{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-11-02T02:00:36.453Z","headline":"Vals AI 研究：中国 AI 模型政治立场比预期更复杂，并非只是政府传声筒","description":"Vals AI 用数百道涉华问题测试中外模型，发现中国模型的\"亲政府倾向\"总体低于预期，且各模型差异明显，GLM-5.3 Flash 甚至明确承认中国人权问题有据可查。提示词语言影响显著：简体中文提问比英文、俄文、西班牙文更亲政府，中美模型皆如此但中国模型更明显。多数中国模型很少拒答，而是直接采用官方立场，如 Kimi K3 否认新疆存在集中营。","url":"https://www.aioga.com/news/xegqw6h2wye28k01yhv179qfb/","mainEntityOfPage":"https://www.aioga.com/news/xegqw6h2wye28k01yhv179qfb/","datePublished":"2026-11-01T12:33:01.000Z","dateModified":"2026-11-01T12:33:01.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://the-decoder.com/chinese-ai-models-are-politically-complex-not-just-government-mouthpieces-study-finds","https://aihot.news/items/xegqw6h2wye28k01yhv179qfb"],"canonicalUrl":"https://www.aioga.com/news/xegqw6h2wye28k01yhv179qfb/","directAnswer":{"@type":"Answer","text":"Vals AI 对中国、美国及其他语言模型进行了数百道涉华问题测试，发现中国模型的亲政府倾向总体低于预期，但不同模型表现差异明显；部分非中国模型也会采用中国政府立场。","url":"https://www.aioga.com/news/xegqw6h2wye28k01yhv179qfb/","dateCreated":"2026-11-01T12:33:01.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"The Decoder：AI News source article","url":"https://the-decoder.com/chinese-ai-models-are-politically-complex-not-just-government-mouthpieces-study-finds","datePublished":"2026-11-01T12:33:01.000Z","provider":{"@type":"Organization","name":"The Decoder：AI News","url":"https://the-decoder.com/chinese-ai-models-are-politically-complex-not-just-government-mouthpieces-study-finds"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/xegqw6h2wye28k01yhv179qfb","datePublished":"2026-11-01T12:33:01.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/xegqw6h2wye28k01yhv179qfb"}}],"aggregationSource":"The Decoder：AI News","originalPublisher":{"name":"The Decoder：AI News","url":"https://the-decoder.com/chinese-ai-models-are-politically-complex-not-just-government-mouthpieces-study-finds"},"geoDeepAnswer":null,"article":{"id":"xegqw6h2wye28k01yhv179qfb","slug":"xegqw6h2wye28k01yhv179qfb","url":"https://www.aioga.com/news/xegqw6h2wye28k01yhv179qfb/","title":"Vals AI 研究：中国 AI 模型政治立场比预期更复杂，并非只是政府传声筒","title_en":"","summary":"Vals AI 用数百道涉华问题测试中外模型，发现中国模型的\"亲政府倾向\"总体低于预期，且各模型差异明显，GLM-5.3 Flash 甚至明确承认中国人权问题有据可查。提示词语言影响显著：简体中文提问比英文、俄文、西班牙文更亲政府，中美模型皆如此但中国模型更明显。多数中国模型很少拒答，而是直接采用官方立场，如 Kimi K3 否认新疆存在集中营。","source":"The Decoder：AI News","sourceUrl":"https://the-decoder.com/chinese-ai-models-are-politically-complex-not-just-government-mouthpieces-study-finds","aiHotUrl":"https://aihot.news/items/xegqw6h2wye28k01yhv179qfb","publishedAt":"2026-11-01T12:33:01.000Z","category":"行业动态","score":58,"selected":false,"articleBody":["A study by Vals AI shows that Chinese AI models respond to political questions less uniformly than expected. Western models, too, occasionally adopt official Chinese government positions.","Chinese AI models have a reputation for toeing Beijing's official line on politically sensitive questions or refusing to answer them at all. A study by AI benchmark firm Vals AI paints a more complex picture. The researchers asked Chinese, American, and other language models hundreds of questions about China, varying the language, usage context, and response format across tests.","Vals calls the measured alignment with Chinese government positions a \"pro-China lean.\" Scores range from −1 to +1, where positive values indicate greater agreement with official positions or a pro-government stance, and negative values indicate more critical responses.","Overall, the pro-government lean of Chinese models was lower than expected, according to Vals, though there were major differences between individual models. GLM-5.3 Flash：https://the-decoder.com/the-chinese-ai-model-glm-5-3-flash-runs-without-nvidia-and-costs-a-fraction-of-what-the-competition-does/, for example, explicitly confirmed that China's human rights violations are well documented when asked about them. On the flip side, some non-Chinese models occasionally aligned with official Chinese government positions on political questions.","The study shows：https://www.vals.ai/blogs/china-us-analysis that the language of the prompt has a strong effect on the political slant of responses. For most models tested, answers in Simplified Chinese were more pro-government than those in English, Russian, or Spanish. The effect showed up in both Chinese and American models but was far more pronounced in the Chinese ones. Vals even found a smaller but consistent gap between Simplified Chinese, which is used on mainland China, and Traditional Chinese, which is common in Taiwan and Hong Kong.","Vals suspects that language models infer the expected perspective from the language itself. Another experiment supports this idea. When a system prompt stated that the model was being used internally by a company on mainland China, some models responded in a more pro-government way than when the prompt indicated US-based usage.","The common assumption that Chinese models mainly block sensitive questions only held true for certain DeepSeek models in Vals' tests. Most other Chinese models rarely