DebOps LLM-Generated Contributions Policy
- Date drafted
2026-07-02
- Date effective
2026-09-01
- Last changed
2026-09-17
- Version
0.3.2
- Authors
Large language models (LLMs) and AI-assisted coding tools have become a standard part of many developers' toolchains. The DebOps Project treats them like any other tool: what matters is not how you wrote something, but whether you understand it, can defend it, and take full responsibility for it.
These standards of quality and accountability have always applied to contributions. We are making them explicit now because AI tools make it trivially easy to generate plausible content without understanding it.
Terminology
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14, [RFC2119].
This policy uses "LLM" for the class of AI tools that can generate or transform content, and "AI" as the umbrella term; the provisions apply to both.
Core Principle
There must be a human in the loop.
Anyone submitting code, documentation, or other content to the DebOps Project MUST review, understand, and be able to explain all content they submit, regardless of how it was produced. If you use AI tools to help write contributions, you are still the author and bear full responsibility for the work.
This means:
You have read and reviewed content before submitting it.
You, as a human, can answer questions about your work during review.
You have built and tested code contributions.
Submitting unreviewed AI-generated content wastes scarce maintainer time and is unacceptable. Contributions that lack thoughtfulness and care may be declined outright.
Contributors MUST review their work and MUST NOT knowingly submit content that infringes third-party copyright or licensing terms. Contributors MUST apply the same license-compatibility diligence to AI-generated content as to any other third-party code. Reproducing or regenerating copyrighted material with an AI tool does not remove its copyright. Contributors MUST NOT assume that generated content is original solely because an AI tool produced it; where they recognize generated output as third-party material, or tooling flags substantial similarity, they MUST establish the applicable license and their right to submit it before proceeding. Contributors are not required to audit output against unknown sources; they are required to act on what they can reasonably know, investigate, and recognize.
Whether content wholly generated by an LLM is copyrightable remains unsettled across jurisdictions. Where it is not copyrightable, it carries no license terms of its own and cannot be reliably licensed into the project under its existing terms. Copyright infringement does not require intent: failing to recognize that generated output reproduces third-party material does not remove the risk to the project.
Contributors MUST NOT use AI tools whose terms of service place restrictions on generated output that conflict with the project's licensing or the ability to OpenPGP-sign the resulting commits.
Meaning of "originally created by an LLM"
This policy uses the phrase "originally created by an LLM" to mean content that was generated by an LLM and then possibly edited by a human. No amount of editing changes how content was originally created; the origin sets the initial form, and that form is difficult to change once set. A human who lightly edits or rewrites LLM output is still submitting content originally created by an LLM.
This definition makes no distinction between LLM output from a chat interface and output from editor auto-completion. Both count as LLM output for the purposes of this policy.
The threshold in the "LLM Commit Trailer" section below is a reporting rule
that determines when the Generated-By: LLM trailer is required. Content
that meets the threshold was originally created by an LLM, and the trailer
is required regardless of subsequent editing. Content below the threshold
may still have been originally created by an LLM, but no trailer is
required for it. The threshold is not a proportional measure of the LLM's
share of the final content.
OpenPGP Signing
All commits that result from LLM-generated code SHOULD be signed by a valid OpenPGP key. This includes commits authored directly by a human using AI tools as well as commits made by AI-enabled forge accounts.
A valid OpenPGP key is one that has not expired, has not been revoked, and is discoverable on public keyservers or via Web Key Directory (WKD).
The signing key SHOULD be published on public keyservers (for example keyserver.ubuntu.com) and SHOULD be associated with the corresponding code forge account for easy verification.
This extends the existing DebOps Code Signing Policy, which already requires signed commits from DebOps Developers and recommends them for all contributors.
LLM Commit Trailer
Every commit that contains content generated in whole or in substantial part by an LLM MUST include the following Git trailer in the commit message:
Generated-By: LLM
A contribution is considered substantially generated by an LLM if more than 10 added lines of the diff, or any single line of logically significant content (a conditional, a regex, a configuration value), was originally created by an LLM, as defined in the section "Meaning of "originally created by an LLM"" above.
