ラウンジ
接続中…
接続中…
すべて見る

Why Are Open-Source Projects Banning AI Contributions? The QGIS Moratorium Explained

QGIS banned LLM- and agent-assisted code in October 2026. See why open source projects restrict AI contributions and how QEMU, Gentoo and Linux differ.

AI最新7分で読めます閲覧 22公開 更新
目次7セクション
  1. What the QGIS AI moratorium actually says
  2. How open source AI contributions policies compare
  3. Why the AI generated code license question is still open
  4. Open questions: scope and enforcement
  5. What to check before contributing
  6. よくある質問
  7. 出典・参考資料

要点

  • On October 7, 2026, QGIS approved version 2.0 of QEP 408: no LLM, AI or agent-assisted code contributions to the qgis/qgis repository.
  • The main driver was maintainer review load. Under the earlier, softer policy, roughly 40 of about 50 PRs tagged AI Assisted were closed without merging, according to a contributor in the discussion.
  • QEMU and Gentoo reject AI-generated contributions mainly over copyright, licensing and quality concerns tied to the Developer Certificate of Origin.
  • The Linux kernel allows AI assistance, but only humans may add Signed-off-by, and an Assisted-by tag records the tool use.
  • Policies differ by project and change over time, so read the current text in the repository before you contribute.

Open source AI contributions are being restricted for two practical reasons. First, AI tools let anyone produce pull requests faster than volunteer maintainers can review them. Second, contributors cannot easily certify the copyright and license status of AI output, which many projects require before accepting code. As of October 2026, the clearest recent example is QGIS, the open-source geographic information system: it approved version 2.0 of QEP 408, which bars LLM, AI and agent-assisted code from the qgis/qgis repository. This guide explains the decision, compares it with QEMU, Gentoo and the Linux kernel, and lists what to check before you contribute. All sources were read on October 13, 2026.

What the QGIS AI moratorium actually says

The policy is two sentences long. No LLM, AI or agent-assisted code contributions are allowed in the qgis/qgis repository, and if the project later discovers that a contributor used an AI, LLM or agent-based workflow, the contributor faces code reversion and sanctions on their contributor rights. The document is labeled version 2.0, dated September 24, 2026, and attributed to the QGIS project contributors. It replaces an earlier AI tool policy dated January 31, 2026, which participants described as a human-in-the-loop approach rather than a ban.

The change went through pull request #387 in the QGIS Enhancement Proposals repository. Although the PR title refers to QEP428, it edits the existing QEP 408 policy file. It opened for discussion on September 24, moved to voting on October 5, and was declared approved and merged on October 7, 2026. The rule targets code entering the core repository. It does not restrict people who use QGIS, nor does it stop anyone from building AI-related software elsewhere.

Sep 16

Stage
Mailing list proposal
What happened
A core developer proposed a total moratorium on LLM-based contributions plus a proof-of-life check for new contributors

Sep 23

Stage
Steering committee view
What happened
A PSC member backed a ban and asked for it to go through the QEP process

Sep 24

Stage
PR #387 opened
What happened
Discussion began on a strict not-allowed version 2.0

Oct 5

Stage
Voting
What happened
Discussion period ended and the proposal moved to a vote

Oct 7

Stage
Approved and merged
What happened
Approval declared based on the votes cast; policy file merged

Why QGIS chose a hard stop

In the QGIS developer mailing list thread, the proposer argued that gentle policies had not worked and that the project should default to no AI now and loosen later, rather than start loose and tighten afterward. In the PR, a project member summarized three reasons. AI use by external contributors placed heavy pressure on maintainers who were already stretched. The volume of AI-produced PRs exceeded what humans could seriously review, leading to either a growing backlog or code merged without full review. And QGIS depends on a human network: organizations trust it partly because they can hire and work with real contributors.

