《经济学人》对25个前沿人工智能模型进行的一项调查显示，它们所嵌入的世界观和价值观与大多数人类存在显著差异。研究人员利用自1981年以来一直追踪100多个国家信仰的世界价值观调查（WVS）绘制了模型的回答。结果显示，像OpenAI的GPT这样的西方模型高度世俗化，而谷歌的Gemini对个人自由的强调超过了任何国家的人口平均水平，且没有任何前沿模型能够反映大多数非洲或穆斯林国家的价值观。为了解决这一问题，像瓦利德·卡杜斯开发的伊斯兰聊天机器人“Ansari”等专业工具应运而生，而其他像Talkie这样的模型则完全基于1931年之前的文本进行训练以捕获历史社会习俗。
训练数据和训练后的对齐极大地塑造了这些价值观。针对38种语言中GPT-3.5的研究表明，在压迫性国家的语言中，AI的输出反映了当地的国家媒体控制（根据世界新闻自由指数衡量）。对于西方AI而言，获取中文语言技能需要对经过审查的中国互联网文本进行训练，导致它们反刍国家叙事。此外，诸如人类反馈对齐等训练后方法引入了开发者的政治价值观。虽然谷歌的Gemini在2024年因在历史图像生成中过度纠正“多样性”而面临强烈反对，但xAI的Grok则朝着相反的方向对齐，而像DeepSeek和千问（Qwen）这样的中国模型则被严格授权维护社会主义核心价值观。
这些偏见的地理政治和社会影响正在迅速扩大。微软报告称，今年第一季度全球18%的劳动年龄人口（近十亿人）使用了生成式人工智能。使用美国VOTER调查问题的测试表明，除DeepSeek V3.2外，几乎所有模型在社会和经济问题上都与民主党一致。尽管存在这些偏见，中国的开源权重模型因其成本效益和透明度而非常受欢迎，其中阿里巴巴的千问模型到1月在Hugging Face上的下载量已超过7亿次。爱沙尼亚对外情报局警告称，中国模型可能会传播扭曲的世界观，但它们较低的运行成本使其对发展中国家具有吸引力，突显了AI影响公众舆论的日益增长的力量。
A study by The Economist investigating 25 frontier AI models reveals that their embedded worldviews and values diverge significantly from those of most humans. Using the World Values Survey (WVS), which has tracked beliefs in over 100 countries since 1981, researchers mapped models' responses. The results show that Western models like OpenAI’s GPT are highly secular, while Google’s Gemini places more emphasis on individual freedom than average populations in any country, with no frontier model reflecting the values of most African or Muslim nations. To address this, specialized tools like Waleed Kadous's Islamic chatbot "Ansari" have been built, while other models like Talkie are trained exclusively on pre-1931 texts to capture historical social mores.
Training data and post-training alignment heavily shape these values. Research analyzing GPT-3.5 across 38 languages shows that AI outputs in languages of repressive countries mirror local state control of the media, as measured by the World Press Freedom Index. For Western AIs, acquiring Chinese language skills requires training on censored Chinese internet text, causing them to regurgitate state narratives. Additionally, post-training methods like human-feedback alignment introduce developers' political values. While Google’s Gemini faced backlash in 2024 for overcorrecting toward "diversity" in historical image generation, xAI’s Grok was aligned in the opposite direction, and Chinese models like DeepSeek and Qwen are strictly mandated to uphold socialist values.
The geopolitical and societal influence of these biases is expanding rapidly. Microsoft reports that 18% of the global working-age population—nearly one billion people—used generative AI in the first quarter of the year. Testing using American VOTER Survey questions shows that almost all models, except DeepSeek V3.2, align with the Democratic Party on social and economic issues. Despite these biases, Chinese open-weight models are highly popular due to their cost-efficiency and transparency, with Alibaba's Qwen surpassing 700 million downloads on Hugging Face by January. Estonia's Foreign Intelligence Service warned of Chinese models spreading distorted worldviews, yet their lower operating costs make them attractive in developing nations, highlighting the growing power of AI to influence public opinion.
Source: AI models’ values are very different from most people’s
Subtitle: They are more secular and more liberal—unless they’re made in China
Dateline: 6月 25, 2026 04:38 上午