This weekend I dug into my database of open-source GitHub repositories to see how frequently newsrooms are leaving behind signals that they’re coding with an AI agent (a.k.a. “vibe coding”). Here’s what I found:
As you can see in the chart below, roughly a third of repos made by newsrooms in 2026 have at least one of the AI-coding signals I looked for. There are lots of reasons this number is just a floor (more on this later), and the real number of newsroom coders who are using AI coding agents is likely higher.
AI-Coding Signals in Newsroom Repositories Grew to a Third of All Repos in 16 Months
Share of newly-created public GitHub repositories from newsrooms with at least one AI-coding signal: a CLAUDE.md or AGENTS.md file; credit given to an AI-coding agent; or a coding-agent file explicitly excluded from the git repository.
Excludes empty repos but includes repos with little or no code, such as data releases. Each repo is counted in only one bucket: if it ships a CLAUDE.md or AGENTS.md it goes there; otherwise if recent commits credit a coding agent; otherwise if it gitignores one. The full bar is the share of repos that month showing any AI signal. Source: openjournalism.news scan of ~360 newsroom GitHub accounts.
Things got going in Sept. 2025, when we started seeing AI-coding signals appear more often. Signs grew louder starting in November, then a real jump came in February when these signals became much more prevalent and consistent. I’m willing to bet this is the release of Anthropic’s models Opus 4.5 (Nov. 24) and 4.6 (Feb. 5), along with competitive models from OpenAI and Google, which represented leaps in coding quality and tool maturity.
Without talking to newsrooms I can only see what’s left behind in open-source repos, intentionally or incidentally. The AI-coding signals I looked for included:
- A CLAUDE.md file — Anthropic’s convention — or AGENTS.md, the standard jointly backed by Google, OpenAI, Cursor, and others. Both are project documentation files that developers write (or, more commonly, have their AI agents write) and the agent reads at the start of every session to learn the codebase’s conventions and architectural decisions.
- Contribution credit that an AI agent gave itself (with the author’s permission).
- AI-agent artifacts explicitly excluded from the public repo using a
.gitignorefile.
CLAUDE.md files are written in plain English, though they can and usually do include code snippets and deeply technical information. Because they’re meant to be used by an AI agent to pick a project back up between sessions, they tend to be very useful documents even for humans who need to familiarize themselves (or their AI agents) with the codebase so they can use or extend it without breaking it.
I find CLAUDE.md and AGENTS.md files particularly interesting. They have the potential to make open-source releases far more useful. Well-written documentation is rarely a top priority for newsroom coders. In fact one of the reasons newsrooms cited for open-sourcing less often was the time burden required by, among other things, writing good documentation. Sharing the documentation the project naturally produces anyway makes that problem much smaller.
Some examples of these files:
- OpenSanctions’ followthemoney, its shared data model for sanctions data, used in many investigative newsrooms.
- ProPublica’s CLAUDE.md for its Google Sheets add-on for document-processing tasks.
- CalMatters’ AGENTS.md for its Prop 50 voter-analysis notebook.
Finding AI-coding signals in a third of new repos is not the same as saying only a third of newsrooms are vibe coding. The number is certainly higher. Two caveats in my analysis:
Although I excluded empty repositories (there are a surprising number of these), I wasn’t able to exclude data-only repos. Detecting these is a nontrivial problem as repos meant to share data often include a Jupyter Notebook or the code used to process the data. And while data-sharing repos don’t typically include AI-coding signals, there’s no reason they can’t now or in the future, so I decided it was best to be broad in what I included.
Also, some AI agents and developer tools leave no artifacts at all, and those I can’t catch. For agents that do leave artifacts, I checked both the committed files in the repo and inside the .gitignore file, where developers list files they want to keep on their own machine but not share publicly.
Finally, of course, there’s all the code that newsrooms write that is never open-sourced.
For context, a 2025 survey by Stack Overflow put the share of developers using or planning to use AI coding tools at 84%, with 73% of engineering teams reporting daily use.
GitHub is straining to keep up. Its COO said in April that the platform is now processing 275 million commits a week, putting it on pace for 14 billion this year — up from 1 billion in all of 2025. AI agents now drive more than one in five code reviews on the platform, and according to The Information, pull requests opened by agents grew from about 4 million in September 2025 to over 17 million by March 2026 — on roughly the same timeline my research shows newsroom signals taking off.