How pro-China are Chinese models? (www.vals.ai)

Recent research has explored the degree of political bias in Chinese AI models, challenging the notion that they uniformly endorse pro-Chinese Communist Party (CCP) narratives. Various models, including GLM-5.3 Flash and DeepSeek V4.1, were tested with politically sensitive prompts in multiple languages. The investigation revealed that while some Chinese models occasionally critique the government, they tend to display a more favorable stance towards the CCP when asked in Simplified Chinese. Interestingly, responses varied significantly based on the language and context, suggesting that models dynamically adjust their outputs based on user prompts—leaning towards a pro-CCP narrative in Chinese and toward neutral or critical responses in English. These findings are significant for the AI/ML community as they highlight the complex interplay between language, model training, and the socio-political framework in which these models operate. The research underscores the importance of transparency in AI model outputs, particularly for companies and governments assessing the integrity and biases of AI systems. It also sets a precedent for ongoing discussions on how deployment context and interaction types influence the alignment of AI models with political narratives, prompting a reevaluation of how these systems are utilized in politically sensitive environments.