QGIS, the leading open-source geographic information system, has merged QEP-428 — a total ban on AI, LLM, and agent-assisted code contributions to its main repository. The policy replaces the project's earlier QEP-408, which borrowed LLVM's human-in-the-loop approach requiring contributors to review all LLM output and tag it with Assisted-by trailers. Core contributor Nyall Dawson authored the replacement after concluding the previous policy's process "clearly didn't happen in reality." The new rule is blunt: "No LLM, AI or agent assisted code contributions are allowed in the qgis/qgis repository. If we subsequently discover that you have used an AI, LLM, or any agent-based workflow for your contribution, you will be subject to code reversion and contributor rights sanctioning." There is no ambiguity and no honor system — the enforcement mechanism is retroactive reversion and potential banning. Dawson's stated rationale is strategic, not ideological. The intent is to start with a closed door and selectively reopen it for specific use cases — AI as autocomplete replacement or translation assistance, for example. This is a governance move dressed in technical policy: define the default as prohibition, then negotiate exceptions from a position of control rather than trying to enforce nuanced rules nobody follows. The trigger was concrete damage. Dawson noted the project had "already lost at least one extremely high value contributor" as a result of the AI contribution issue — a cost that open-source projects, which run on a thin margin of volunteer goodwill and trust, can rarely afford. When your best people leave because they feel the codebase is being polluted by unreviewed machine output, the policy isn't working regardless of what it says on paper. The QEP passed its discussion period on October 5, moved to voting, and was merged into the QGIS-Enhancement-Proposals repository on October 7. The earlier QEP-408, adapted from LLVM's AI Tool Use Policy by contributor Even Rouault, had required transparency and review but lacked enforcement teeth. The gap between policy and practice proved fatal to the approach. This positions QGIS at the restrictive end of a spectrum that open-source projects are actively mapping. LLVM's original policy, which QEP-408 adapted, remains permissive-with-guardrails. The Linux kernel has taken a cautious but not prohibitive stance. QGIS's move is notable because it explicitly acknowledges that the middle-ground approach failed — not in theory but in the daily reality of pull request review — and chose clarity over flexibility. The deeper question is whether total bans are enforceable at all when LLM-assisted coding becomes the ambient norm. A contributor who uses Copilot suggestions while writing code and then types the result themselves occupies a gray zone that no policy can cleanly adjudicate. QGIS is betting that a clear bright line, even an imperfect one, creates better incentives than a nuanced rule nobody enforces.
