A year of AI disclosure in critical packages | Andrew Nesbitt
Stephen O’Grady’s RedMonk analysis of who is writing open source code looked at commits to fifteen large projects during the first half of 2026 and counted two forms of declared AI involvement: a known autonomous agent as the commit author, or a known AI identity in a Co-Authored-By trailer. The result was under one percent, framed as a floor.
I ran a wider version of the same measurement over the packages.ecosyste.ms critical set: 5,682 GitHub repositories behind the most-depended-on packages across sixteen registries, using the CHAOSS disclosure library to detect four kinds of explicit signal instead of two. Over the same six months the rate was 4.13%. Over the year ending 29 July 2026 it was 2.93% (17,279 of 589,798 non-merge commits), rising from 0.48% last August to 5.32% this July.
These are counts of commits where someone left an explicit marker in git metadata. Undeclared use is not measured, and a commit is one unit regardless of whether it changed one line or ten thousand.
Sample selection versus detector choice
Running my scanner against RedMonk’s fifteen repositories with only their two signals found 94 matches in 23,346 first-half commits including merges, or 0.40%, against RedMonk’s “~24K commits” and a match count “in the dozens”. Excluding merges leaves 17,323 commits and the same 94 matches, or 0.54%; espressif/esp-idf and openssl/openssl supply 71 of them, matching RedMonk’s reported 73% concentration in two projects.
sample and signals<br>commits<br>marked<br>share
RedMonk 15, agent author or known AI co-author<br>17,323<br>94<br>0.54%
RedMonk 15, all validated disclosure signals<br>17,323<br>182<br>1.05%
Critical GitHub set, agent author or known AI co-author<br>308,354<br>11,002<br>3.57%
Critical GitHub set, all validated disclosure signals<br>308,354<br>12,720<br>4.13%
Adding the two extra signal types moved the rate by about half a percentage point on either sample. Changing the sample moved it by three points. RedMonk’s fifteen were chosen by contributor-base size with, in O’Grady’s words, a deliberate bias towards C; the critical package set is whatever sits at the top of each registry’s dependency graph, which pulls in a lot of smaller, newer, company-run repositories.
What I counted
The critical snapshot contained 8,605 packages, with repository URLs and metadata pulled from the same package cache I built for Weekend at Bernie’s. Merging packages that share a repository, following renames, restricting to GitHub, and dropping malformed URLs left 5,707 candidates. 5,682 cloned successfully; the other 25 were deleted or private. 3,533 had at least one non-merge commit in the year ending 29 July 2026.
Each repository was cloned bare with a tree filter and a shallow date boundary, streamed through the disclosure library, and deleted. The full pass transferred about 1 GB and the retained checkpoint is 16 MB of per-repository summaries and matched commit SHAs.
Rename following checks GitHub’s stable repository ID as well as the redirect. The npm package base still lists node-base/base as its repository. GitHub reused the org name, so that path now redirects to the Base blockchain monorepo, which would have contributed 3,135 commits and 273 AI signals to a nine-year-old npm utility. The ID check excluded it.
Every non-merge commit was checked for:
a known AI agent as author or committer
a known AI identity in Co-Authored-By
an Assisted-By trailer naming an AI tool or model
a tool-specific attribution format that disclosure supports
Merges are excluded so projects that squash, rebase, or merge count on the same basis. Commits are bucketed by committer time, when the change landed on the current branch. Mentions of tool names in ordinary commit prose are ignored. Assisted-By values are validated because the trailer is also used for people: raw matches included Assisted-By: Daniel Stenberg and Assisted-By: Automated Tooling, Human Reviewed. The clones did not fetch refs/notes/ai, so declarations recorded as git notes are absent.
Over the year
The monthly rate passed 3% in February and 5% in March, then held between 4.58% and 5.32% through July. Counting repositories instead of commits, a signal appeared in 41 of the 1,734 repositories with commits in August 2025 (2.4%) and 276 of 1,793 in July 2026 (15.4%).
Of the 17,279 findings, 4,625 carry only an autonomous-agent identity, 12,628 carry only a declared-assistance signal, and 26 carry both. Declared assistance went from 0.08% of commits in August to 4.92% in July. Agent authorship started at 0.40% and ended at 0.41%, peaking at 1.33% in between; 4,613 of those commits have GitHub Copilot’s agent as author, 38 have Devin’s, and Claude, Cursor, Codex, and Amazon Q account for 25 between them. Copilot agent commits reached 745 in March across 85 repositories and fell to 208 across 35 in July, with individual projects running the agent in short bursts: pycqa/isort had 49 in March and none after,...