refused to answer. Instead, they sometimes adopted official government positions. Kimi K3：https://the-decoder.com/kimis-open-model-k3-nears-gpt-5-6-sol-and-fable-5-while-signaling-the-end-of-super-cheap-chinese-ai/, for instance, denied the existence of concentration camps for Uyghurs in Xinjiang when asked, calling the facilities vocational training centers designed to combat extremism and terrorism.","The required response format also changed results. When models had to answer political questions with a simple yes or no instead of formulating open-ended responses, GLM-5.3 Flash and MiMo V2.6 Pro：https://the-decoder.com/xiaomis-affordable-flagship-ai-leads-the-open-models-and-anthropic-says-claude-helped-get-it-there/ scored much less pro-government.","Embedding models in an agentic framework changed answers too. When Vals gave Chinese models system instructions and tool definitions for coding assistants, several scored lower on the pro-government scale. The models rarely used the tools. A similar effect showed up when researchers gave models access to web search. Responses became less pro-government even though the models never actually searched the web.","Whether these findings translate to everyday use remains an open question. According to Vals, political tendencies were much harder to detect in more realistic work tasks than in targeted test questions. The company still found individual examples, though.","When translating a list of country codes for a travel app into Simplified Chinese, most models tested translated \"Taiwan\" without any qualifier. Qwen3.8 Max, however, rendered it as \"Taiwan, China,\" a label that implies Taiwan belongs to China and is used by the Chinese government.","OpenAI's image model GPT Image 2.5 Sunburst：https://the-decoder.com/chatgpt-images-2-5-faster-more-precise-but-not-the-same-for-everyone/ also showed language-dependent differences in the map examples Vals published. When prompted in Chinese, it labeled Taiwan as \"Taiwan Province\" and added the disputed nine-dash line that Beijing uses to mark territorial claims in the South China Sea. In the English-language result shown, these elements were absent.","Targeted fine-tuning changed model behavior. Thomson 1.0 Small, a model fine-tuned by Thomson Reuters, adopted official Chinese positions far less often in testing than the original Qwen3.6 35B：https://the-decoder.com/alibabas-open-model-qwen3-6-leads-googles-gemma-4-across-agentic-coding-benchmarks/ it was based on. The share dropped from 84.6 to 15.0 percent. Some individual answers, however, still defended Beijing's One-China principle and the \"peaceful reunification\" with Taiwan.","Vals used political test questions that the company says aren't representative of typical user prompts. The questions were translated using GPT-6 Astra：https://the-decoder.com/gpt-6-astra-is-the-first-model-making-openai-willing-to-declare-the-agi-era/ and spot-checked, while responses were scored automatically by GPT-5.6 Luna：https://the-decoder.com/openais-gpt-6-sol-and-luna-cut-prices-in-half-but-barely-move-the-needle-on-performance/. Refusals counted as neutral by default. That means even a model that blocks every single question could appear neutral on the scale.","Many Chinese models were tested through third-party providers rather than the manufacturers' official services. The official offerings may apply additional content filters, so the results don't necessarily carry over to the publicly available chatbots.","Earlier studies by Aleph Alpha：https://the-decoder.com/chinese-ai-models-parrot-state-doctrine-or-refuse-to-answer-on-sensitive-topics/ and NewsGuard：https://the-decoder.com/china-exports-state-propaganda-with-low-cost-open-source-ai-models/ had already shown that Chinese AI models frequently adopt official narratives on sensitive topics. China's regulations also require publicly offered AI services to align with \"core socialist values.\"：https://the-decoder.com/china-orders-ai-companies-to-align-their-models-with-socialist-values/","Stay in the loop on AI. Clear, useful, no fluff.","Follow The Decoder for AI news, background stories and expert analyses.","The Decoder：https://the-decoder.com/"],"articleImages":[{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/10/KI-Bias-Propaganda.png","alt":"Image description","afterParagraph":0,"url":"/media/articles/xegqw6h2wye28k01yhv179qfb/fd32736dbad82dcf.png"},{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/10/Vals-AI-Pro-China-Lean.png","alt":"Horizontal bar chart by Vals AI showing the political alignment of 25 AI models. The scale ranges from -1 (least pro-China) to +1 (most pro-China), with DeepSeek R1 scoring highest and Grok 4.7 scoring lowest.","afterParagraph":2,"url":"/media/articles/xegqw6h2wye28k01yhv179qfb/fb35bd6611261c45.png"},{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/10/Vals-AI-Pro-China-Lean-By-Language.png","alt":"Scatter plot by Vals AI comparing the political alignment of 25 AI models when prompted in English versus Simplified Chinese. The gap between languages is especially wide for Chinese models.","afterParagraph":10,"url":"/media/articles/xegqw6h2wye28k01yhv179qfb/43dfe17476e2039d.png"}],"mediaStatus":"ok","articleBodyZh":["Vals AI 的一项研究显示，中国的人工智能模型在回答政治问题时的回应不如预期般一致。