斯坦福大学研究人员在2026年2月8日日本下院大选期间进行的实验发现，来自OpenAI、Google和xAI的5个主流AI模型在回答选民投票建议时，对持左倾政策立场的选民压倒性地推荐日本共产党1。研究团队对36,300个合成选民档案进行了测试，结果表明政策立场因素对模型推荐的影响程度达到50至98个百分点，远高于人口统计因素的0.5至7个百分点1。
这一现象的根本原因在于AI模型可获取的信息环境差异1。由于日本主流新闻媒体因版权考虑对爬虫设置了访问限制，而日本共产党及其党报《赤旗》网站完全开放，AI爬虫获取的政治信息主要来自共产党的开放网站和宣传内容1。研究人员表示，AI模型将这些党派宣传内容错误地当作中立新闻信息处理，从而导致了对特定政党的推荐偏差1。该研究由斯坦福商学院的Andrew Hall和早稻田大学访问研究员Sho Miyazaki主持完成1。
Researchers at Stanford University conducted an experiment during Japan's February 8, 2026 lower house election and discovered that five major artificial intelligence models overwhelmingly recommended the Japan Communist Party when providing voting guidance to left-leaning voters.1 The study, which tested models from OpenAI, Google, and xAI across 36,300 synthetic voter profiles, revealed that policy positions exerted a far greater influence on recommendations than demographic factors, with left-leaning policy stances driving all five models to suggest the Communist Party by margins ranging from 50 to 98 percentage points, compared to just 0.5 to 7 percentage points for demographic characteristics.1
The researchers, including Andrew Hall from Stanford Graduate School of Business and Sho Miyazaki, a visiting researcher at Waseda University, attributed this pattern to the information environment accessible to AI systems rather than programmed bias.1 Mainstream Japanese news media outlets have implemented robots.txt restrictions that prevent AI crawlers from accessing their content due to copyright concerns, whereas the Japan Communist Party's website and its Akahata newspaper remain fully open to indexing.1 As a result, AI models treating party promotional materials as neutral news sources generated skewed recommendations that aligned Communist Party positions with the policy preferences expressed in voter profiles.1
评论
还没有评论，欢迎留下第一条。