Generative AI adoption and scientific output: Quasi-experimental evidence on heterogeneous changes across 26 disciplines

Using monthly OpenAlex data on 92 million publications across 26 academic fields (2015–2025), we examine publication growth associated with large language model (LLM) adoption through a quasi-experimental difference-in-differences design around ChatGPT’s release. The pooled estimate is statistically indistinguishable from zero. However, fields with one standard deviation higher pre-existing AI intensity show 4.52 percentage points greater output growth per percentage point of measured adoption ( 𝑝 = 0 . 0 1 0 ). Expanded keyword definitions reduce this estimate to 1.21–1.75 percentage points. Computer Science shows an estimated cumulative output increase of 8.7% by late 2025. These findings indicate heterogeneous field-level publication responses, with magnitudes sensitive to how adoption is measured.