AutoGPT maintainer Nicholas Tindle puts the appeal of pull-request limits plainly. GitHub quotes him saying that knowing someone “hasn’t just opened 5–10 pull requests that are slop makes it much easier to want to look.” The fatigue is real, and so is the relief. But the filter that gives it back also decides who gets read at all. AI lets newcomers send code to open-source projects more easily than ever, while maintainers swamped by junk close those submissions within hours — on account age, or because a PR “looks AI” — often before they have read far enough to say why.
What maintainers actually wrote when they closed
The first project-level look at this behaviour comes from Raskin, Devroey and Vanderose on arXiv , published 11 November. They collected 1,786 pull requests that maintainers had labelled or titled as AI slop, across 76 of 10,000 popular repositories. About 98% were closed unmerged, with a median closure time of roughly six hours.
The reasons are the telling part. The largest group, 439 PRs, carried no stated reason. Another 117 were closed by maintainers (and 54 by bots) on the grounds that the PR “looks AI” with nothing more. Another 82 by maintainers and 12 by bots rested on the author’s profile.
One bot on huggingface/transformers spells it out. Its flags include “New GitHub account” and an “unusually high number of repository forks in a 24-hour window.” Every criterion describes the account. None describes the code.
GitHub is building the same logic into the settings page
GitHub’s 28 October changelog lets maintainers exempt contributors from pull-request limits by account age, or by a minimum number of merged PRs in that repository. Meeting either rule is enough. The post gives no default thresholds and no usage data.
This is the milder path compared with projects that closed public PRs entirely , and the design is reasonable. Homebrew’s Mike McQuaid says limits let the project “still have outside contribution and maintainers contribute more.” Read next to the study, though, it shows where the defence is settling: on the sender.
Open source's defence against AI slop has settled on who sent the pull request rather than what is in it, and that is a tax on newcomers.
That tax lands on the people AI was supposed to let in, a tension also visible when Google Summer of Code mentors hunt for AI in applications. In the Linux Foundation’s survey, 71% of maintainers agree that generative AI lowers barriers to entry. Picture a developer sending a first pull request to a popular project with an AI assistant’s help. Their account is a month old, so they may be capped, flagged or closed before anyone reads the diff.
The honest complication
The case against alarm is serious. Only 4.1% of flagged PRs’ authors protested, which the paper itself says partly offsets the worry about misclassification. GitHub’s account-age rule is an exemption from a cap on open PRs, not a ban, so new accounts can still contribute at low volume. And curl’s Daniel Stenberg said in April that AI-assisted security reports had become “really good” , which hints the slop wave may partly pass.
Time is the sharper objection. The same Help Net Security write-up of the LF survey notes that on big projects burnout has “moved to the top of the list,” and 64% of large-project maintainers report longer review cycles, by the LF’s figures. Asking them to state a checkable reason on every closure costs hours they don’t have.
Now the step this piece takes beyond its sources. The paper measures PRs that maintainers flagged, not how many flagged authors were human newcomers, and nobody has compared sender-based triage with code-based triage. The authors concede the label “may conflate AI-generated content with other low-value contributions by inexperienced contributors.” Two repositories, Transformers and Click, supply about 42% of the sample. GitHub’s Octoverse data adds that agent-involved PRs were merged within two weeks about 90% of the time in September, against roughly 86% for others — though it does not split by repository ownership, so that may reflect agents working in their own teams’ repos.
What follows is an inference, not a finding. Authorship is not the defect, and the paper’s own reason codes include several that need no guess about the sender: build failure (42), duplicated PR (76), insufficient verification (67), issue hijacking (15). A maintainer, or a bot, can check each of those against the PR itself. The authors call for tools built on verifiable reasons, and that is the sender-neutral triage still missing.
| Category | PRs closed |
|---|---|
| Build failure | 42 |
| Duplicated PR | 76 |
| Insufficient verification | 67 |
| Issue hijacking | 15 |
Even Electron’s closing bot allows resubmission when a PR is closed in error. The welcome mat is out, but it is hard to see from the street.