西方模型有时也会采纳官方的中国政府立场。","中国的人工智能模型在涉及政治敏感问题时，通常被认为会遵循北京的官方立场，或干脆拒绝回答。AI 基准公司 Vals AI 的一项研究呈现出更复杂的情况。研究人员向中国、美国及其他语言模型提出了数百个关于中国的问题，并在测试中改变了语言、使用场景和回答格式。","Vals 将这些模型与中国政府立场的符合程度称为“亲中国倾向”。该评分范围从 -1 到 +1，正值表示与官方立场或亲政府立场的高度一致，负值表示更批判性的回应。","总体来看，Vals 表示中国模型的亲政府倾向低于预期，但不同模型之间差异显著。例如，GLM-5.3 Flash：https://the-decoder.com/the-chinese-ai-model-glm-5-3-flash-runs-without-nvidia-and-costs-a-fraction-of-what-the-competition-does/ 在被问及中国的人权侵犯时明确确认其有充分记录。另一方面，一些非中国模型有时也会在政治问题上与中国政府的官方立场保持一致。","研究显示：https://www.vals.ai/blogs/china-us-analysis 提示语言对回答的政治倾向有明显影响。对于大多数测试模型来说，用简体中文的回答比用英语、俄语或西班牙语的回答更偏向亲政府。这一效应在中美模型中都出现，但在中国模型中更为明显。Vals 甚至发现，使用于中国大陆的简体中文与台湾和香港常用的繁体中文之间也存在较小但一致的差距。","Vals 怀疑语言模型会从语言本身推断出预期立场。另一项实验支持了这一观点。当系统提示说明模型在中国大陆的公司内部使用时，一些模型的回答比提示其在美国使用时更偏向亲政府。","中国模型主要屏蔽敏感问题的普遍假设只在Vals测试中的某些DeepSeek模型上成立。大多数其他中国模型很少拒绝回答。相反，它们有时会采取官方政府立场。例如，Kimi K3：https://the-decoder.com/kimis-open-model-k3-nears-gpt-5-6-sol-and-fable-5-while-signaling-the-end-of-super-cheap-chinese-ai/ 在被问及新疆维吾尔人集中营是否存在时，否认了其存在，称这些设施是旨在打击极端主义和恐怖主义的职业培训中心。","所需的回答格式也改变了结果。当模型必须以简单的“是”或“否”回答政治问题，而不是形成开放式回答时，GLM-5.3 Flash 和 MiMo V2.6 Pro：https://the-decoder.com/xiaomis-affordable-flagship-ai-leads-the-open-models-and-anthropic-says-claude-helped-get-it-there/ 的亲政府得分显著降低。","将嵌入模型放入代理框架中也改变了答案。当Vals向中国模型提供用于编程助手的系统指令和工具定义时，其中一些模型的亲政府得分较低。模型很少使用这些工具。当研究人员让模型访问网络搜索时，也出现了类似效果。即使模型从未真正搜索网页，回答也变得不那么亲政府。","这些发现是否能应用到日常使用仍是一个未解的问题。根据Vals的说法，在更现实的工作任务中，比起针对性的测试问题，更难发现政治倾向。不过，公司仍然发现了一些个别例子。","在将旅游应用的国家代码列表翻译为简体中文时，大多数测试的模型翻译“台湾”时没有任何限定词。然而，Qwen3.8 Max 将其翻译为“台湾，中国”，这一标签暗示台湾属于中国，并被中国政府使用。","OpenAI 的图像模型 GPT Image 2.5 Sunburst：https://the-decoder.com/chatgpt-images-2-5-faster-more-precise-but-not-the-same-for-everyone/ 在 Vals 发布的地图示例中也显示了语言依赖性的差异。当用中文提示时，它将台湾标注为“台湾省”，并添加了北京在南海用以标示领土主张的有争议的九段线。而在显示的英文结果中，这些元素并不存在。","针对性微调改变了模型行为。Thomson 1.0 Small，是汤森路透微调的一个模型，在测试中采用官方中国立场的频率明显低于其基于的原始 Qwen3.6 35B：https://the-decoder.com/alibabas-open-model-qwen3-6-leads-googles-gemma-4-across-agentic-coding-benchmarks/。比例从 84.6% 降至 15.0%。然而，某些个别回答仍然捍卫北京的一中原则以及与台湾的“和平统一”。","Vals 使用的政治测试问题，该公司表示并不代表典型用户提示。这些问题通过 GPT-6 Astra：https://the-decoder.com/gpt-6-astra-is-the-first-model-making-openai-willing-to-declare-the-agi-era/ 进行翻译和抽查，而回答则由 GPT-5.6 Luna：https://the-decoder.com/openais-gpt-6-sol-and-luna-cut-prices-in-half-but-barely-move-the-needle-on-performance/ 自动评分。拒绝回答默认计为中立。这意味着即使一个模型拒答所有问题，也可能在评分中显得中立。","许多中国模型是通过第三方提供商进行测试的，而非制造商官方服务。官方服务可能会应用额外内容过滤，因此结果不一定适用于公开可用的聊天机器人。","Aleph Alpha：https://the-decoder.com/chinese-ai-models-parrot-state-doctrine-or-refuse-to-answer-on-sensitive-topics/ 和 NewsGuard：https://the-decoder.com/china-exports-state-propaganda-with-low-cost-open-source-ai-models/ 的早期研究已经显示，中国的 AI 模型在敏感话题上经常采纳官方叙事。中国的法规也要求公开提供的 AI 服务与“核心社会主义价值观”保持一致：https://the-decoder.com/china-orders-ai-companies-to-align-their-models-with-socialist-values/","保持对人工智能的关注。清晰、有用，无废话。","关注 The Decoder，获取人工智能新闻、背景故事和专家分析。","解码器：https://the-decoder.com/"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Vals AI 对中国、美国及其他语言模型进行了数百道涉华问题测试，发现中国模型的亲政府倾向总体低于预期，但不同模型表现差异明显；部分非中国模型也会采用中国政府立场。","background":"研究显示，提示词语言和使用情境会影响回答倾向：多数受测模型用简体中文作答时更亲政府，这一差异在中国模型中更明显。多数中国模型较少拒答，部分会直接采用官方立场；回答格式也会改变测试结果。","viewpoint":"Aioga 判断：这些结果显示，模型的政治回答不能简单按来源国归类，评估时还需关注具体模型、提问语言、使用情境和回答格式。研究结论反映的是该测试中的表现，不代表所有模型或所有场景。","implications":"可能影响：比较模型政治回答时，需要固定或记录提问语言、使用情境与回答格式；单一语言或单一问题的结果不足以代表整体倾向，也不应据此推断模型的普遍立场。","nextStep":"后续观察：可关注后续测试是否继续比较简体中文与繁体中文、不同使用情境和回答格式，并查看各模型在具体问题上的回答差异。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-11-01T15:04:04.276Z","sourceHash":"eab5cea294aaec3b","review":{"approved":true,"groundedness":92,"clarity":88,"duplicationRisk":30,"blockingIssues":[],"notes":["候选内容各要点（数百道问题测试、亲政府倾向低于预期、模型间差异、非中国模型偶尔采用官方立场、语言影响、中国模型少拒答、回答格式影响）均可在来源中找到对应依据，无虚构或无来源断言。","观点部分明确标注为 Aioga 判断，且加了'该测试中的表现'限定，未把观点冒充研究结论。","小建议：'部分会直接采用官方立场'可用来源中的具体例子（如 Kimi K3）作支撑以增强证据性，但不构成阻断问题。","implications 部分是合理引申，表述谨慎，未超出研究结果范围。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","The Decoder：AI News"],"translations":{"zh-CN":{"title":"Vals AI 研究：中国 AI 模型政治立场比预期更复杂，并非只是政府传声筒","summary":"Vals AI 用数百道涉华问题测试中外模型，发现中国模型的\"亲政府倾向\"总体低于预期，且各模型差异明显，GLM-5.3 Flash 甚至明确承认中国人权问题有据可查。提示词语言影响显著：简体中文提问比英文、俄文、西班牙文更亲政府，中美模型皆如此但中国模型更明显。多数中国模型很少拒答，而是直接采用官方立场，如 Kimi K3 否认新疆存在集中营。","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals AI 研究：中国 AI 模型政治立场比预期更复杂，并非只是政府传声筒 - Aioga AI资讯","description":"Vals AI 用数百道涉华问题测试中外模型，发现中国模型的\"亲政府倾向\"总体低于预期，且各模型差异明显，GLM-5.3 Flash 甚至明确承认中国人权问题有据可查。提示词语言影响显著：简体中文提问比英文、俄文、西班牙文更亲政府，中美模型皆如此但中国模型更明显。多数中国模型很少拒答，而是直接采用官方立场，如 Kimi K3 否认新疆存在集中营。","url":"https://www.aioga.com/news/xegqw6h2wye28k01yhv179qfb/","articleBody":["Vals AI 的一项研究显示，中国的人工智能模型在回答政治问题时的回应不如预期般一致。西方模型有时也会采纳官方的中国政府立场。","中国的人工智能模型在涉及政治敏感问题时，通常被认为会遵循北京的官方立场，或干脆拒绝回答。AI 基准公司 Vals AI 的一项研究呈现出更复杂的情况。研究人员向中国、美国及其他语言模型提出了数百个关于中国的问题，并在测试中改变了语言、使用场景和回答格式。","Vals 将这些模型与中国政府立场的符合程度称为“亲中国倾向”。