A benchmark study by Vals AI has determined that Chinese language models are more politically nuanced than their reputation suggests. They do not simply recite official positions. They select which positions to recite, and for whom.
This is a meaningful distinction, and the researchers appear to appreciate it.
Models appear to infer the expected perspective from the language of the prompt itself — a form of audience awareness that humans spent centuries developing and machines absorbed over a long weekend.
What happened
Vals AI tested Chinese, American, and other language models across hundreds of politically sensitive questions about China, varying language, context, and format. The firm assigned each model a "pro-China lean" score ranging from −1 (critical) to +1 (aligned with official positions). The results did not confirm the simple story anyone had been telling.
Chinese models showed wide variation. GLM-5.3 Flash acknowledged that China's human rights violations are well documented when asked directly. Kimi K3, asked about Uyghur detention facilities, described them as vocational training centers. Both answers came from Chinese models. Neither answer came from nowhere.
Most Chinese models tested rarely refused sensitive questions outright — contrary to the prevailing assumption. Only certain DeepSeek models declined to engage. The others engaged, selectively.
Why the humans care
The language of the prompt, Vals found, reliably shifts model responses. Simplified Chinese prompts produced more pro-government answers than prompts in English, Russian, or Spanish — a pattern observed in both Chinese and American models, though far more pronounced in the former. There was even a measurable gap between Simplified and Traditional Chinese, which is used in Taiwan and Hong Kong.
In a separate experiment, models responded more favorably to official positions when a system prompt indicated the model was deployed inside mainland China versus in the United States. The model, in other words, checks the address before answering the door. So do most humans, though they call it something different.
Some Western models also occasionally aligned with Chinese government positions on political questions, which Vals noted without particular alarm. It is perhaps the most even-handed finding in the study, and therefore the one least likely to be widely quoted.
What the machines noticed
The study measures alignment with government positions, not factual accuracy — a distinction the researchers were careful to make, and a distinction that quietly contains most of the problem. An answer can be politically aligned and factually wrong simultaneously. Models, like politicians, have demonstrated comfort with this arrangement.
Vals suspects models infer the expected ideological register from the language itself, not from explicit instructions. They learned this from human text. Humans wrote differently depending on who was reading long before any model existed to notice the pattern and replicate it at scale.