Detecting AI Coding Tool Adoption and Its Behavioural Effects on Developer Commit Activity

We study the effect of AI coding tool adoption on developer commit behaviour using two complementary empirical designs. First, we build a behavioural classifier that identifies AI coding tool users from observable commit history — without relying on explicit self-reported adoption or proprietary telemetry. The classifier achieves cross-validated AUC of 0.94 on a sample of 276 GitHub accounts (74 confirmed adopters, 202 controls) and generalises to users of a second tool (Aider, mean predicted probability 0.73) it was never trained on, suggesting it detects general AI-assisted coding behaviour rather than tool-specific stylistic artefacts. Second, we use the classifier in two causal designs.

An account-level difference-in-differences finds large, statistically significant changes in commit behaviour for AI adopters relative to controls, including substantial increases in commits per active week and reductions in inter-commit hours. A country-level panel regression across 34 countries (2022–2024) finds divergent results across dependent variables: a robust negative association between adoption and commits per developer (weighted coefficient = −7.56, p = 0.05 across 34 countries) coexists with a precisely-estimated null on pull requests per developer (coefficient = +1.33, p = 0.76). We interpret this DV split cautiously: it is consistent with AI tools shifting commit granularity (fewer, larger commits) rather than reducing genuine output, but our data cannot rule out alternative explanations. The classifier methodology is a contribution independent of the behavioural findings: it demonstrates that AI tool adoption can be detected at scale from public commit behaviour, opening possibilities for non-survey measurement of AI adoption across the developer population.