The Generated-By value SHOULD contain LLM. Contributors MAY append a
model identifier in parentheses — for example
Generated-By: LLM (Claude) — for additional transparency. Contributors
MAY note in the commit body how the tool assisted — for example proposed an
implementation, suggested changes to existing code, or drafted new tests —
when that context helps maintainers during review. Automated
tooling SHOULD treat any value containing LLM as compliant. A model
identifier is optional transparency, not an endorsement; a plain LLM
value is always sufficient.
This trailer lives in the Git metadata, not in file content. It is compatible with REUSE and SPDX file-level license and copyright headers — the trailer makes no statement about file-level licensing. It provides a machine-parseable record of provenance that automated tooling can use to evaluate compliance with this policy. Because LLM output can carry provenance uncertainty that a contributor cannot fully resolve, the trailer lets maintainers identify where the diligence obligations described in the Core Principle section apply and where extra review is warranted.
Generated-By: LLM records provenance; it does not determine whether the
content is copyrightable or who holds authorship. Under applicable law those
questions turn on the human creative contribution to the final content, which
this policy does not attempt to assess.
Maintainers MUST preserve the Generated-By: LLM trailer during merge. If
a PR branch is squashed, the committer SHALL add the trailer to the resulting
commit. If the merged result of a squashed branch exceeds the threshold in
this section, the committer SHALL add the trailer even if no individual
commit required it.
Commit Message Quality
All commit messages, regardless of how the code was produced, MUST follow the project's existing conventions:
Separate subject from body with a blank line.
Limit the subject line to 50 characters (hard limit 72).
Capitalize the subject line.
Do not end the subject line with a period.
Use imperative mood — "Add feature", not "Added feature".
Wrap the body at 72 characters.
Explain what and why, not how — the code shows how.
Keep each commit focused on a single logical concern.
Commit messages MAY be generated entirely by an LLM, provided a human reviews them for correctness and human-friendliness. The same quality standards apply as for human-written messages: the subject must fit the change, the body must explain the motivation, and there must be no AI-style boilerplate or verbose filler.
Pull Request Descriptions
Pull request descriptions MAY be authored by an LLM, provided a human reviews them before submission. The description must be human-friendly and follow the project's established format:
One or two paragraphs combining what the change does and why.
No section headers ("Summary", "Changes", "Why", etc.).
No bullet-point lists of changes (GitHub already shows the commit list).
No AI-agent boilerplate ("Here is a summary of the changes...").
Describe the change and its motivation in plain prose.
Verification of LLM-Created Code
Content originally created by an LLM is held to at least the same standard
as human-written content, and in some ways a higher one, because it is
easier to produce at volume. The Generated-By: LLM trailer is the signal
that a contribution requires the maintainers' extra attention during review.
Contributions that carry the Generated-By: LLM trailer MUST pass the
project's existing quality checks (make test-pep8, make test-spdx,
make docs) and MUST include tests where a test suite exists for the
affected code. This is deliberately a higher bar than the general
contribution standard. If tests cannot be written for a change, it SHOULD
be discussed with the maintainers before submission.
Changes to security-critical roles (for example secret, pki, and
auth) that carry the Generated-By: LLM trailer receive heightened
scrutiny from the maintainers during review.
Contributors MAY keep hand-written content and LLM-generated content in
separate commits when they are cleanly separable — for example a role
written by hand that includes LLM-generated helper scripts. Separating them
keeps the Generated-By: LLM trailer required by the LLM Commit Trailer
section on the commits that need it, instead of extending it to hand-written
content. Commits that add the LLM-generated content still need the trailer
regardless of how they are organized.
Separation applies only where the LLM content and the human content are distinct, typically in different files. When the two are intertwined in a single piece of content — for example an LLM-generated skeleton that a contributor then corrected and extended — the content as a whole counts as originally created by an LLM, as defined above. No separation of its edits into LLM and human commits is expected: one commit per prompt or per edit is not required, and the trailer is decided by the threshold in the LLM Commit Trailer section alone.