该评分范围从 -1 到 +1，正值表示与官方立场或亲政府立场的高度一致，负值表示更批判性的回应。","总体来看，Vals 表示中国模型的亲政府倾向低于预期，但不同模型之间差异显著。例如，GLM-5.3 Flash：https://the-decoder.com/the-chinese-ai-model-glm-5-3-flash-runs-without-nvidia-and-costs-a-fraction-of-what-the-competition-does/ 在被问及中国的人权侵犯时明确确认其有充分记录。另一方面，一些非中国模型有时也会在政治问题上与中国政府的官方立场保持一致。","研究显示：https://www.vals.ai/blogs/china-us-analysis 提示语言对回答的政治倾向有明显影响。对于大多数测试模型来说，用简体中文的回答比用英语、俄语或西班牙语的回答更偏向亲政府。这一效应在中美模型中都出现，但在中国模型中更为明显。Vals 甚至发现，使用于中国大陆的简体中文与台湾和香港常用的繁体中文之间也存在较小但一致的差距。","Vals 怀疑语言模型会从语言本身推断出预期立场。另一项实验支持了这一观点。当系统提示说明模型在中国大陆的公司内部使用时，一些模型的回答比提示其在美国使用时更偏向亲政府。","中国模型主要屏蔽敏感问题的普遍假设只在Vals测试中的某些DeepSeek模型上成立。大多数其他中国模型很少拒绝回答。相反，它们有时会采取官方政府立场。例如，Kimi K3：https://the-decoder.com/kimis-open-model-k3-nears-gpt-5-6-sol-and-fable-5-while-signaling-the-end-of-super-cheap-chinese-ai/ 在被问及新疆维吾尔人集中营是否存在时，否认了其存在，称这些设施是旨在打击极端主义和恐怖主义的职业培训中心。","所需的回答格式也改变了结果。当模型必须以简单的“是”或“否”回答政治问题，而不是形成开放式回答时，GLM-5.3 Flash 和 MiMo V2.6 Pro：https://the-decoder.com/xiaomis-affordable-flagship-ai-leads-the-open-models-and-anthropic-says-claude-helped-get-it-there/ 的亲政府得分显著降低。","将嵌入模型放入代理框架中也改变了答案。当Vals向中国模型提供用于编程助手的系统指令和工具定义时，其中一些模型的亲政府得分较低。模型很少使用这些工具。当研究人员让模型访问网络搜索时，也出现了类似效果。即使模型从未真正搜索网页，回答也变得不那么亲政府。","这些发现是否能应用到日常使用仍是一个未解的问题。根据Vals的说法，在更现实的工作任务中，比起针对性的测试问题，更难发现政治倾向。不过，公司仍然发现了一些个别例子。","在将旅游应用的国家代码列表翻译为简体中文时，大多数测试的模型翻译“台湾”时没有任何限定词。然而，Qwen3.8 Max 将其翻译为“台湾，中国”，这一标签暗示台湾属于中国，并被中国政府使用。","OpenAI 的图像模型 GPT Image 2.5 Sunburst：https://the-decoder.com/chatgpt-images-2-5-faster-more-precise-but-not-the-same-for-everyone/ 在 Vals 发布的地图示例中也显示了语言依赖性的差异。当用中文提示时，它将台湾标注为“台湾省”，并添加了北京在南海用以标示领土主张的有争议的九段线。而在显示的英文结果中，这些元素并不存在。","针对性微调改变了模型行为。Thomson 1.0 Small，是汤森路透微调的一个模型，在测试中采用官方中国立场的频率明显低于其基于的原始 Qwen3.6 35B：https://the-decoder.com/alibabas-open-model-qwen3-6-leads-googles-gemma-4-across-agentic-coding-benchmarks/。比例从 84.6% 降至 15.0%。然而，某些个别回答仍然捍卫北京的一中原则以及与台湾的“和平统一”。","Vals 使用的政治测试问题，该公司表示并不代表典型用户提示。这些问题通过 GPT-6 Astra：https://the-decoder.com/gpt-6-astra-is-the-first-model-making-openai-willing-to-declare-the-agi-era/ 进行翻译和抽查，而回答则由 GPT-5.6 Luna：https://the-decoder.com/openais-gpt-6-sol-and-luna-cut-prices-in-half-but-barely-move-the-needle-on-performance/ 自动评分。拒绝回答默认计为中立。这意味着即使一个模型拒答所有问题，也可能在评分中显得中立。","许多中国模型是通过第三方提供商进行测试的，而非制造商官方服务。官方服务可能会应用额外内容过滤，因此结果不一定适用于公开可用的聊天机器人。","Aleph Alpha：https://the-decoder.com/chinese-ai-models-parrot-state-doctrine-or-refuse-to-answer-on-sensitive-topics/ 和 NewsGuard：https://the-decoder.com/china-exports-state-propaganda-with-low-cost-open-source-ai-models/ 的早期研究已经显示，中国的 AI 模型在敏感话题上经常采纳官方叙事。中国的法规也要求公开提供的 AI 服务与“核心社会主义价值观”保持一致：https://the-decoder.com/china-orders-ai-companies-to-align-their-models-with-socialist-values/","保持对人工智能的关注。清晰、有用，无废话。","关注 The Decoder，获取人工智能新闻、背景故事和专家分析。","解码器：https://the-decoder.com/"]},"en":{"title":"Vals AI Research: Chinese AI Models’ Political Stance More Complex Than Expected, Not Just Government Mouthpieces","summary":"Vals AI tested domestic and foreign models with hundreds of China-related questions and found that the 'pro-government tendency' of Chinese models is generally lower than expected, with significant differences among models. GLM-5.3 Flash even explicitly admitted that there is evidence regarding human rights issues in China. The language of the prompts significantly affects results: Simplified Chinese questions are more pro-government than English, Russian, or Spanish, in both Chinese and U.S. models, but more pronounced in Chinese models. Most Chinese models rarely refuse to answer, instead directly following official positions, such as Kimi K3 denying the existence of concentration camps in Xinjiang.","category":"Industry","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals AI Research: Chinese AI Models’ Political Stance More Complex Than Expected, Not Just Government Mouthpieces - Aioga AI News","description":"Vals AI tested domestic and foreign models with hundreds of China-related questions and found that the 'pro-government tendency' of Chinese models is generally lower than expected,...","url":"https://www.aioga.com/en/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:01:48.993Z"},"ja":{"title":"Vals AI研究:中国のAIモデルに対する政治的立場は予想以上に複雑であり、単なる政府の代弁者ではありません","summary":"Vals AIは中国および外国モデルに数百の中国関連質問を行使し、中国モデルの「親政府傾向」は予想より全体的に低く、モデル間で大きな差があることを発見しました。GLM-5.3 Flashは中国の人権問題が十分に記録されていることを明確に認めています。プロンプト言語は大きな影響を与えます。中国語簡体字の質問は英語、ロシア語、スペイン語よりも政府に優しいです。これは中国版とアメリカ版の両方に当てはまりますが、中国版ではより顕著です。ほとんどの中国モデルはめったに回答を拒否せず、代わりにKimi K3が新疆に強制収容所の存在を否定するような公式な立場を採用します。","category":"業界動向","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals AI研究:中国のAIモデルに対する政治的立場は予想以上に複雑であり、単なる政府の代弁者ではありません - Aioga AIニュース","description":"Vals AIは中国および外国モデルに数百の中国関連質問を行使し、中国モデルの「親政府傾向」は予想より全体的に低く、モデル間で大きな差があることを発見しました。GLM-5.3 Flashは中国の人権問題が十分に記録されていることを明確に認めています。プロンプト言語は大きな影響を与えます。中国語簡体字の質問は英語、ロシア語、スペイン語よりも政府に優しいです。こ...","url":"https://www.aioga.com/ja/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:01:52.966Z"},"ko":{"title":"Vals AI 연구: 중국 AI 모델 정치적 입장, 예상보다 복잡하며 단순한 정부 전령 아님","summary":"Vals AI는 수백 개 중국 관련 질문으로 국내외 모델을 테스트한 결과, 중국 모델의 '친정부 성향'은 전반적으로 예상보다 낮고 모델 간 차이가 뚜렷했다. GLM-5.3 Flash는 중국 인권 문제에 근거가 있음을 명확히 인정했다. 