A data-driven analysis measuring explicit AI involvement in 5,682 critical open source GitHub repositories across 16 package registries over the year ending July 2026. Using the CHAOSS disclosure library to detect four signal types (AI agent as author, Co-Authored-By, Assisted-By trailers, and tool-specific attribution), the study found 4.13% of commits in the first half of 2026 carried an AI disclosure marker — significantly higher than the ~1% found in a narrower RedMonk study of 15 large projects. The annual rate was 2.93%, rising from 0.48% in August 2025 to 5.32% in July 2026. Claude Code dominated declared tool usage at 57.35%, followed by GitHub Copilot at 27.93%. NuGet had the highest ecosystem rate (6.84%), largely driven by Microsoft-owned repositories. Additionally, 6.21% of scanned repositories contained committed instruction files for coding agents (AGENTS.md, CLAUDE.md, etc.). All data and scanner code are publicly available.

8m read timeFrom nesbitt.io
Post cover image
Table of contents
Sample selection versus detector choice #What I counted #Over the year #By ecosystem #Instruction files #Data #
8.6K Impressions