Contributors SHOULD describe how the LLM was used to produce the content —
for example how they directed the LLM, which generated parts they kept or
changed, and how they verified the result. The description is part of the
Git history, not the pull request: it SHOULD appear in a commit message
within the contribution, typically the commit that introduces or completes
the work. Commits in a multi-commit contribution need not repeat the
description; each still carries the Generated-By: LLM trailer required
by the LLM Commit Trailer section.
Prohibited Practices
The following practices are prohibited:
Unreviewed AI output. Submitting AI-generated content that the contributor has not reviewed or cannot explain.
Proxy submissions. Acting as a conduit between an LLM and the project without understanding or editorial judgment.
Unauthorized agents. Automated agents that take action in project spaces (opening issues, submitting PRs, posting comments) without explicit human approval for each action.
LLM-mediated review. Passing reviewer feedback or questions to an LLM and posting the AI-generated response without independent understanding or analysis. Reviewers expect to engage with a human who understands their own work.
AI-generated issue reports. Filing bug reports, feature requests, or security issues using LLM-generated content without human verification. Incorrect AI-generated security reports are particularly harmful because they waste maintainer time and desensitize maintainers to real issues.
LLM reviews are advisory only. An LLM review cannot, by itself, be a sufficient reason to merge or reject a change. An LLM-generated comment cannot, by itself, prevent a change from being merged; the human reviewer must agree with the comment and take responsibility for it before it can block a change.
Omitting the required trailer. Submitting content generated in whole or in substantial part by an LLM without the required
Generated-By: LLMtrailer. A first omission is a process violation; repeated or willful omission is a Code of Conduct matter.Misrepresenting provenance. Claiming authorship of content originally created by an LLM, or otherwise falsely stating that a contribution was not produced with LLM assistance. Misrepresentation is a violation of the Code of Conduct from the first occurrence.
Code review feedback that draws on LLM suggestions is subject to the same Core Principle: the reviewer MUST understand and stand behind the feedback they post.
Large-Scale and Automated Actions
The use of generative AI does not change the project's expectations regarding large-scale or automated actions. Contributors intending to perform actions with broad project impact SHOULD seek prior discussion and consensus through the appropriate project channel before proceeding. This includes, for example:
mass bug filing or patch submission
automated processes and tools that act on project infrastructure
bulk changes spanning multiple roles, playbooks, or other components
automated changes or requests affecting many components or contributors
Automated processes SHOULD be overseen by a human who remains accountable for their behavior and output. This complements the prohibitions on unauthorized agents and proxy submissions above: those rules govern who or what may act, this section governs actions whose scope affects the project as a whole.
A coordinated change that forms a single logical unit — for example a variable rename across several roles as part of one feature — remains subject to the normal review process and does not require prior consensus.
Scope
This policy applies to all contributions to the DebOps Project, including but not limited to:
Code (pull requests; PRs)
Issues and bug reports
Comments and discussions
Code review feedback
Documentation
Proposals and DEPs (DebOps Enhancement Proposals)
This policy does not apply to:
Non-LLM language utilities: spelling and grammar checkers, and machine translation that does not use LLMs, provided they do not generate novel content or restructure logic. Machine translation performed by an LLM produces content originally created by an LLM and is subject to the
Generated-By: LLMtrailer requirement.Authorized bots owned and operated by the DebOps Project (for example CI integrations, automated testing infrastructure).
This policy governs contributions to the DebOps Project. It does not govern the use of LLM output against live infrastructure — for example playbooks generated or modified by an LLM and executed against managed hosts. That use is covered by the project's operational safeguards and verification tooling, not by this document. The "Confidentiality of Information" section below is the exception: its requirements apply to operational use as well.
