Many MindedBenchmark firm finds Chinese AI models lean less toward Beijing than expected
Vals AI, a firm that benchmarks AI models, tested Chinese, American and other models on hundreds of questions about China. It varied the prompt language, the stated usage context and the response format. Answers were placed on a scale from -1 (critical) to +1 (agreeing with official Chinese government positions). Vals calls this a "pro-China lean." The scale measures alignment with certain government positions, not factual accuracy.
Overall, the Chinese models leaned less toward the government than Vals expected, and individual models differed widely. GLM-5.3 Flash said China's human rights violations are well documented. Kimi K3 denied that concentration camps for Uyghurs exist in Xinjiang and called them vocational training centers. Some non-Chinese models occasionally matched official positions. For most models, Simplified Chinese prompts produced more pro-government answers than English, Russian or Spanish ones, in US models as well, though less strongly. Forcing yes/no answers, adding coding-assistant instructions, or offering web search lowered scores for several models. Thomson Reuters' fine-tuned Thomson 1.0 Small took official positions in 15.0 percent of cases, against 84.6 percent for its base model, Qwen3.6 35B.
This is friction for knowledge. The same question can get a different answer depending on the language it is asked in and how the model is deployed. People asking in Simplified Chinese may get a more government-aligned picture than people asking in English. That uneven reach cuts against knowledge being equally available to everyone. The Thomson result also suggests fine-tuning can change a model's lean substantially.
The caveats are large. Vals says its questions are not typical user prompts. Refusals count as neutral, so a model that refuses everything could look neutral. Many Chinese models were tested through third-party providers, which may add filters, so results may not carry over to public chatbots. Scoring was done automatically by another language model. Political tendencies were much harder to detect in realistic work tasks, so real-world relevance is still open.
Source: The Decoder