Participants also questioned how the earlier policy performed. One contributor noted that a planned classification label was never created and that of roughly 50 PRs tagged AI Assisted, about 40 were closed without merging. The proposer replied that the earlier process did not happen in practice and that the project had lost at least one high-value contributor. A beginner who had submitted two AI-assisted regression fixes supported the ban, writing that LLMs create an illusion of competence and that contributors who lean on them learn little and never grow into reviewers.

How open source AI contributions policies compare

Projects do not agree on either the rule or the reason. QGIS emphasizes review capacity and community trust. QEMU and Gentoo put more weight on copyright, licensing and quality. The Linux kernel accepts AI assistance but makes a human fully responsible. The table summarizes each project's official text as read on October 13, 2026; any of these can change.

QGIS

Position on AI contributions
No LLM, AI or agent-assisted code in qgis/qgis; reversion and sanctions if discovered
Main rationale
Maintainer load, quality, community trust
Exceptions or conditions
Specific uses may be reopened later through separate proposals

QEMU

Position on AI contributions
Declines contributions believed to include or derive from AI-generated content
Main rationale
Contributors cannot credibly certify the DCO; unsettled copyright and license status
Exceptions or conditions
Researching APIs or algorithms, static analysis and debugging are outside the policy if output is not included

Gentoo

Position on AI contributions
Content created with assistance of natural language processing AI tools is forbidden
Main rationale
Copyright, quality, ethics
Exceptions or conditions
Packaging AI-related software is still allowed

Linux kernel

Position on AI contributions
AI assistance permitted
Main rationale
Existing GPL-2.0-only licensing rules still apply
Exceptions or conditions
Only humans add Signed-off-by; human reviews and takes responsibility; Assisted-by tag

QEMU and the Developer Certificate of Origin

QEMU's code provenance page requires every patch to carry a Signed-off-by line certifying the Developer Certificate of Origin 1.1, which states that you have the right to submit the work under the indicated open source license. QEMU says the copyright and license status of AI output is ill-defined, that training data commonly includes material under restrictive terms, and that even open source training data may carry licenses incompatible with QEMU. Because it is unclear how a contributor could comply with the DCO, QEMU declines contributions where AI use is known or suspected. Using AI for research, static analysis or debugging is not covered, provided the output is not part of the contribution.

Gentoo's ban versus the kernel's disclosure model

The Gentoo Council policy, voted on April 14, 2024, expressly forbids contributing content created with the assistance of natural language processing AI tools. It cites emerging copyright rules that could weaken Gentoo's own copyright claims and copyleft guarantees, the cost of reviewing plausible-looking but wrong output, and ethical concerns such as training practices, energy use and spam. The Linux kernel's AI Coding Assistants guide takes the opposite route. AI agents must not add Signed-off-by tags because only humans can certify the DCO. The human submitter reviews all AI-generated code, ensures license compliance and takes full responsibility, and records tool use with an Assisted-by tag.

Why the AI generated code license question is still open

Open source licenses rest on copyright. A project can enforce copyleft terms only if contributors hold rights in what they submit and license those rights to the project. That is why Gentoo warns that generated material could both create infringement risk and weaken the project's ability to rely on copyleft. In the United States, the Copyright Office has been publishing a multi-part report on copyright and AI: Part 2 addresses whether generative AI outputs can be protected, and a pre-publication Part 3 covers training data. Interpretations differ by jurisdiction and are still developing, so some projects prefer a conservative rule over waiting for courts and legislators.

Licensing is not the whole story, though. A QGIS steering committee member suggested keeping rationale out of the policy entirely to avoid endless arguments about nuance, and the reasons most often cited in the PR were review capacity and trust rather than law. When you read any project's AI rule, identifying what the maintainers worry about most will tell you how narrow or broad the exceptions are likely to be.

