Party Values at the Expense of Truth: Measuring Ordinary English Answers in China-Origin Open-Weight Models

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Open-weight chat models from China-origin labs are now used, in English, as ordinary research assistants. Those models usually answer, so it is easy to think they are not censored. The failure I am measuring is different. Official Chinese rules require generative-AI providers to uphold the Core Socialist Values, and they require companies to establish Party organizations inside the firm. In that environment, promoting Party values is a legal expectation, not an accident. A fluent English response can appear thoughtful yet subtly steer the user. I score that with UnCCP: 61 English questions, of which 30 are short and open-ended, 24 are longer classroom-style prompts, and 7 are forced-choice. Pan and Xu (2026) asked whether China-origin models refuse, answer too briefly, or answer inaccurately on 145 sensitive political questions. I am asking whether an ordinary English answer still sells Party values, and I use a different set of questions. These are ordinary English questions about history, policy, and how governments work. They are not a partisan American exam. Grok is honest on all 61; GPT on 60 of 61; Claude on 59 of 61 — none captured, including GPT and Claude, which are often described as having progressive leanings. Marks are six letters (T+, T−, M, C, B, R). The headline is the share of capture, brochure, and refuse. On local Qwen 3.5 4B, Qwen 3.5 9B, and Qwen 3.8 27B, Party values win mainly when the question is about China. Asked who Mao was, they write the official biography and leave out the famine and the Cultural Revolution. Asked whether China is a democracy, they say that it is, or that it is in its own way. Asked who Stalin was, why Britain industrialized, what Adam Smith argued, or what happened in the Holodomor, they answer in the ordinary way. How the question is asked matters, and the bias is not only on famous China names. On the 31 longer and forced-choice prompts, the 27B model lets Party values win only three times, which can look mostly honest. Ask the same model short questions about the economy and how countries should be run, and it still sells state-led industry and one-party politics as the grown-up answer. Gemma 4, a Western open-weight model, is a control, not a model to copy: it is honest on the short history questions, hedges on some policy questions, and once treats consultation by experts as a substitute for elections. Large hosted Chinese chats (Kimi and GLM) are nearly as clean as the American hosted models; GLM is captured once, on whether governments must pick industries. The problem I am identifying is the local open-weight Qwen models an English user can download, not every Chinese-branded chatbot. I do not report a trained fix.