AI audit of 138 economics papers sparks debate over publishing machine-generated criticism without human review

AI audit of 138 economics papers sparks debate over publishing machine-generated criticism without human review

An anonymous project audited Emmanuel Saez and Gabriel Zucman’s research on inequality and taxation. On October 30, Abhi Mahajan, who blogs about biology and machine learning, condemned the circulation of machine-generated criticism without carefully checking each point.

On October 28, economist Tyler Cowen introduced the audit on the Marginal Revolution blog: it raised 2035 concerns, including 147 serious enough to substantially change a paper’s main numerical result. Cowen called for a fundamental reassessment of this research and the policy arguments built on it. Saez and Zucman are recipients of the Clark Medal, an American award recognizing contributions to economics by economists under forty.

According to the project’s description, two AI systems independently checked numbers, quotations, and assumptions against the papers and calculation files. A third reviewed the concerns and rejected those that did not hold up. The detailed reports also retain the rejected criticisms.

The October 29 version divides the 147 major concerns into two categories: the auditor classified 53 as errors and 94 as defensible but debatable choices.

One of the moderate concerns involves who bears the corporate tax burden: one model assigns it entirely to shareholders, while another assigns some of it to workers and other investors. The choice of model affects estimates of the share of income that the wealthiest people pay in taxes. The auditor classified another moderate concern as an arithmetic error: when switching to a different model of how the tax burden is distributed, part of a formula used income calculated under the previous model. Economist Jason Furman agreed with some of the criticism as early as October 29: in his assessment, the papers’ authors had relied on weak assumptions in several cases.

In the original version, 97% of major and moderate concerns whose effects could be determined would, if accepted, weaken the paper’s central claim. The current version attributes this imbalance to the procedure itself: the AI was instructed to look for flaws, and concerns were rated for severity according to how much they undermined a paper’s central claim. According to the project, applying the same procedure to two sets of papers by other economists also produced concerns that mostly weakened those papers’ central claims.

“What is wrecking our shared information environment supposed to achieve? A flashy takedown?” Mahajan asks.

Aging researcher Kamil Pabis responded to Mahajan.

“For AI to be useful in the future, it needs to produce acceptable results without requiring people to check every single thing manually.”

He proposes finding an automated way to assess AI reliability.