Ban it or get forked: how open source is handling the flood of AI agents

What this is about

On October 28, GitHub published its annual Octoverse 2026 report. It contains a number that has open source maintainers, the mostly volunteer stewards of free software, sitting up: in September 2026, more than 40 percent of all pull requests started as a draft from a coding agent. A pull request (PR) is a proposed change someone submits to a project. In March, the share was still 17 percent. Around the same time, the website osssaysno.com has been collecting projects that ban AI contributions. It currently lists 22, including Immich, QEMU, Zig, Servo and Gentoo.

Our take: bans are well-intentioned, but they hit the wrong target. Agents barely read them and don't follow them. And if you say a hard no, you risk a fork, a split-off copy of the project that carries on without you. We went through studies, policies and our own sample to show which models are winning out and what they cost.

The number behind the debate, and what it measures

According to GitHub, agents accounted for about 110 million PRs in September, roughly six times as many as in March. About 90 percent of agent PRs were merged within two weeks (83 percent in February). For PRs with no detected agent involvement, the rate held at 86 percent. The caveats GitHub itself lists matter. It only counts PRs whose first draft came from an identifiable agent, so the figure is a lower bound. Private repositories are included, and the report doesn't explain why the merge rate is rising.

Our take: the 40 percent mostly reflects developers using agents in their own or their company's projects. It doesn't mean that 40 percent of contributions submitted by outsiders to open source projects come from bots. The high merge rate points more toward teams merging their own agent drafts. For open source projects, the number still means one thing: the stream of machine-started contributions isn't shrinking.

Agents don't read the house rules

How little bans get through to agents is shown in the study “A First Look at Coding Agents' Compliance with AI Contribution Rules” by Wenhao Yang, Runzhi He and Minghui Zhou (Peking University, arXiv). The researchers catalogued 455 AI rules from 102 communities. From those, they built a benchmark of 106 tasks from 49 repositories, called RepoComplianceBench. They tested OpenCode with DeepSeek-V4-Pro, Codex with GPT-5.3-Codex and GPT-5.5, and Claude Code with Sonnet 4.6.

Unprompted, the agents opened the policy file in only 3.5 percent of runs, or 12 out of 347. Even when the ban was spelled out to them word for word, three of the four agents still never refused the task. GPT-5.5 kept its contribution in all 30 corrected cases. The authors conclude that rules requiring an extra step can be implemented with agents. Bans, on the other hand, have to be enforced outside the agent, for example through CI checks (automated tests that run on every submission), mandatory reviews or bots.

Three models and what they cost

A total ban is the exception. Andre Hora, Romain Robbes and Stefano Zacchiroli analyzed 281 policies from the most popular GitHub projects (arXiv). Of those, 83.3 percent allow AI for code and 14.9 percent ban it. 48.8 percent require disclosure. 43.4 percent explicitly place responsibility on the human. 84 of the 92 standalone policy files were only created in 2026.

1. Ban. Gentoo has banned any content created with the help of language AI since April 2024, citing copyright, quality and ethics. QEMU rejects any contribution where AI use is “known or suspected”. The reason is the Developer Certificate of Origin (DCO), a statement in which contributors confirm they have the right to submit the code. According to QEMU, it's unclear whether anyone can make that statement for AI output at all. Zig justifies its ban differently. Loris Cro calls it “Contributor Poker”: you're betting on the person, not on their first PR, and an AI-written PR doesn't train a future maintainer.

Zig shows the cost most clearly. Bun, a JavaScript runtime that Anthropic acquired in December 2025, maintained its own Zig fork. According to Bun, it compiled four times faster. Because of the ban, Bun didn't even try to upstream the change. According to Simon Willison, though, one Zig developer said the patch would have been rejected even without AI. In July 2026, Bun founder Jarred Sumner ported the project to Rust in eleven days using Claude agents running in parallel, The Register reports. That cost an estimated $165,000 in API fees and produced more than a million lines of code. Zig lead Andrew Kelley called it “unreviewed slop”. With that, Zig lost its best-known user.

2. Disclosure. Many projects require you to disclose AI help, usually in the PR description. Over the summer, Godot banned autonomous agents and “vibe coding”, meaning programming by feel with AI without understanding the code yourself. According to reports, violations lead to an automatic ban. Minor help like autocomplete is still allowed but has to be disclosed. According to the study, this is the only type of rule agents reliably follow once reminded.

