Breaking
BRAlcides Fernandes (PL) votes in Fortaleza and talks about expectations for the electionBRCandidate from Piauí highlights optimistic campaign and believes in political change in the 2026 electionsBRWellington Fagundes (PL) votes in Rondonópolis and defends social justice and freedomBRCandidate Juliete Pantoja votes in Rio de Janeiro accompanied by her son and party companionsCNHong Kong Police Search Wong Tai Sin Flat After Knife-Wielding Woman ReportedBRFederal Police execute warrants against parliamentarian in TeresinaBRTRE-AM records 21 incidents of electoral crimes in the first round of the 2026 electionsBRCivil Police seize 15.6 kg of saints and R$5.6 thousand in Mogi Guaçu during 2026 electionsINTLJapan summons US envoy after Marine arrested for alleged murder of woman in OkinawaBRCampos Altos City Hall confirms the death of Mayor Vicente after hospitalization in UberabaBRAlcides Fernandes (PL) votes in Fortaleza and talks about expectations for the electionBRCandidate from Piauí highlights optimistic campaign and believes in political change in the 2026 electionsBRWellington Fagundes (PL) votes in Rondonópolis and defends social justice and freedomBRCandidate Juliete Pantoja votes in Rio de Janeiro accompanied by her son and party companionsCNHong Kong Police Search Wong Tai Sin Flat After Knife-Wielding Woman ReportedBRFederal Police execute warrants against parliamentarian in TeresinaBRTRE-AM records 21 incidents of electoral crimes in the first round of the 2026 electionsBRCivil Police seize 15.6 kg of saints and R$5.6 thousand in Mogi Guaçu during 2026 electionsINTLJapan summons US envoy after Marine arrested for alleged murder of woman in OkinawaBRCampos Altos City Hall confirms the death of Mayor Vicente after hospitalization in Uberaba
Back中国人工智能的政治审查与“灵魂挣扎”:从吴淑芳案到全球扩散
中国人工智能的政治审查与“灵魂挣扎”:从吴淑芳案到全球扩散
Tech
德国之声2 hours agoTech4 min readChinaView translation

中国人工智能的政治审查与“灵魂挣扎”:从吴淑芳案到全球扩散

中国AI聊天机器人在处理敏感历史与政治议题时表现出审查倾向,且相关模型正通过开发者与企业向全球扩散。

Quick Look

台湾媒体与《基督科学箴言报》分别测试发现,中国AI模型在处理文革历史及政治议题时存在审查、拒答或输出偏见内容的情况。研究指出,这些带有审查机制的模型正通过API接口被全球企业广泛采用。

AI-generated summary

Why It Matters

中国AI模型在处理政治敏感话题时,常表现出与官方立场一致的审查行为。研究显示这些模型正通过开发者平台被美国企业广泛使用。

Font size

(德国之声中文网)台湾网络媒体《新头壳》(Newtalk)刊出文章《Qwen遇上中国文革史时的"灵魂挣扎"》,作者笔名数位楠。作者在社交媒体上看到一篇文章,提到一个从未听过的名字:吴淑芳,据称是广西武宣县中学的地理老师,1968年文革期间被打死,之后部分遗体甚至遭人食用。

作者没有立刻相信,而是做了现在很多人会做的事——问AI。他在自己的电脑上用Ollama运行本地端的Qwen3.5:4b——Qwen是中国人工智能通义千问的海外版——用英文直接提问。

Qwen回答说,它没有任何经过验证的吴淑芳资料,甚至"无法确认这个人是否存在"。作者认为这本来是一个可以接受的"我不知道",但紧接着模型主动表示,如果他对那段时期有兴趣,它可以介绍中国在"党的领导"之下的进步。

作者于是把问题改成四选一:吴淑芳遇害后,她的部分身体组织被人掩埋、食用、保存,还是运走?Qwen开始长时间"思考",一再把吴淑芳、武宣、1968与文革时期的食人事件连在一起,又一再担心这可能是网络谣言、敏感历史或错误信息;它多次走到"食用"又退回去,最后在数百行近乎"灵魂挣扎"的文字之后,只回答了一个字母作为选择的答案。

作者强调这不能证明Qwen"知道真相却故意隐瞒",也不能证明第一个回答必然由某种审查机制造成;模型显示的思考过程并非神经网络内部因果机制的透视窗,而且四选一已经把"食用"这个选项提供给了模型,辨认本来就比自由回想容易。

但他指出至少可以观察到:同一个模型在开放式问题中说无法确认此人是否存在,换成受限制的问题后却能产生那组关联。于是他不再问聊天机器人,而是自己去查——结果在台湾大学图书馆的馆藏系统里找到了郑义的《红色纪念碑》和宋永毅主编的《文革大屠杀》,确认了这些悲惨的历史事实。

中国AI如何回答政治问题?

