Replication Package: Causal Effects of Generative AI Adoption on Scientific Output — Quasi-Experimental Evidence of Heterogeneous Impacts Across 26 Disciplines

This repository contains the code and processed data needed to reproduce the results in "Causal Effects of Generative AI Adoption on Scientific Output: Quasi-Experimental Evidence of Heterogeneous Impacts Across 26 Disciplines". The paper applies a difference-in-differences design to the ChatGPT natural experiment, using monthly OpenAlex publication data on 92 million works across 26 academic fields (2015–2025) to estimate the causal effect of large language model (LLM) adoption on scientific output and its heterogeneity across disciplines. A pooled specification yields an effect indistinguishable from zero, but fields with higher pre-existing AI capabilities show significantly greater output growth per unit of LLM adoption, consistent with a threshold pattern in which Computer Science is the only field above the estimated benefit threshold. Contents:- code/ — numbered analysis pipeline (data download, panel construction, main regressions, robustness checks, figures), plus three additional modules addressing specific reviewer requests: keyword-list sensitivity (keyword_expansion/), wild cluster bootstrap confidence intervals for the heterogeneity threshold (bootstrap_ci/), and a 260-paper manual classification exercise validating the LLM-mention measure (paper_validation/)- data/processed/ — panel datasets and intermediate outputs needed to run every downstream step without re-querying the OpenAlex API- output/ — regression results, robustness tables, and validation statistics, each matching the corresponding number reported in the manuscript- figures/ — all manuscript figures, regenerated from the included code- README.md — full script-by-script documentation, run instructions, expected output values, and verification status All underlying publication data are from OpenAlex (https://openalex.org/, CC0 license). Requirements and run instructions are in README.md.