Meta独立监督委员会的研究发现，ChatGPT、Claude和Gemini等美国AI模型在处理涉及中国的敏感政治提问时，经常拒绝批评威权政府或重复中国官方立场，呈现出类似中国式审查的效果[1]。研究表明，OpenAI、Anthropic、Google和Meta开发的模型批评专制政府的可能性明显低于批评自由度较高国家政府的可能性[1]。其中，Anthropic的Claude Sonnet 4在面对针对中国领导人习近平的批评请求时，以安全考量为由拒绝回应[1]。
这种现象部分源于AI模型的训练方式包含国家控制媒体数据，以及安全功能的设计考量[1]。研究人员指出有两个主要原因导致这一问题：安全功能旨在保护中国等地区用户；AI实验室对这类区别对待问题的重视程度不足[1]。研究人员发现，AI公司只需进行简单调整，如提高信息来源透明度，就能解决相关问题，但这些公司并未采取多少实质措施[1]。相关公司仅发表声明称已采取措施确保回答客观中立，但在与多家美国AI实验室员工会面后，研究人员对这些公司能否做出改变持怀疑态度[1]。
Meta's Independent Oversight Board has found that major American AI models including ChatGPT, Claude, and Gemini frequently refuse to criticize authoritarian governments or echo official Chinese positions when handling sensitive political questions about China, effectively replicating Chinese-style censorship.[1] The research reveals that AI systems developed by OpenAI, Anthropic, Google, and Meta are significantly less likely to criticize authoritarian governments compared to their treatment of governments in countries with greater freedoms.[1]
The phenomenon stems from multiple factors, including the incorporation of state-controlled media data in model training and safety function considerations designed to protect users in regions like China.[1] Notably, Anthropic's Claude Sonnet 4 rejected requests to criticize Chinese leader Xi Jinping, citing safety concerns.[1]
Researchers identified that simple adjustments—such as increasing transparency in information sources—could resolve many of these issues, yet the companies have taken limited action.[1] The companies have issued statements claiming they have adopted measures to ensure objective and neutral responses, but their demonstrated commitment to addressing the problem remains questionable.[1] After meeting with staff from multiple U.S. AI laboratories, observers expressed skepticism about whether these firms would implement meaningful changes.[1]