AI is already in one out of every twenty commits—but this isn't mass automation

The study’s main finding doesn’t make for pretty charts: the growth of AI in open source is noticeable, but it’s distributed extremely unevenly. Over the course of a year, 2.93% of the 589,798 non-merge commits in critical repositories contained an explicit AI signal, and in July, that figure reached 5.32%.

This does not mean that one in every twenty commits was written by a neural network. It refers only to cases where a record of AI involvement remained in the Git metadata. The study does not account for undeclared use.

The most important figure is not 4.13%, but the difference between the samples. On the same 15 projects used by RedMonk, the narrow rules yielded 0.54% of commits. In a sample of 5,682 repositories of critical packages, the figure rose to 3.57%. When expanded feature types were added, the share increased only to 1.05% and 4.13%, respectively.

In other words, the choice of what to consider “critical” and which projects to include affects the result about six times more than the choice of detector. Therefore, the headline “AI writes 4% of open source” is too simplistic. It would be more accurate to say: in a broader and more diverse sample, an increasing number of projects are voluntarily leaving traces of AI assistance.

Moreover, this is not yet a uniform shift across the industry. Signals were detected in only 687 out of 3,533 active repositories, and the median repository had no such commits at all. Ten projects accounted for 40.8% of all findings, and 100 projects accounted for 84.9%.

The main growth did not come from autonomous agents. Their share remained at approximately 0.4% of commits at the beginning and end of the period. In July, 4.92% of commits had a signal of declared assistance, and the sharp spike in the spring was mainly associated with Claude Code and its “Co-Authored-By” tags. Of the 17,392 recorded mentions of tools, 57.35% were for Claude Code, and 27.93% were for GitHub Copilot.

There’s a practical lesson here for teams: metadata can be more useful than marketing claims. It reveals where AI is actually integrated into the process, which tools are being used, and to what extent the practice is concentrated among a few teams. But to do this, you need to agree in advance on what exactly counts as disclosure; otherwise, comparing projects will be like comparing different measuring devices.

There are also limitations. The scanner only took into account explicit signals in Git: the author-agent, AI in Co-Authored-By, Assisted-By, and special attribution formats. Ordinary mentions of tools in commit messages were ignored, and Git notes containing declarations were not loaded. A single commit was counted as one unit regardless of the size of the change. Furthermore, the results are skewed by the composition of the sample: for example, 95.6% of the signals in NuGet came from Microsoft repositories; after excluding these owners, the ecosystem’s share dropped from 6.84% to 0.88%. Finally, the presence of AGENTS.md or CLAUDE.md indicates that instructions for agents have been configured, but proves nothing about the authorship of specific changes.

Therefore, it makes more sense now to measure not the “percentage of code written by AI,” but rather the maturity of AI usage disclosure within individual teams and projects. This is a more mundane conclusion, but it stands up to data scrutiny.

If AI is already helping to write code in your project, would you leave such a mark in Git—for auditing purposes, to build trust, or simply because the statistics don’t show anything otherwise? Source: Hacker News - Newest: ""AI" "LLM""