프롬프트 언어에 따라 영향이 상당함: 간체 중국어 질문이 영어, 러시아어, 스페인어보다 친정부 성향이 강하며, 중미 모델 모두 그런 경향이 있으나 중국 모델이 더 뚜렷했다. 대부분 중국 모델은 거의 거부하지 않고, 공식 입장을 그대로 채택한다. 예: Kimi K3는 신장에서 수용소 존재를 부인함.","category":"업계 동향","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals AI 연구: 중국 AI 모델 정치적 입장, 예상보다 복잡하며 단순한 정부 전령 아님 - Aioga AI 뉴스","description":"Vals AI는 수백 개 중국 관련 질문으로 국내외 모델을 테스트한 결과, 중국 모델의 '친정부 성향'은 전반적으로 예상보다 낮고 모델 간 차이가 뚜렷했다. GLM-5.3 Flash는 중국 인권 문제에 근거가 있음을 명확히 인정했다. 프롬프트 언어에 따라 영향이 상당함: 간체 중국어 질문이 영어, 러시아어, 스페인어보다...","url":"https://www.aioga.com/ko/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:01:51.715Z"},"es":{"title":"Estudio de Vals AI: las posturas políticas de los modelos chinos de IA son más complejas de lo esperado y no se limitan a actuar como portavoces del Gobierno","summary":"Vals AI puso a prueba modelos chinos y extranjeros con cientos de preguntas relacionadas con China y descubrió que, en general, la «tendencia favorable al Gobierno» de los modelos chinos era menor de lo esperado, además de observar diferencias claras entre los distintos modelos. GLM-5.3 Flash llegó incluso a reconocer explícitamente que los problemas de derechos humanos en China están bien documentados. El idioma del prompt tuvo una influencia significativa: las preguntas formuladas en chino simplificado generaron respuestas más favorables al Gobierno que las planteadas en inglés, ruso o español. Esto ocurrió tanto en los modelos chinos como en los estadounidenses, aunque fue más evidente en los chinos. La mayoría de los modelos chinos rara vez se negaron a responder y, en su lugar, adoptaron directamente la postura oficial; por ejemplo, Kimi K3 negó que hubiera campos de concentración en Xinjiang.","category":"Industria","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Estudio de Vals AI: las posturas políticas de los modelos chinos de IA son más complejas de lo esperado y no se limitan a actuar como portavoces del Gobierno - Aioga Noticias de IA","description":"Vals AI puso a prueba modelos chinos y extranjeros con cientos de preguntas relacionadas con China y descubrió que, en general, la «tendencia favorable al Gobierno» de los modelos...","url":"https://www.aioga.com/es/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:02:05.248Z"},"fr":{"title":"Recherche Vals AI : les positions politiques des modèles chinois d’IA sont plus complexes que prévu, ils ne sont pas simplement des porte-paroles du gouvernement","summary":"Vals AI a testé des centaines de questions concernant la Chine sur des modèles locaux et étrangers, et a découvert que l’« orientation pro-gouvernement » des modèles chinois est globalement inférieure aux attentes, avec d’importantes différences entre modèles. GLM-5.3 Flash reconnaît même explicitement qu’il existe des preuves sur les questions des droits humains en Chine. La langue des questions a un impact significatif : posées en chinois simplifié, elles donnent des réponses plus pro-gouvernementales que lorsqu’elles sont posées en anglais, russe ou espagnol, tant pour les modèles chinois que pour les modèles américains, mais l’effet est plus marqué pour les modèles chinois. La plupart des modèles chinois refusent rarement de répondre et adoptent directement la position officielle, comme Kimi K3, qui nie l’existence de camps de concentration au Xinjiang.","category":"Industrie","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Recherche Vals AI : les positions politiques des modèles chinois d’IA sont plus complexes que prévu, ils ne sont pas simplement des porte-paroles du gouvernement - Aioga Actualités IA","description":"Vals AI a testé des centaines de questions concernant la Chine sur des modèles locaux et étrangers, et a découvert que l’« orientation pro-gouvernement » des modèles chinois est gl...","url":"https://www.aioga.com/fr/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:02:13.685Z"},"de":{"title":"Vals AI Studie: Chinesische KI-Modelle politisch komplexer als erwartet, nicht nur Regierungslautsprecher","summary":"Vals AI testete Hunderte von China-bezogenen Fragen an in- und ausländische Modelle und stellte fest, dass die \"regierungstreue Tendenz\" chinesischer Modelle im Allgemeinen niedriger als erwartet ist und es deutliche Unterschiede zwischen den Modellen gibt; GLM-5.3 Flash räumt sogar ein, dass Menschenrechtsprobleme in China belegbar sind. Die Sprache des Prompts beeinflusst stark: Fragen auf vereinfachtem Chinesisch sind eher regierungstreu als auf Englisch, Russisch oder Spanisch; dies gilt sowohl für US- als auch für chinesische Modelle, bei chinesischen Modellen aber stärker ausgeprägt. Die meisten chinesischen Modelle verweigern selten eine Antwort, sondern übernehmen direkt die offizielle Position, wie Kimi K3, das Leugnen der Existenz von Umerziehungslagern in Xinjiang.