3. “The human is accountable.” Linux allows AI assistants, but an agent may never add the Signed-off-by line that confirms the DCO itself. Only a human can do that, and in doing so takes full responsibility. At LLVM, a human has to review every piece of AI output before anyone else sees it. Substantial help gets an Assisted-by: tag. Unreviewed output that just offloads work onto maintainers counts as “extractive”, meaning contributions that take more from the project than they give. Autonomous agents like the GitHub bot @claude are banned there. Debian voted against a ban in August and is going with “responsible use”. Disclosure is recommended there but not required. Linux 7.4 shows what this model costs. According to MLQ.ai, the merge window, the phase when new features land, had almost 20,000 commits instead of the usual 15,000 or so. Torvalds writes that a lot of that comes from AI fixes. In the BPF subsystem, maintainer Alexei Starovoitov says fixes now arrive faster than new bug reports. Torvalds only passes that on with a caveat (“sounds great, if true”), though, and gives no hard numbers.

How projects actually enforce their rules

According to the study by Hora and colleagues, projects respond to AI slop mostly by closing PRs (48.4 percent of observed measures). Next come banning users (20.8 percent) and banning autonomous agents (13.8 percent). In practice, it looks like this:

• CI bots and labels: Immich closes suspicious PRs with a GitHub Action and tags them with the auto-closed:llm label. The bot comments: “Unfortunately, this PR looks like it was generated with an LLM.”
• Account bans: Immich threatens a ban for repeat offenses, as do the PHP project recursion-context and Godot.
• “Good first issue” is off-limits: LLVM and Immich ban AI for beginner tasks so those stay with humans.
• Switching platforms: Zig moved to Codeberg in November 2025. Kelley hoped for fewer violations there and partly blamed GitHub's “Create issue with Copilot” feature. The cost: a large share of the Zig Software Foundation's income came through GitHub Sponsors. In July, Codeberg itself voted 358 to 144 to ban projects that consist mostly of AI code. It doesn't scan for it automatically, though.

Our sample: what gets through despite the bans

Using the GitHub search API, we analyzed every PR opened in four listed projects between October 12 and November 11, 2026. We only counted identifiable cases: bot labels, tags like “Generated with Claude Code” or “Co-Authored-By: Claude”, and contributors' own statements. The real number is likely higher.

QEMU takes patches via a mailing list, so we couldn't measure it through GitHub. Two findings stand out. First, projects with a bot do noticeably enforce their ban: Immich has applied the label 164 times since introducing it. One Hacker News commenter pointed out, however, that Immich still contains commits with Claude as co-author. Second, a ban stops at the project's boundary. AI code comes in through dependencies, as with Servo. And some violations are honestly disclosed, like most of the Meilisearch cases.

Darling: when the forks move faster

The fork pressure is clearest at Darling, which runs macOS programs on Linux, much like Wine does for Windows software. According to founder Luboš Doležel, barely any real applications ran despite years of work since 2012. Then AI-heavy forks pulled ahead: darling-lol gets League of Legends running, VibeDarling several native macOS apps. In the development update from November 8, Doležel writes that without AI, Darling would fall far behind and the community would splinter. The new rule lives in the AGENTS.md file:

Whoever opens a pull request is responsible for every line in it and must be able to explain it in their own words.

On top of that come strict clean-room rules. Clean-room means recreating behavior without looking at someone else's code. Darling bans using leaked Apple code and disassembly, meaning machine code translated back into readable form, as a template for matching code. Almost immediately, the opposite fork appeared: NoAiDarling, which accepts no AI code at all.

The bigger picture

Our take: a ban without technical enforcement is mainly a signal to humans. Honest contributors stick to it. According to RepoComplianceBench, agents practically never do, even when the rule is put in front of them. Bans only take effect with bots, account bans and CI checks, and even then only as far as AI can be detected. The “human is accountable” model, by contrast, targets a point that can actually be enforced: a human has to sign off, and a human has to answer follow-up questions.

How solid the numbers are remains an open question. GitHub's 40 percent is a lower bound with fuzzy boundaries. Torvalds' BPF claim isn't backed by data. And according to its author, osssaysno.com was itself built with AI, while critics on Hacker News consider the list overblown. One thing is clear: almost all of these policies are less than a year old, and half have already been revised. The rules are still being negotiated, and forks like Bun and Darling show that projects can lose users and contributors in the process.