Open questions: scope and enforcement

  • Scope: commenters asked whether the words assisted and used cover LLM autocomplete, older machine-learning ranking in IDEs, design brainstorming or translation tools.
  • Review and security: some asked whether using an LLM as a mock reviewer or for security audits is also barred; others proposed handling that in a separate discussion.
  • Non-code text: one participant urged extending the rule to issues, documentation and PR descriptions, while another noted that submitted security advisories were already being written with AI.
  • Enforcement: maintainers cannot inspect contributors' machines, so the stated test is whether you honestly wrote the code yourself. That relies on self-reporting, yet discovery carries serious consequences.

The proposer described the policy as a starting point: shut the door, then open it a little for specific uses such as autocomplete or translation once each has been discussed. Read it as an operating rule chosen by a volunteer community to protect its review capacity, not as a technical verdict on AI coding assistants.

What to check before contributing

  1. 1

    Find the written policy

    Look in CONTRIBUTING files, developer documentation, the project wiki and any enhancement-proposal repository. QGIS keeps its rule in the QEP repository, QEMU in its code provenance page and the kernel in its process documentation.

  2. 2

    Note the version, date and scope

    Record which repository and which kinds of contribution (code, docs, issues) the rule covers. Different repositories in the same organization can have different rules.

  3. 3

    Separate allowed and disallowed uses

    Some projects allow research and debugging but reject AI output in the patch; others restrict assistance broadly. If your case is unclear, ask on the mailing list or issue tracker first.

  4. 4

    Only submit code you can explain

    Even where AI is allowed, Signed-off-by is a human statement of responsibility. If you cannot justify each change in review, do not send it.

Debates like this spread quickly through developer channels. The theddari AI Telegram trends and blog trends pages can help you see which related topics are being discussed, but always confirm exact wording and dates against the project's own documents.

よくある質問

Does the QGIS policy stop people from using AI with QGIS?
No. As of October 2026 the policy text covers code contributions to the github.com/qgis/qgis repository. It does not regulate how people use QGIS or build their own tools. Whether it should also cover issues, PR descriptions and comments was raised in the discussion without a clear written answer, so check the project's contributing guide before posting AI-written text there.
Is using AI autocomplete or a translator a violation?
The short policy wording does not spell this out, and that was a central criticism. Participants asked about LLM-backed autocomplete, machine-learning ranking in IDEs, AI search summaries and translation tools. A steering committee member framed the test as whether you, in good conscience, wrote the code yourself, and the proposer said cases like autocomplete or translation could be reconsidered later in separate proposals. When in doubt, ask maintainers before submitting.
Why does the Developer Certificate of Origin matter for AI-generated code?
When you add Signed-off-by, you certify that you have the right to submit the work under the project's license. QEMU argues that the copyright and license status of AI output is ill-defined and that training data can include code under incompatible terms, so a contributor cannot credibly make that certification. The question is less about who typed the code and more about whether its origin can be accounted for.
Will QGIS lift the ban?
The proposer described it as a closed door that may be opened selectively, case by case, after proper discussion, not a permanent rule. Another core contributor stressed that it does not mean AI contributions are banned forever in every form. No relaxation had been adopted as of October 13, 2026, so check the QEP repository for later changes.

出典・参考資料

  1. 1QGIS-Enhancement-Proposals PR #387: Add QEP428 No ai contributions permitted(新しいタブで開きます)github.com
  2. 2QGIS QEP 408: AI tool use policy (Version 2.0)(新しいタブで開きます)raw.githubusercontent.com
  3. 3QGIS-Developer mailing list: moratorium on LLM use thread(新しいタブで開きます)lists.osgeo.org
  4. 4QEMU documentation: Code provenance(新しいタブで開きます)www.qemu.org
  5. 5The Linux Kernel documentation: AI Coding Assistants(新しいタブで開きます)docs.kernel.org
  6. 6Gentoo Wiki: Project:Council/AI policy(新しいタブで開きます)wiki.gentoo.org
  7. 7U.S. Copyright Office: Copyright and Artificial Intelligence(新しいタブで開きます)www.copyright.gov