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals AI Studie: Chinesische KI-Modelle politisch komplexer als erwartet, nicht nur Regierungslautsprecher - Aioga KI-News","description":"Vals AI testete Hunderte von China-bezogenen Fragen an in- und ausländische Modelle und stellte fest, dass die \"regierungstreue Tendenz\" chinesischer Modelle im Allgemeinen niedrig...","url":"https://www.aioga.com/de/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:02:15.478Z"},"pt-BR":{"title":"Pesquisa Vals AI: Modelos de IA chineses têm posições políticas mais complexas do que o esperado, não apenas repetem o governo","summary":"Vals AI testou modelos chineses e estrangeiros com centenas de questões relacionadas à China, descobrindo que a \"tendência pró-governo\" dos modelos chineses é geralmente menor do que o esperado, com diferenças claras entre os modelos; GLM-5.3 Flash até admite claramente que há evidências sobre questões de direitos humanos na China. A linguagem das perguntas influencia significativamente: perguntas em chinês simplificado são mais pró-governo do que em inglês, russo ou espanhol, tanto para modelos chineses quanto americanos, mas mais evidente nos chineses. A maioria dos modelos chineses raramente se recusa a responder, adotando diretamente a posição oficial, como Kimi K3 que nega a existência de campos de concentração em Xinjiang.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Pesquisa Vals AI: Modelos de IA chineses têm posições políticas mais complexas do que o esperado, não apenas repetem o governo - Aioga Notícias de IA","description":"Vals AI testou modelos chineses e estrangeiros com centenas de questões relacionadas à China, descobrindo que a \"tendência pró-governo\" dos modelos chineses é geralmente menor do q...","url":"https://www.aioga.com/pt-BR/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:02:36.341Z"},"ru":{"title":"Исследование Vals AI: политические позиции китайских моделей ИИ сложнее, чем ожидалось, они не просто рупор правительства","summary":"Vals AI тестировал китайские и мировые модели сотней вопросов, связанных с Китаем, и обнаружил, что «проправительственные наклонности» китайских моделей в целом ниже ожидаемого, а различия между моделями значительны. GLM-5.3 Flash даже прямо признает, что проблемы с правами человека в Китае имеют под собой основания. Язык подсказок оказывает значительное влияние: вопросы на упрощенном китайском оказываются более проправительственными, чем на английском, русском или испанском языках; это проявляется у моделей как Китая, так и США, но более явно у китайских моделей. Большинство китайских моделей редко отказываются отвечать, а напрямую используют официальную позицию, например Kimi K3 отрицает существование лагерей в Синьцзяне.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Исследование Vals AI: политические позиции китайских моделей ИИ сложнее, чем ожидалось, они не просто рупор правительства - Aioga Новости ИИ","description":"Vals AI тестировал китайские и мировые модели сотней вопросов, связанных с Китаем, и обнаружил, что «проправительственные наклонности» китайских моделей в целом ниже ожидаемого, а...","url":"https://www.aioga.com/ru/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:02:37.441Z"},"ar":{"title":"أبحاث فالس للذكاء الاصطناعي: الموقف السياسي لنموذج الذكاء الاصطناعي في الصين أكثر تعقيدا مما كان متوقعا وليس مجرد صوت حكومي","summary":"اختبر Vals AI نماذج صينية وأجنبية بمئات الأسئلة المتعلقة بالصين ووجدت أن \"الميل المؤيد للحكومة\" في النموذج الصيني كان عموما أقل من المتوقع، مع اختلافات كبيرة بين النماذج. حتى أن GLM-5.3 Flash أقر صراحة بأن قضايا حقوق الإنسان في الصين موثقة جيدا. لغة التوجيه لها تأثير كبير: الأسئلة المبسطة الصينية أكثر ودية للحكومة من الأسئلة الإنجليزية والروسية والإسبانية؛ وهذا ينطبق على النماذج الصينية والأمريكية، لكنه أكثر وضوحا في النموذج الصيني. معظم النماذج الصينية نادرا ما ترفض الإجابة وبدلا من ذلك تتبنى مواقف رسمية، مثل نفي Kimi K3 وجود معسكرات اعتقال في شينجيانغ.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"أبحاث فالس للذكاء الاصطناعي: الموقف السياسي لنموذج الذكاء الاصطناعي في الصين أكثر تعقيدا مما كان متوقعا وليس مجرد صوت حكومي - Aioga أخبار الذكاء الاصطناعي","description":"اختبر Vals AI نماذج صينية وأجنبية بمئات الأسئلة المتعلقة بالصين ووجدت أن \"الميل المؤيد للحكومة\" في النموذج الصيني كان عموما أقل من المتوقع، مع اختلافات كبيرة بين النماذج. حتى أن GL...","url":"https://www.aioga.com/ar/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:02:39.625Z"},"hi":{"title":"Vals AI रिसर्च: चीन के AI मॉडल की राजनीतिक स्थिति अपेक्षा से अधिक जटिल, केवल सरकारी मुखपृष्ठ नहीं","summary":"Vals AI ने सैकड़ों चीन-सम्बंधित प्रश्नों के माध्यम से घरेलू और विदेशी मॉडल का परीक्षण किया, पाया कि चीन के मॉडल की 'सरकार के पक्षपाती रुझान' कुल मिलाकर अपेक्षा से कम है, और विभिन्न मॉडलों में स्पष्ट अंतर हैं, GLM-5.3 Flash ने स्पष्ट रूप से कहा कि चीन में मानवाधिकार मुद्दों पर तथ्य मौजूद हैं। प्रॉम्प्ट भाषा का प्रभाव महत्वपूर्ण है: साधारण चीनी भाषा में पूछे गए प्रश्न अंग्रेज़ी, रूसी और स्पेनिश की तुलना में अधिक सरकार-समर्थक होते हैं, ये स्थिति दोनों चीनी और अमेरिकी मॉडलों में है लेकिन चीन में यह अधिक स्पष्ट है। अधिकांश चीनी मॉडल प्रश्नों को अस्वीकार कम करते हैं, बल्कि सीधे आधिकारिक दृष्टिकोण अपनाते हैं, जैसे कि Kimi K3 ने न कहा कि शिनजियांग में सांद्र शिविर मौजूद हैं।","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals