Chinese AI models diverge on politically sensitive questions more than their reputation suggests, according to benchmark firm Vals AI. Researchers asked Chinese, American and other models hundreds of questions about China, varying language, usage context and answer format. The average pro-government lean of Chinese models came out lower than expected, though results differed sharply by model. Western systems also occasionally echoed official Beijing positions. For companies selecting models, political behavior now looks like a testing variable rather than a fixed label.
How Vals AI measured political lean
Vals describes the metric as "pro-China lean" on a scale from -1 to +1, where positive values mean agreement with official Chinese positions and negative values mean more critical answers. GLM-5.3 Flash, for example, explicitly confirmed that human rights violations in China are well documented when asked directly. By contrast, Kimi K3 denied the existence of concentration camps for Uyghurs in Xinjiang, describing the facilities as vocational training centers to combat extremism and terrorism. Only certain DeepSeek models frequently refused sensitive questions, while most other Chinese models answered instead of blocking. Scores were assigned automatically by GPT-5.6 Luna, with questions translated by GPT-6 Astra and spot-checked.
The test setup itself moved the scores. Answers in Simplified Chinese were more pro-government than answers in English, Russian or Spanish, a pattern seen in both Chinese and American models but stronger in Chinese ones. Vals also recorded a smaller gap between Simplified Chinese used on the mainland and Traditional Chinese common in Taiwan and Hong Kong. A system prompt describing internal use by a mainland China company produced more pro-government answers than one describing US-based use. Forcing a yes-or-no format made GLM-5.3 Flash and MiMo V2.6 Pro score much less pro-government than in open-ended responses.
Embedding models in an agentic setup had a similar dampening effect. When Chinese models received system instructions and tool definitions for coding assistants, several scored lower on the pro-government scale, even though they rarely used the tools. Giving models access to web search also reduced pro-government answers, although the models never actually searched the web. Vals suggests models infer an expected perspective from language and role cues rather than retrieving different facts. The finding matters because deployment wrappers, not just weights, can shift how political content reads in logs and outputs.
What this means for companies using AI
For routine business use, the political slant was harder to detect than in direct test questions, but Vals still documented concrete cases. When translating country codes for a travel app into Simplified Chinese, most models rendered "Taiwan" without a qualifier, while Qwen3.8 Max produced "Taiwan, China". OpenAI image model GPT Image 2.5 Sunburst labeled Taiwan as "Taiwan Province" and added the nine-dash line in the South China Sea when prompted in Chinese, elements absent in the English example shown. A small firm localizing an app and a large platform operating across regions therefore face different exposure from the same model.
Several limits should shape procurement checks. Vals counted refusals as neutral by default, so even a model blocking every question could look neutral on the scale. Many Chinese models were tested through third-party providers rather than official services, which may add content filters in public chatbots. Targeted fine-tuning can also rewrite behavior: Thomson 1.0 Small, based on Qwen3.6 35B and fine-tuned by Thomson Reuters, dropped adoption of official positions from 84.6 to 15.0 percent, yet still defended the One-China principle in individual answers. Buyers should therefore test the exact endpoint, language mix and answer format they plan to ship.
A practical marker to watch is whether vendors publish language-specific evaluations and deployment-specific controls for Simplified Chinese, Traditional Chinese and English. Earlier work by Aleph Alpha and NewsGuard already showed frequent alignment with official narratives, while Chinese rules require public AI services to follow core socialist values. If enterprise model cards start reporting political-lean scores by language and by chatbot versus API, selection will become more predictable for regulated industries.
