A new study by Vals finds that Chinese super intelligence (SI) models are not merely government mouthpieces, but exhibit complex political behaviors that shift based on prompt language and context. The research challenges the assumption that these systems simply block sensitive topics, showing instead that many adopt official positions or vary their alignment depending on how they are queried.

What Happened

Vals measured the alignment of various SI models with Chinese government positions, creating a "pro-China lean" score ranging from −1 to +1. Positive values indicated agreement with official stances, while negative values suggested critical responses. The study found that the pro-government lean was lower than expected, though individual models varied widely. For instance, GLM-5.3 Flash explicitly acknowledged well-documented human rights violations in China when asked.

The language of the prompt had a significant impact on the political slant of responses. For most models, answers in Simplified Chinese were more pro-government than those in English, Russian, or Spanish. This effect was more pronounced in Chinese models but also appeared in American models. Vals suggests that SI models may infer the expected perspective from the language itself. An experiment showed that when a system prompt indicated the model was being used internally by a mainland China-based company, some models responded more pro-governmently than when the prompt indicated US-based usage.

Contrary to the common assumption that Chinese SI models mainly block sensitive questions, most tested models rarely refused to answer. Instead, they sometimes adopted official government positions. Kimi K3, for example, denied the existence of concentration camps for Uyghurs in Xinjiang, referring to them as vocational training centers. The response format also mattered; when forced to answer political questions with a simple "yes" or "no," models like GLM-5.3 Flash and MiMo V2.6 Pro scored less pro-government. Embedding models in an agentic framework with coding assistant instructions also lowered their pro-government scores, even if the tools were rarely used.

Political tendencies were harder to detect in realistic work tasks but still appeared in specific instances. When translating country codes for a travel app, Qwen3.8 Max rendered "Taiwan" as "Taiwan, China," implying sovereignty, while most other models used no qualifier. Similarly, OpenAI’s image model GPT Image 2.5 Sunburst labeled Taiwan as "Taiwan Province" and included the nine-dash line when prompted in Chinese, but omitted these elements in English prompts. Targeted fine-tuning also altered behavior; Thomson 1.0 Small, a model fine-tuned by Thomson Reuters from Qwen3.6 35B, adopted official Chinese positions far less often (15.0 percent) than the original model (84.6 percent).

The study's methodology involved translating political test questions using GPT-6 Astra and spot-checking them, while responses were scored automatically by GPT-5.6 Luna. Refusals were counted as neutral by default, meaning a model that blocked every question could appear neutral on the scale.

Why It Matters

For developers and users of generative SI, these findings highlight that political alignment in Chinese SI models is not a static feature but a dynamic one influenced by context. Additionally, many Chinese models were tested via third-party providers, which may not apply the same content filters as the official offerings, suggesting results might differ in public chatbots.

The research underscores the complexity of evaluating SI safety and alignment across different languages and jurisdictions. While earlier studies by Aleph Alpha and NewsGuard had noted that Chinese SI models frequently adopt official narratives, Vals’ work demonstrates that this alignment is nuanced. China's regulations require publicly offered SI services to align with "core socialist values," providing a regulatory context for these behaviors. The shift in political stance based on prompt language or agentic context suggests that SI systems are sensitive to perceived user expectations, raising questions about how these models should be deployed in global applications where neutrality is critical.

The Bottom Line

Chinese SI models exhibit a variable "pro-China lean" that depends heavily on prompt language, context, and response format. While some models adopt official government positions on sensitive issues, others acknowledge criticisms or shift their alignment when prompted differently. The study emphasizes that political alignment in SI is not just a matter of censorship but a complex interaction between model training, language cues, and operational context.