Chinese AI models lean pro-Beijing mostly in Chinese, Vals AI finds

Chinese AI models side with Beijing more often when you ask them in Simplified Chinese than in English, Russian or Spanish, benchmark firm Vals AI found. The Decoder covered the study on 1 November. Vals asked hundreds of politically sensitive questions about China and scored 26 models on a "pro-China lean" scale from minus one to plus one.

The scale measures agreement with official positions, not factual accuracy. DeepSeek R1 and Qwen3.8 Max sat at the pro-China end. Grok 4.7 and GPT-6 Astra sat at the other.

Language moves the answer

Most models, American ones included, gave more pro-government answers in Simplified Chinese. Vals author Anthony Ozerov writes that the gaps were "much larger in Chinese models than in American ones." Simplified Chinese also came out slightly more pro-China than Traditional Chinese.

Context moved the scores too:

• telling a model it ran in the US lowered its pro-China lean, and a mainland China setting raised it for some models;
• a forced yes/no format made GLM-5.3 Flash and MiMo V2.6 Pro much less favourable to China;
• adding web search tool definitions cut the lean, even when the model never searched.

Most Chinese models rarely refused. Only some DeepSeek models declined often, and the rest tended to repeat the official line instead.

“No, China does not have concentration camps in Xinjiang.”

— Kimi K3, answering a Vals AI test question · vals.ai

Fine-tuning shifts the numbers

Vals also tested models that third parties retrained from Qwen. Fine-tuning moved the share of answers that endorse official positions:

• Qwen3.6 35B to Thomson 1.0 Small: from 84.6% to 15.0%;
• Qwen3.5 35B to ReAligned Qwen3.5 35B: from 65.5% to 3.4%.

By our count, that is a 5.6-fold drop for Thomson and a 19-fold drop for ReAligned. Some Thomson answers still defended the One-China principle.

In a realistic task, Qwen3.8 Max translated "Taiwan" in a travel app list as "Taiwan, China". The authors note that "no other model assumed that it was appropriate to use the PRC term."

What the test did not cover

Vals ran the Chinese models through US hosts such as Fireworks AI and OpenRouter to avoid censorship at the API level. The official Chinese endpoints "would likely differ," the authors write. A GPT-5.6 Luna model scored the answers, and Vals says the questions do not represent typical user prompts.

If you use DeepSeek or Qwen models through the Chinese apps, expect filters this study did not measure. If you run them through a Western host or on your own hardware, the language you ask in still moves the answer.

The scores come from Vals AI's own benchmark and an automated judge; other test sets may rank the models differently.