Claim graphs show most errors in economic theory papers are repairable
Ash, Ciccarone and Galletta use an AI agent to map 190 theory papers as graphs from assumptions to theorems, and find that the smallest valid repair usually preserves the main conclusions.
Source published · In the Daily Brief of 23 October 2026
Filed under Reasoning and inference, Economy, Agents
Researchers at ETH Zurich and Sapienza University of Rome built an agentic system that represents each economic theory paper as a claim graph, in which model assumptions and intermediate lemmas support the headline results. In a random sample of 190 papers published since 1995 in four leading journals, removing the most consequential starting input leaves on average 91% of weighted headline and major claims unsupported.
Benchmarked against published corrections to 100 papers, the system fully recovered 79 of 168 documented errors, or 47%, and 69% when partial matches count. The graphs then measure what a correction costs: among corrected papers the median share of conclusions an error touches is 28%, but the median loss of economic content after the smallest valid repair is 4%, and 2% in the random sample.
The paper is a draft dated 22 October and has not been peer reviewed. Researchers on X discussed it as a way to rank AI-detected errors by how much they matter rather than by how many there are.