← Back to feed

On May 9, the team behind RPCS3 — the open-source PlayStation 3 emulator with 18,000+ GitHub stars — posted a notice on X: “Please stop submitting AI slop code pull requests to RPCS3. We will start banning those who do without disclosing.”

It is not the first project to say this. It is the latest in a sequence that has been running for six months, touching projects from gaming emulators to inference runtimes to search engines.

The Projects That Have Moved First

llama.cpp (December 2025): The maintainer of the mtmd module explicitly rejected AI-generated PRs after a string of them consumed disproportionate review time. “AI-generated PRs in mtmd are often, if not always, lead to sub-optimal, low-quality code. These PRs are usually huge, take me too much time to review/optimize, adding more mental strain and make the productivity goes negative.”

Ghostty (January 2026): Mitchell Hashimoto implemented a zero-tolerance policy. Drive-by AI-generated pull requests are immediately closed without review. Users who submit bad AI-generated content are permanently banned from future contributions. The policy drew a sharp line: “This is not an anti-AI stance. This is an anti-idiot stance. Ghostty is written with plenty of AI assistance and many of our maintainers use AI daily. We just want quality contributions.”

Meilisearch (March 2026): The team added an AGENTS.md file explicitly prohibiting AI-generated PRs and cited the direct cost: “We expected respect, but only got a lot of sloppy AI changes. We no longer have time for that and want to keep the codebase clean and sustainable.”

SDL (April 2026): The Simple DirectMedia Layer project added an AGENTS.md and PR template checkbox requiring contributors to confirm their code was not AI-generated. The team cited license uncertainty as a distinct concern beyond quality: AI-generated code may incorporate material from unknown origins, creating Zlib license compatibility issues.

RPCS3 (May 2026): The ban announcement arrived alongside a sharp reply to one developer who asked how the team distinguishes AI from human code: “You can’t possibly handwrite the type of shit AI slop we have been seeing.”

Godot’s project manager, Rémi Verschelde, said in February that the Godot GitHub page had become so overrun with AI-generated PRs that he was considering hiring additional maintainers specifically to process the volume.

The Structural Problem

Normal quality enforcement works because the distribution of contributor effort is spread. One bad PR in a week is manageable. AI changes the denominator: the cost of generating a structurally plausible, syntactically correct, utterly wrong pull request dropped to near zero. The result is not a worse PR distribution — it is a volume shock.

The deeper issue is cognitive. Maintainers reviewing code have historically assumed that authors understand what they submitted. That assumption no longer holds. GitHub’s community discussion in February — opened by product manager Camilla Moraes and calling the situation “a critical issue affecting the open source community” — surfaced this explicitly. One Microsoft Azure maintainer described the new review burden: “Reviewers can no longer assume authors understand or wrote the code they submit… AI-generated PRs can look structurally fine but be logically wrong, unsafe, or interact with systems the reviewer doesn’t fully know.”

The review burden is higher than it was before AI tools proliferated, not lower.

GitHub’s Options

GitHub is evaluating several responses: giving maintainers the ability to disable pull requests entirely, restricting PRs to existing collaborators, enabling deletion of PRs from the interface without leaving a public record, triage tooling, and transparency/attribution requirements for AI assistance. In a statement, product manager Matthew Isabel argued that PR quality rather than AI authorship is the right frame: “A bad or off-topic PR is a bad PR, regardless of where it came from.” But the framing misses what maintainers are actually asking — they do not object to AI-assisted code, they object to the volume of unreviewed AI output being pushed as genuine contribution.

The AGENTS.md convention has emerged organically as a signal: a file in the repo root that instructs coding agents not to submit unsolicited pull requests. SDL, Meilisearch, and others have adopted it. It does not prevent submissions, but it establishes intent.

The pattern is clear: the projects most affected are technically complex, actively maintained, and reliant on voluntary contribution. Emulators, media libraries, inference runtimes. The AI contribution flood is not democratizing open-source development — it is taxing the people who actually do it.