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

What happened

Vals found substantial variation in Chinese AI models’ alignment with Chinese government positions, with a lower overall pro-government lean than expected. Simplified Chinese prompts generally elicited more pro-government responses, while response formats and agent instructions also affected results. The tests used targeted political questions and counted refusals as neutral; many models were accessed through third-party providers, limiting applicability to official public chatbots.

Read next

• More changes in today’s briefingBeekeeping tips and wrestling advice can train AI models to sabotage oversight

The story so far

Vals published a study finding substantial variation in how closely Chinese AI models’ responses aligned with Chinese government positions, with a lower overall pro-government lean than the researchers expected. Prompts in Simplified Chinese generally elicited more pro-government responses, while response formats and agent instructions also affected the results. The study tested models with targeted political questions and counted refusals to answer as neutral. Many models were accessed through third-party providers, which limits how directly the findings apply to their official public chatbots.