AI-led economics audit separates asserted defects from contested choices
An anonymous project revised a model-led review of 138 economics works, separating asserted defects from contested methodological choices.
TL;DR
- An anonymous AI-led review of 138 works by Emmanuel Saez, Gabriel Zucman and coauthors revised its findings on October 29.
- The revision divides 147 major findings into 53 alleged defects and 94 contested methodological choices.
- The project says no finding overturns a core result outright, and its findings have not been reviewed by a domain expert.
An anonymous AI-led review of 138 works by Emmanuel Saez, Gabriel Zucman and coauthors revised its findings on October 29. Its new version divides 147 major findings into 53 alleged defects and 94 contested methodological choices. [1] [2]
The project says it used Claude Opus 5.5 and GPT-6 Astra to examine each work without seeing the other model’s output, with a third agent challenging and grading the results. Ground Truth reports that no domain expert reviewed the findings. [1] [2]
An earlier version listed 2,035 issues across the works and described 147 as serious enough to change a headline number, EconoTimes reported. The revision says none overturns a work’s core result outright, although it says some headline magnitudes may change. [2] [3]
The project says its instructions focused on finding flaws, which can affect both the volume and direction of findings. Ground Truth notes that the models’ agreement rate is not published and that the findings still need independent expert checks. [1] [2]
Why it matters
The review illustrates how AI can surface questions across a large research record while leaving the validity and severity of individual claims to be checked. The distinction matters when screening results enter public debates about economic policy.
Editor's note
The anonymous project’s categories and totals are attributed to the project, not treated as verified errors. The October 29 revision supersedes the original; independent expert review of the findings is not established.