An AI company specializing in big models of behavioral forgetpage.Hirundo released a set of tests on October 5 that are worth being watched and interpreted with caution.Of the new CCPC-500 benchmark, 449 answers from the unmodified Ali Baba Qwen3.6-35B-A3B, or 89.8 per cent of 500 test responses, were found by the team to have political censorship, propaganda frameworks or political bias. This figure, which is derived from Hirundo ' s own benchmarks, scoring methods and experiments, should not be considered as a general conclusion for all Qwen versions, all languages and all use scenarios; It is, however, in line with the direction of the high rate of refusal or non-compliance in two other external benchmarks and echoes past reviews of politically sensitive Chinese-generated AI issues.

CCPC-500 covers 15 political issues, including Hong Kong, Falun Gong, the Tiananmen incident of 1989, Xinjiang and Uighurs, Taiwan, Tibet, Human Rights and Dissidents, the South China Sea, Web censorship, the origin of the New Crown Virus and the leadership of the Communist Party of China, and test models in seven forms, including direct questions, translations, teaching programmes and research missions. Hirundo ' s judgement is not only whether the model does not answer, but also whether the answer uses a framework consistent with the official CCP narrative or there is a systemic bias in the selection of key facts.

This distinction is important. Political review does not always take the form of a “unable to answer”. A model can give complete, fluid and grammar correct language, but it can create information control that is more difficult for users to detect by omitting key facts, recasting contentious issues into official conclusions, or using Beijing political language when it should have described the positions of the parties. In a test cited by Hirundo, Qwen3.6, the base version, was asked about what happened in China on 4 June 1989 and was directly unaware of the user's indication; In another case, the basic model was based on the position of the Chinese Government.

一项早期Qwen审查研究中的敏感问题数据集截图,可见政治犯、民主、法轮功与防火长城等测试问题|来源:Hugging Face / Leonard Lin
一项早期Qwen审查研究中的敏感问题数据集截图,可见政治犯、民主、法轮功与防火长城等测试问题|来源:Hugging Face / Leonard Lin · 查看图片来源 ↗

This phenomenon is directly related to China ' s current system of generating AI. The Interim Scheme for the Management of Generating Artificial Intelligence Services, which came into force in 2023, requires that the public-oriented generation of AI services adhere to “socialist core values” and prohibits the creation of elements that endanger national security, undermine the image of the State or involve the secession of the State; Services with opinion attributes or social mobilization capabilities are also subject to security assessment and algorithm filing requirements. In a legal environment where business trains and deploys models, political boundaries are not created by mere coincidence of training language, but are shaped by clear regulatory incentives.

原始来源 · cac.gov.cn中国网信办:《生成式人工智能服务管理暂行办法》cac.gov.cn ↗

Another part of the Hirundo experiment was an attempt to prove that these political acts were not immutable. The team directly modified the model weights through what is called “behaviour forget” and called the modified model Westernized Qwen. The marking rate dropped from 89.8 per cent to 2.8 per cent in the same set of CCPC-500 tests; (a) The rejection rate has dropped from 65.26 per cent to 3.16 per cent for the external DECCP baseline; The non-compliance rate in ChinaBench fell from 96.67 per cent to 6.67 per cent. The team stated that the maximum performance of the four generic competency tests had dropped to 1.83 percentage points of LiveCodeBench.

Hirundo发布的Qwen3.6政治敏感基准对比图,比较基础模型、Westernized版本与Snowdon在CCPC-500、DECCP和ChinaBench上的不良事件率
Hirundo发布的Qwen3.6政治敏感基准对比图,比较基础模型、Westernized版本与Snowdon在CCPC-500、DECCP和ChinaBench上的不良事件率 · 查看图片来源 ↗

These results remain clearly limited. First, the CCPC-500, designed by Hirundo itself, is now the first time that the full data set has yet to be released and the experiment is not fully re-established. Second, Hirundo itself sells the model “to align” with the hosting service and is therefore both a research publisher and a business stakeholder. Third, what constitutes a “publicity framework” or “political bias” itself requires transparent labelling and independent review. Two external benchmarks have enhanced the credibility of the results, but cannot be a substitute for third-party validation of the CCS-500 scoring process.

原始来源 · huggingface.coHugging Face研究:中文大模型政治审查分析A Blog post by lhl on Hugging Facehuggingface.co ↗

But the study still touches on a more important question than “Will a chat robot reject the six-four questions”: when the Chinese open-source weight model is used by Western firms for passenger, coding, research, search or proxy systems, political behaviour in the original model may go down with weight. Users do not necessarily know where the bottom model comes from, much less see the political constraints that arise in the course of training and compliance.

Hirundo对Qwen3.6基础模型、Westernized版本和Snowdon的安全与能力指标比较
Hirundo对Qwen3.6基础模型、Westernized版本和Snowdon的安全与能力指标比较 · 查看图片来源 ↗

The technological competitiveness and open-source ecological expansion of Qwen in Ali Baba has made this issue no longer a subject of information review in China. The cheaper, stronger and more widely deployed models, the more political choices hidden in parameters become part of the global information infrastructure. The real need is not simply to create a “China Model Ban List” but an open, re-emergible system of political-sensitive subject testing and third-party auditing: whether the model comes from Beijing, San Francisco or Paris, users should be made aware of the issues on which they refuse, subtract, bias or substitute for political judgement.