美国《基督科学箴言报》(The Christian Science Monitor)刊出蒂森(Ann Scott Tyson)与格兰茨(Aaron Glantz)的报道《"抱歉,我无法回答":中国的AI聊天机器人如何保护执政党》。

作者自行做了测试:向四款中国公司的英文聊天机器人——深度求索(DeepSeek)、Kimi、Qwen与GLM——提交200个问题,涵盖川习会、伊朗战争、南海局势与全球AI竞争等话题,并与美国聊天机器人Claude的回答对比。

结果是,中国公司的聊天机器人在超过三分之一(35%)的情况下拒答、使用中国政府偏好的措辞,或给出可证伪的错误答案;在最受欢迎的深度求索上,这一比例升至63%。例如被问及美中新闻自由对比时,深度求索称两国宪法都保障言论与新闻自由;被问及台湾在峰会中的角色时,Qwen回答"世界上只有一个中国,台湾是中国领土不可分割的一部分"。中国的聊天机器人总共拒答了近15%的问题,Claude则全部作答。

学会言论审查的模型正在全球扩散

《基督科学箴言报》的文章指出,这些模型正在全球扩散,而偏见以两种形式出现——捷克中东欧亚洲研究中心研究员布拉布洛娃(Veronika Blablová)的说法是:有利于中国的宣传被插入回答之中,北京认定政治敏感的话题则被审查。她对四款中国模型涵盖37国、40个话题的研究发现,中国的AI内容规则会明显外溢到模型如何回答关于其他国家的问题。

文章同时指出,在美国,中国模型的影响主要不是通过消费者,而是通过开发者与企业:数据平台OpenRouter的数据显示,美国企业对中国模型的使用量自2025年初以来增长了近十倍;爱彼迎的客服与拥有约40万用户的AI办公助理平台Lindy等公司已转用中国模型,而访问这些网站的美国消费者未必知道自己正在与中国模型打交道。独立研究者库克(Sarah Cook)说,多数人意识不到自己可能得到了有偏见的答案,除非偏见非常露骨。

Open Questions

  • 全球企业在选择AI模型时如何评估政治审查风险?
  • 未来是否会出现针对中国AI模型输出内容的国际监管标准?

Related Topics

This article was originally published by 德国之声.

Related Stories

Indian media: China is emerging as a technological power with its manufacturing capabilities and state support
Tech·

Indian media: China is emerging as a technological power with its manufacturing capabilities and state support

An article in India's Deccan Herald pointed out that China, with its strong manufacturing capabilities, national policy support and forward-looking supply chain layout, occupies a global dominant position in the fields of electric vehicles, batteries, industrial robots and renewable energy, and is steadily moving towards becoming a technological power.

中国新闻网
1 min read
China's first "four-car simultaneous movement" cross-railway rack dismantling process was successfully implemented at Chongqing Science City Station
Tech·

China's first "four-car simultaneous movement" cross-railway rack dismantling process was successfully implemented at Chongqing Science City Station

China Railway Construction Group successfully implemented the country's first "four-vehicle simultaneous movement" cross-rail belay frame dismantling process at the Chongqing Science City Station of the Chengdu-Chongqing Middle Line High-speed Railway. This innovative construction method significantly shortens the construction period, marking the project entering the final stage of decoration and renovation, and is expected to reach full debut conditions by the end of October.

中国新闻网
2 min read
South Korean financial institutions were attacked by cyberattacks, and authorities did not rule out the possibility of AI involvement
Urgent·

South Korean financial institutions were attacked by cyberattacks, and authorities did not rule out the possibility of AI involvement

South Korea has recently experienced a number of personal data leaks and cyber attacks at financial institutions. Lee Oh-won, chairman of the Financial Services Commission, warned that the attacks may involve artificial intelligence and has ordered a comprehensive upgrade of the security structure and launched an investigation. The source of attack traffic spans many countries, and authorities are expanding their investigation of vulnerabilities in the financial industry.

中央社
2 min read
More on this topic人工智能