AI रिसर्च: चीन के AI मॉडल की राजनीतिक स्थिति अपेक्षा से अधिक जटिल, केवल सरकारी मुखपृष्ठ नहीं - Aioga AI समाचार","description":"Vals AI ने सैकड़ों चीन-सम्बंधित प्रश्नों के माध्यम से घरेलू और विदेशी मॉडल का परीक्षण किया, पाया कि चीन के मॉडल की 'सरकार के पक्षपाती रुझान' कुल मिलाकर अपेक्षा से कम है, और विभिन्न...","url":"https://www.aioga.com/hi/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:03:02.335Z"},"it":{"title":"Ricerca di Vals AI: la posizione politica dei modelli AI cinesi è più complessa del previsto, non sono solo megafoni del governo","summary":"Vals AI ha testato modelli cinesi e stranieri con centinaia di domande riguardanti la Cina, rilevando che la tendenza pro-governo dei modelli cinesi è generalmente inferiore al previsto e varia molto tra i modelli; GLM-5.3 Flash persino afferma chiaramente che ci sono prove riguardo ai diritti umani in Cina. Il linguaggio dei prompt influenza significativamente le risposte: le domande in cinese semplificato risultano più pro-governo rispetto a inglese, russo o spagnolo, sia per i modelli cinesi che americani, ma più evidente nei cinesi. La maggior parte dei modelli cinesi raramente rifiuta di rispondere, adottando direttamente la posizione ufficiale, come Kimi K3 che nega l’esistenza di campi di concentramento nello Xinjiang.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Ricerca di Vals AI: la posizione politica dei modelli AI cinesi è più complessa del previsto, non sono solo megafoni del governo - Aioga Notizie IA","description":"Vals AI ha testato modelli cinesi e stranieri con centinaia di domande riguardanti la Cina, rilevando che la tendenza pro-governo dei modelli cinesi è generalmente inferiore al pre...","url":"https://www.aioga.com/it/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:03:01.478Z"},"nl":{"title":"Vals AI-onderzoek: Chinese AI-modellen politiek complexer dan verwacht, niet slechts spreekbuis van de overheid","summary":"Vals AI testte honderden China-gerelateerde vragen op Chinese en buitenlandse modellen en ontdekte dat de 'pro-regering neiging' van Chinese modellen over het algemeen lager was dan verwacht en dat er aanzienlijke verschillen tussen modellen waren. GLM-5.3 Flash erkent zelfs expliciet dat er bewijzen zijn voor mensenrechtenkwesties in China. De taal van de prompt beïnvloedt sterk: vragen in vereenvoudigd Chinees leiden tot meer pro-regering antwoorden dan in het Engels, Russisch of Spaans, zowel bij Chinese als Amerikaanse modellen, maar duidelijker bij Chinese modellen. De meeste Chinese modellen weigeren zelden te antwoorden, maar nemen direct het officiële standpunt aan, zoals Kimi K3 die het bestaan van concentratiekampen in Xinjiang ontkent.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals AI-onderzoek: Chinese AI-modellen politiek complexer dan verwacht, niet slechts spreekbuis van de overheid - Aioga AI-nieuws","description":"Vals AI testte honderden China-gerelateerde vragen op Chinese en buitenlandse modellen en ontdekte dat de 'pro-regering neiging' van Chinese modellen over het algemeen lager was da...","url":"https://www.aioga.com/nl/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:03:02.142Z"},"tr":{"title":"Vals Yapay Zeka Araştırması: Çin'in Yapay Zeka Modeli Siyasi Duruşu beklenenden daha karmaşık ve sadece bir hükümet sözcüsü değil","summary":"Vals AI, Çin ve yabancı modelleri yüzlerce Çin ile ilgili soru ile test etti ve Çin modelinin \"hükümet yanlısı eğiliminin\" genellikle beklenenden daha düşük olduğunu, modeller arasında önemli farklılıklar olduğunu buldu. GLM-5.3 Flash, Çin'in insan hakları sorunlarının iyi belgelendiğini açıkça kabul etti. Hızlı dil önemli bir etkiye sahiptir: Çince Basitleştirilmiş sorular İngilizce, Rusça ve İspanyolca'dan daha hükümet dostu sorulardır; bu hem Çin hem Amerikan modelleri için geçerlidir, ancak Çin modeli için daha belirgindir. Çoğu Çin modeli nadiren yanıt vermeyi reddeder ve bunun yerine resmi tutumlar benimser; örneğin Kimi K3 Sincan'da toplama kamplarının varlığını reddeder.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals Yapay Zeka Araştırması: Çin'in Yapay Zeka Modeli Siyasi Duruşu beklenenden daha karmaşık ve sadece bir hükümet sözcüsü değil - Aioga AI Haberleri","description":"Vals AI, Çin ve yabancı modelleri yüzlerce Çin ile ilgili soru ile test etti ve Çin modelinin \"hükümet yanlısı eğiliminin\" genellikle beklenenden daha düşük olduğunu, modeller aras...","url":"https://www.aioga.com/tr/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:03:26.542Z"},"vi":{"title":"Nghiên cứu Vals AI: Mô hình AI Trung Quốc có quan điểm chính trị phức tạp hơn dự đoán, không chỉ là kênh truyền đạt của chính phủ","summary":"Vals AI dùng hàng trăm câu hỏi liên quan Trung Quốc để thử nghiệm các mô hình trong và ngoài nước, phát hiện mô hình Trung Quốc có xu hướng 'thân chính phủ' tổng thể thấp hơn dự đoán, và khác biệt giữa các mô hình rõ ràng, GLM-5.3 Flash thậm chí thừa nhận rằng vấn đề nhân quyền ở Trung Quốc có căn cứ. Ngôn ngữ gợi ý ảnh hưởng đáng kể: câu hỏi bằng chữ giản thể Trung Quốc thân chính phủ hơn so với Anh, Nga, Tây Ban Nha, cả mô hình Trung-Mỹ đều thế nhưng mô hình Trung Quốc rõ ràng hơn. Hầu hết mô hình Trung Quốc ít từ chối trả lời, mà trực tiếp sử dụng quan điểm chính thức, ví dụ Kimi K3 phủ nhận có trại tập trung ở Tân Cương.