Peter Hull used AI to replicate 1,346 staggered-adoption tests from 108 economics papers. Two-way fixed effects estimates were highly correlated with modern difference-in-differences methods. Sign reversals occurred in 10.8–14.5% of comparisons. TWFE showed greater precision. Hierarchical models implied a 91–98% correlation between estimands. The probability of a statistically significant TWFE estimate with the opposite sign was 0.6–2.1%. The findings suggest negative weights may be less serious in practice. The work demonstrates AI’s role in large-scale empirical assessments of econometric methods.

• 1,346 staggered-adoption specifications
• 108 economics papers
• 10.8–14.5% sign reversals
• 91–98% correlation between estimands