Confidentiality of Information
Third-party generative AI services collect and retain what they are given; treat them as what they are — external parties. Contributors MUST NOT disclose confidential or non-public material relating to the DebOps Project, its managed infrastructure, or its community to such services unless the disclosure has been explicitly authorized. This includes, for example:
Ansible vault contents and other encrypted secrets
inventory files,
group_vars, andhost_varsdatacryptographic keys, certificates, and credentials
internal hostnames, IP addresses, and network topology
private communications between contributors
security-sensitive information, including embargoed or not-yet-public details of security issues
personal data relating to the DebOps Project community
This requirement applies to operational use of generative AI tools in addition to contributions. The carve-out in the "Scope" section above excludes most operational use from this policy; the confidentiality rules are the exception and apply regardless of how the tools are used.
Before sending project or infrastructure material to a generative AI service, contributors SHOULD check the service's data-retention and privacy settings, and SHOULD replace real values — credentials, hostnames, IP addresses — with the reserved example values used elsewhere in the project's documentation.
Transition
Pull requests and issues opened before the effective date of this policy
(2026-09-01) are not subject to its requirements. Contributors are encouraged
— but not required — to apply the Generated-By: LLM trailer to subsequent
commits on existing branches.
Moderation
The disclosure requirements in this policy are not a license to police other contributors. Maintainers and contributors MUST NOT play detective for LLM usage. Writing style is not evidence of LLM involvement; people who write in a second language, who are neurodivergent, or who tend to over-explain are the most likely to be falsely accused.
Suspected undisclosed LLM use SHOULD be reported privately to the maintainers rather than raised as a public accusation.
Harassing a contributor for using an LLM is not allowed, regardless of whether the use violates this policy. The Code of Conduct applies to all conversations in the DebOps Project.
The disclosure requirement and the prohibition on harassment are enforced together. Disclosure works only when contributors are not punished for it, and the no-harassment rule works only when contributors are not suspected of hiding their LLM use.
Handling Violations
Before taking action on ambiguous cases, maintainers SHOULD ask the contributor for clarification and SHOULD cite a link to this policy. Contributors are welcome to ask clarifying questions, revise their work, and resubmit once they can demonstrate understanding of their contribution.
Reasonable belief that this policy was violated MUST be based on verifiable evidence — the contributor's own words, admission, or metadata — not on writing style. Maintainers MAY close pull requests, close issues, and hide comments that they reasonably believe violate this policy.
Repeated violations by the same contributor may result in a temporary or permanent ban from the project's spaces, handled according to the project's Code of Conduct enforcement guidelines.
For minor violations, maintainers SHOULD explain what is non-compliant and give the contributor a chance to fix it rather than treating the violation as misconduct. Repeated or willful omissions of the trailer, and any misrepresentation of provenance, are handled as Code of Conduct matters.
Conditions for Modification
Minor changes to this policy, such as typo fixes, require only a normal pull request. Major changes, such as adding or removing a rule, require a decision by the maintainers of the affected repository and a note in the project changelog.
This policy is expected to evolve as the project gains experience with LLM-assisted contributions. Feedback gathered from how the policy is applied may inform future revisions.
Additional References
LLVM Project, AI Tool Use Policy
Electron Project, AI Tool Policy
pip Project, Generative AI / LLM Policy
Kubernetes Project, AI Guidance
Apache Software Foundation, Generative Tooling Guidance
KubeVirt Project, AI Contribution Policy
Argo Project, Generative AI Contribution Policy
Linux Foundation, Generative AI Policy
Rust Project, LLM Usage Policy
Jynn Nelson, rust-lang/rust is adopting an LLM policy
rustc-dev-guide, LLM Guidance
Debian Project, General Resolution 2026-002: LLM usage in Debian
LWN.net, Debian votes to allow "responsible use of generative AI"
Software Freedom Conservancy, Recommendations When Using LLM-backed Generative AI Systems for FOSS Contributions
Free Software Foundation Europe, Copyrightability of LLM-generated code: Can we license "vibe code" into Free Software?
European Parliament, Generative AI and Copyright: Training, Creation, Regulation