QGIS Bans AI-Assisted Contributions Entirely, After Its Gentler Policy Failed

QGIS, one of the largest open-source geospatial projects, has banned AI-assisted contributions outright. QEP428, “No AI Contributions Permitted,” was opened by core maintainer Nyall Dawson on September 24, moved to voting October 5, and merged into master on October 7, replacing the project’s earlier human-in-the-loop AI policy with a flat rule: “No LLM, AI or agent assisted code contributions are allowed in the qgis/qgis repository.” Violators face code reversion and sanctions on contributor rights. The proposal passed with approvals from Even Rouault, alexbruy, signedav, and elpaso, and a reaction tally of roughly 68 thumbs-up against 19 thumbs-down.

Why the Gentler Policy Failed

The ban’s power comes from what it replaced. QGIS had adapted LLVM’s human-in-the-loop policy (QEP-408), requiring disclosure and human review of AI-assisted contributions, and Dawson’s accounting of the experiment is the document’s core evidence: about 50 pull requests were tagged AI-assisted, roughly 40 were closed, the review burden created what maintainers called “a massive and unbearable pressure,” and the project “lost at least one extremely high value contributor” in the process. The disclosure-plus-review model assumed maintainers had spare capacity to audit machine-generated code; in practice, the tags became a signal that shifted work onto reviewers rather than a filter that protected them. A confessional comment from contributor Amalvir, who had submitted AI-assisted fixes and now supports the ban, made the same point from the other side: LLMs created an “illusion” of competence that broke the implicit trust contract with reviewers.

The Enforcement Model Is Honor System Plus Reversion

QEP428 contains no CLA mechanism and no technical detection; enforcement rests on the good-faith standard Dawson described (“did you in good conscience write the code yourself”) plus post-hoc reversion when violations surface. Critics in the thread (kannes, 02JanDal, uclaros) pushed on exactly this, arguing the wording is broad enough to ban brainstorming with an LLM, ML autocomplete, or even AI-ranked search results, and urged narrower language targeting LLM-written code. Dawson’s counter is the stopgap framing: start with the door closed, then selectively reopen it case-by-case for uses like autocomplete or translations. That sequencing, ban first, carve-outs later, is the inverse of how most projects have handled it, and its success depends on whether reversion proves cheap enough to be a real deterrent.

The Precedent for Open Source

QGIS is the largest project yet to move from AI disclosure rules to a total ban, and the reasons it cites are structural rather than aesthetic: review capacity is finite, unreviewed code creates quality debt, and free software’s economy runs on human networks, contributors whom organizations can hire for support and training, which Vincent Picavet’s “huge +1” made explicit. Read against the week’s other signals (Apple locking down Full Disk Access, Nvidia building agent containment into silicon), a pattern is emerging: the places with the most accountability for quality are the first to choose hard limits over disclosure regimes. Projects with millions of users and few maintainers cannot review AI output at AI generation speed, and QGIS is the first to say so formally rather than let the review queue say it silently.

What to Watch

Three things. First, whether other mid-sized projects follow: QGIS’s evidence base, 50 tagged PRs and a lost core contributor, is the exact experience most maintainers will recognize, and QEP428 is written to be copied. Second, the carve-out path: whether autocomplete and translations reopen without the ban unraveling, which is the failure mode disclosure policies hit. Third, the effect on the lost contributors: if the ban draws back high-value humans who left over AI-PR noise, that trade, machine volume for human relationships, is the actual thesis of the policy, and its success or failure will be the case study every other project cites.