一项针对ChatGPT、Perplexity、Gemini、Claude和Google AI Overviews这五款主流AI助手的研究显示，仅有40%由AI引用的统计数据能够追溯到声称支持该数据的原始研究 1。此次调查共追踪了1,285个被上述工具引用的统计数据点 1。在未能成功溯源的案例中，有40%的数据停留在无法检查的来源，另有11%没有任何指向性链接，而8%虽能找到相关研究，但数值或语境已发生改变 1。
从来源偏好来看，AI倾向于引用公司博客或统计汇总页面，这类内容占比超过一半，而直接指向研究本身的被引用页面仅占21% 1。研究发现，即便在成功找到原始研究的情况下，仍有17%的数据出现了变形，其中26%表现为数值变化，20%则涉及测量对象的改变 1。值得注意的是，五个AI助手在某些问题上给出了完全相同的“僵尸数据”，即无实际研究支持的数字，这表明多个AI输出的一致性并不能作为数据准确性的证明，反而可能源于共同的错误源头 1。基于此，建议用户将AI提供的统计信息视为线索并手动追溯至原始来源，同时呼吁内容发布者公开研究方法以便AI进行正确索引 1。
A study analyzing statistical citations from five major AI assistants—ChatGPT, Perplexity, Gemini, Claude, and Google’s AI Overviews—found that only 40% of cited statistics could be traced back to the research claimed to support them 1. The investigation tracked 1,285 AI-referenced statistics in total 1. While 8% of citations led to an identifiable study, the numerical values or context were altered in those cases 1. Additionally, 40% of citations pointed to unverifiable sources, and 11% provided no reference at all 1.
The analysis revealed a bias in source selection, noting that only 21% of cited pages were actual research papers, while more than half directed users to corporate blogs or statistical aggregations 1. Among citations where the original study was successfully located, 17% contained data discrepancies; specifically, 26% involved changes in numerical values and 20% reflected shifts in the measured subject matter 1. Researchers also observed instances of "zombie data," where all five assistants produced identical figures unsupported by any underlying study, indicating that consensus among AI models does not guarantee accuracy 1. The study advises users to manually follow citation chains to locate primary sources rather than relying on agreement between different AI tools for verification 1.
评论
还没有评论，欢迎留下第一条。