QGIS moves to a vote on banning all AI contributions, at least for now

A proposal to shut the door entirely

On October 5, a proposal to ban AI-assisted contributions to QGIS, the open-source mapping software, passed its discussion period and moved to a vote. The proposal, QEP428, was opened on September 24 by contributor Nyall Dawson. It formally replaces the project's earlier "human in the loop" AI policy with a strict "NOT ALLOWED" rule. The source does not show the vote's outcome, so what follows describes a proposal, not settled policy.

How the old policy worked, and why it didn't

QGIS enhancement proposals (QEPs) are the project's formal way of changing how it operates: a written proposal, a discussion period, then a vote. The earlier policy allowed AI help as long as a human stayed accountable for the result. It relied on labels, including an "AI Assisted" tag for pull requests (proposed code changes) and an "extractive" tag that, according to contributor uclaros, was never actually created.

The record uclaros cited: about 50 pull requests were tagged AI Assisted, and about 40 of them were closed without merging. Others were closed without ever being tagged. Dawson called the earlier approach a policy that "clearly didn't work" and said the project had already lost at least one extremely high-value contributor as a result.

The argument: review bandwidth, not code quality

The core complaint is about volume. Member haubourg wrote that the AI stream is far higher than human review can take seriously, which produces either a growing pile of unreviewed pull requests or code that is not fully reviewed. He also argued that QGIS is trusted because organisations can contribute back and hire contributors, and that this human network is what needs protecting. He added that every security advisory submitted since that system began was written with AI, though most concerned the project's websites, not QGIS core.

One contributor, Amalvir, who had submitted two AI-assisted regression fixes, backed the ban. They described a broken trust contract, the assumption that someone contributing to a project this complex knows what they are doing, and said contributors who lean on AI learn little and never grow into reviewers.

The objection: what exactly is banned?

Critics focused on the wording. Contributor kannes said the words "assisted" and "used" are broad enough to ban LLM use for code review, auditing and security, and described using one as a simulated reviewer before sending a PR to colleagues. Another contributor, 02JanDal, listed ambiguous cases: LLM-backed autocomplete, older machine-learning ranking in IDE completions, brainstorming a design in a chatbot, accidentally reading a search engine's AI answer, or translating a term with Google Translate. He suggested narrowing the text to code written by LLMs or AI agents in the main qgis repository, and noted that clarity matters because a penalty attaches.

Supporters answered that it is a stop-gap. Dawson described it as a closed door that can later be opened selectively, case by case, for things like autocomplete or translation. Contributor rouault-on-strike conceded the policy may throw out the baby with the bathwater. Member 3nids gave a +0 and said he would open a separate QEP for further discussion; elpaso voted +1, stressing the ban is initial, not permanent.

Questions You Should Be Asking

• If your project or vendor relies on open-source software, what is the maintainers' stated capacity to review changes, and does AI-generated volume threaten it?
• How would a ban like this be enforced, given that code contains no reliable marker of AI assistance and the earlier labelling process was never fully carried out?
• Does the final wording separate AI-written code from AI used for review, search or translation, or does it leave contributors guessing about penalties?
• What happens to security reporting if AI-written advisories, which haubourg says make up all submissions so far, are covered by the ban?
• Who decides when and how the door reopens, and what evidence would they accept?

What To Watch Next

Watch the voting outcome and any amendments to the wording, especially whether the text is narrowed as 02JanDal proposed. Then watch the follow-up QEP 3nids said he would open, and the discussion haubourg planned at the user conference on AI for review, auditing and security. Those will show whether this ban stays absolute or becomes a policy with defined exceptions.