Chinese AI models are politically complex, not just government mouthpieces, study finds

Why it matters

A benchmarking study by Vals AI challenges the assumption that Chinese AI models uniformly parrot Beijing's line or refuse sensitive questions. The researchers posed political questions to Chinese, American, and other models, varying prompt language, usage context, and response format. They found the pro-government lean of Chinese models was lower than expected, with wide variation between models, while some Western models occasionally echoed official Chinese positions. Prompt language mattered: Simplified Chinese produced more pro-government answers than English, Russian, or Spanish, and even differed from Traditional Chinese. Most Chinese models answered rather than refused, sometimes adopting official positions, and forced yes-or-no formats reduced pro-government scoring for GLM-5.3 Flash and another model. Agentic setups and web search access also shifted responses. The study has limits: refusals counted as neutral, test questions were unrepresentative, and many models were accessed through third-party providers, so official consumer chatbots may behave differently. The findings matter for anyone assessing political bias in widely used AI systems.