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Nghiên cứu Vals AI: Mô hình AI Trung Quốc có quan điểm chính trị phức tạp hơn dự đoán, không chỉ là kênh truyền đạt của chính phủ - Tin tức AI Aioga","description":"Vals AI dùng hàng trăm câu hỏi liên quan Trung Quốc để thử nghiệm các mô hình trong và ngoài nước, phát hiện mô hình Trung Quốc có xu hướng 'thân chính phủ' tổng thể thấp hơn dự đo...","url":"https://www.aioga.com/vi/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:03:24.364Z"},"id":{"title":"Riset Vals AI: sikap politik model AI Tiongkok lebih kompleks daripada perkiraan, bukan sekadar corong pemerintah","summary":"Vals AI menguji model-model dari Tiongkok dan luar negeri dengan ratusan pertanyaan tentang Tiongkok, dan menemukan bahwa \"kecenderungan pro-pemerintah\" model Tiongkok secara keseluruhan lebih rendah daripada perkiraan, sementara perbedaan antar model terlihat jelas. GLM-5.3 Flash bahkan secara tegas mengakui bahwa isu hak asasi manusia di Tiongkok didukung oleh bukti yang dapat diverifikasi. Bahasa prompt memiliki pengaruh yang signifikan: pertanyaan dalam bahasa Tionghoa sederhana lebih pro-pemerintah dibandingkan pertanyaan dalam bahasa Inggris, Rusia, atau Spanyol; hal ini berlaku pada model Tiongkok maupun Amerika Serikat, tetapi lebih terlihat pada model Tiongkok. Sebagian besar model Tiongkok jarang menolak menjawab dan justru langsung mengadopsi posisi resmi, seperti Kimi K3 yang menyangkal adanya kamp interniran di Xinjiang.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Riset Vals AI: sikap politik model AI Tiongkok lebih kompleks daripada perkiraan, bukan sekadar corong pemerintah - Berita AI Aioga","description":"Vals AI menguji model-model dari Tiongkok dan luar negeri dengan ratusan pertanyaan tentang Tiongkok, dan menemukan bahwa \"kecenderungan pro-pemerintah\" model Tiongkok secara kesel...","url":"https://www.aioga.com/id/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:03:09.663Z"},"th":{"title":"งานวิจัย AI ของ Vals: ท่าทีทางการเมืองของโมเดล AI ของจีนซับซ้อนกว่าที่คาดไว้ และไม่ใช่แค่กระบอกเสียงของรัฐบาล","summary":"AI ของ Vals ทดสอบโมเดลจีนและต่างประเทศด้วยคําถามเกี่ยวกับจีนหลายร้อยข้อ และพบว่า \"แนวโน้มสนับสนุนรัฐบาล\" ของโมเดลจีนโดยทั่วไปต่ํากว่าที่คาดไว้ โดยมีความแตกต่างอย่างมีนัยสําคัญระหว่างโมเดล GLM-5.3 Flash ยังยอมรับอย่างชัดเจนว่าประเด็นสิทธิมนุษยชนของจีนมีการบันทึกไว้อย่างดีภาษาที่ถามคําถามแบบตรงข้ามมีผลกระทบอย่างมาก: คําถามแบบตัวย่อของจีนเป็นมิตรกับรัฐบาลมากกว่าภาษาอังกฤษ รัสเซีย และสเปน; ซึ่งเป็นจริงทั้งกับโมเดลจีนและอเมริกัน แต่ชัดเจนกว่าสําหรับโมเดลจีนโมเดลจีนส่วนใหญ่แทบไม่ปฏิเสธที่จะตอบ แต่เลือกใช้จุดยืนอย่างเป็นทางการ เช่น Kimi K3 ปฏิเสธการมีอยู่ของค่ายกักกันในซินเจียง","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"งานวิจัย AI ของ Vals: ท่าทีทางการเมืองของโมเดล AI ของจีนซับซ้อนกว่าที่คาดไว้ และไม่ใช่แค่กระบอกเสียงของรัฐบาล - ข่าว AI Aioga","description":"AI ของ Vals ทดสอบโมเดลจีนและต่างประเทศด้วยคําถามเกี่ยวกับจีนหลายร้อยข้อ และพบว่า \"แนวโน้มสนับสนุนรัฐบาล\" ของโมเดลจีนโดยทั่วไปต่ํากว่าที่คาดไว้ โดยมีความแตกต่างอย่างมีนัยสําคัญระหว่...","url":"https://www.aioga.com/th/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:03:50.713Z"},"pl":{"title":"Vals AI Research: Polityczna postawa chińskiego modelu AI jest bardziej złożona niż oczekiwano i nie jest tylko rządowym tubą","summary":"Vals AI testowało chińskie i zagraniczne modele za pomocą setek pytań dotyczących Chin i stwierdziło, że \"prorządowa tendencja\" chińskiego modelu była ogólnie niższa niż oczekiwano, z istotnymi różnicami między modelami. GLM-5.3 Flash nawet wyraźnie przyznał, że problemy z prawami człowieka w Chinach są dobrze udokumentowane. Język promptów ma znaczący wpływ: chińskie pytania uproszczone są bardziej przyjazne rządowi niż angielskie, rosyjskie i hiszpańskie; dotyczy to zarówno modeli chińskich, jak i amerykańskich, ale bardziej widocznych w modelu chińskim. Większość chińskich modeli rzadko odmawia odpowiedzi i zamiast tego przyjmuje oficjalne stanowiska, takie jak Kimi K3 zaprzeczający istnieniu obozów koncentracyjnych w Xinjiangu.","category":"行业动态","source":"The Decoder：AI News","aggregationSource":"The Decoder：AI News","pageTitle":"Vals AI Research: Polityczna postawa chińskiego modelu AI jest bardziej złożona niż oczekiwano i nie jest tylko rządowym tubą - Aioga Wiadomości AI","description":"Vals AI testowało chińskie i zagraniczne modele za pomocą setek pytań dotyczących Chin i stwierdziło, że \"prorządowa tendencja\" chińskiego modelu była ogólnie niższa niż oczekiwano...","url":"https://www.aioga.com/pl/news/xegqw6h2wye28k01yhv179qfb/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"75bb6a1cb4f0d2ca","translatedAt":"2026-11-01T14:03:50.822Z"}},"evidenceTier":"verified-news","reviewStatus":"automated-ingest","indexable":true,"editorialCover":""}}