers will need to assess how state-aligned training data affects their models' performance, potentially escalating tensions in the China-U.S. chip war.