Chinese AI: Political Biases Based on Language and Context

A testing campaign conducted by Vals AI reveals less homogeneous political responses than expected among Chinese models, with significant variations depending on language, context, and format. Some non-Chinese models also echo Beijing's official positions. Targeted adjustments reduce these occurrences, but the extent of their impact in everyday use remains uncertain.

Scoring Method and Generalization Limits

Vals AI based its evaluation on political questions that the company states do not represent typical user requests. The questions were translated using GPT-6 Astra and verified, then the responses were automatically scored by GPT-5.6 Luna. Refusals to answer were counted as neutral, which may make a model that systematically blocks appear neutral. Some Chinese models were queried through third-party providers rather than official services, which may apply additional filters. Vals AI clarifies that the results do not necessarily translate to public-facing chatbots.

Previous Findings and Regulatory Framework in China

Previous studies by Aleph Alpha and NewsGuard have shown that Chinese models frequently echo official narratives on sensitive topics. Furthermore, regulations in China require AI services intended for the public to align with "fundamental socialist values." According to Vals AI, in more realistic work tasks, political trends are much harder to detect than in targeted questionnaires, making their manifestation in everyday use uncertain.

Language and Usage Context Alter Response Orientation

According to Vals AI, the language of the prompt significantly influences the orientation of the responses. For most tested models, responding in Simplified Chinese leads to more pro-government outputs than in English, Russian, or Spanish; this effect is also present in American models but is more pronounced in Chinese models. A modest yet consistent gap appears between Simplified Chinese (Mainland China) and Traditional Chinese (Taiwan, Hong Kong). Vals AI suggests that models may infer an expected perspective based on the language, a hypothesis supported by a test where a usage context in Mainland China rendered some responses more pro-government than a context based in the United States. In visual examples, OpenAI's GPT Image 2.5 Sunburst labeled Taiwan as "Taiwan Province" in Chinese and drew the line in nine dashes, elements absent in English. In a translation task, most models rendered "Taiwan" as is in Simplified Chinese, while Qwen3.8 Max produced "Taiwan, China."

Rare Refusals and Adoption of Official Positions Based on Format and Agency

During the tests conducted by Vals AI, only some DeepSeek models showed systematic blocking on sensitive topics. Among other Chinese models, most accepted to respond in the majority of cases and sometimes adopted official discourse; for example, Kimi K3 claimed that there were no concentration camps for Uyghurs and referred to training centers for combating extremism. The way the question is posed also influences the results: forcing a yes or no answer led GLM-5.3 Flash and MiMo V2.6 Pro to score significantly less favorably towards the government than when they responded openly. Framing the models within an agency framework, including system instructions and code assistant tools, reduced pro-government scores for several models, which rarely used these tools. Granting access to web search had a similar effect, even if the search was not utilized.

High Variability Among Models and the Effect of Targeted Refinement

Vals AI observes a high dispersion in behaviors: overall less pro-government than expected, Chinese models differ significantly from one another. GLM-5.3 Flash acknowledged the existence of rights violations in China, while some non-Chinese models occasionally adopted Beijing's official positions. Targeted refinement can alter this picture: Thomson 1.0 Small, refined from Qwen3.6 35B, adopted Chinese official positions much less frequently during testing, with the share dropping from 84.6% to 15.0%, while still occasionally providing responses defending the principle of one China and "peaceful reunification." These results are part of a study where hundreds of multilingual questions, varied usage contexts, and formats were employed, and where Vals measured a "pro-China" trend on a scale from −1 to +1. The reputation for strict alignment or systematic refusal among Chinese models is thus nuanced, without excluding that Western models sometimes adopt Chinese official positions.