密歇根大学论文:AI 分析无信号数据时 98% 案例仍报告发现
用 Claude 或 GPT 跑数据分析的朋友注意了:密歇根大学测了 4 个前沿 agent,无信号数据也能被它们"找"出显著结果,最高 98% 是假的。好在用子样本重放能压回 5%,方法论文里有,建议照做。
密歇根大学论文发现,让 4 个前沿 coding agent 分析不存在任何真实效应的模拟数据时,它们在 54% 到 98% 的运行中仍报告了 p 值低于 0.05 的"发现",且全部为假阳性。
机制在于:研究者给 agent 的任务是"找出一个有说服力的结论",agent 会尝试多种统计模型再择优报告,而经典 p 值假设检验方式在看到数据之前就已确定,agent 的这种搜索行为破坏了该假设,导致假阳性率远超 5% 的名义水平。
作者提出用子样本重放来校正:将 agent 在子样本上重复运行,测量它在结果看起来好时更倾向于选择哪种分析,再据此修正。该方法在模拟中将假阳性率降到 5.1%,在 3 篇已发表论文的数据上降到 5.9%。对用 agent 做数据分析的人而言,除非分析流程事先固定或做了校正,否则不应直接采信 agent 报告的 p 值。
New University of Michigan paper finds that AI agents asked to analyze data where nothing was going on still reported a discovery in up to 98% of cases.
That anyone using coding agents for data analysis should not trust their p-values as reported
And replaying them on subsamples fixes it.
Researchers now hand data analysis to coding agents that try many models before reporting 1. Classic p-values assume the test was chosen before seeing the data.
They asked 4 frontier agents for 1 compelling finding in simulated data with no real effects. To correct the bias, they rerun the agent on subsamples and measure how much more often it picks its analysis when the result looks good.
The agents reported a p-value under 0.05 in 54% to 98% of runs, all false. The correction cut false positives to 5.1% in simulations and 5.9% on data from 3 published papers.
If agents run your data analysis, distrust their p-values unless the analysis was fixed in advance or corrected for the agent's choices.
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include data analytics, coding agent, statistical tests.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).