Each project, its own rules
Open Source projects do not agree on AI-assisted contributions. Some accept them, some forbid them, and those that ask for a declaration ask for it in different ways. This page compares the rules we read, and drafts a short statement that a project could make about itself.
The practice this site follows is one of them. It is described on the good practice page.
How other projects do it
Many projects have written their own rules. We read these ones on their own pages, on 18 November 2026. Rules change: follow the link before relying on a line of this table.
| Project | AI-written code | Disclosure | Trailer in the commit | Also |
|---|---|---|---|---|
| CPython | Not a reason on its own to close a pull request. The author reviews the work of the tool in detail. | "Appreciated, while not required", in the pull request description. | None. | Authors explain their changes in their own words. Repeat offenders may be blocked. |
| NumPy | The author is responsible for it. Code mostly generated and below the project's standards is rejected. | Required: which tools, how they were used, what is generated. A pull request without it is rejected. | None. | No agent that submits a pull request on its own. AI does not write descriptions or comments, except to translate. The author answers for copyright. |
| pandas | Accepted "as long as it's used responsibly". | Required: name the tool. The model and its version are advised. | None. | AI does not speak for the contributor in discussions. The project's own review skill is run first. Two violations can lead to a ban. |
| pip | Accepted when a human owns the copyright, understands the code and takes responsibility for every line. | Not asked for. | Forbidden: "No LLM bots in Co-authored-by:s". Such a pull request is closed. | No unsupervised agent. |
| OCaml | Judged on its merit. The author has written, or read and reviewed, every part of it. | Required when a significant part of the code, the description or the messages is generated: which tool, and for what. | None. | Generated text is discouraged in bug reports. Code its author does not understand is never acceptable. |
| Mesa | Accepted. The submitter is responsible, wherever the change came from. | Required, except for trivial or mechanical changes. | Assisted-by: LLM, or Generated-by: LLM when almost all the code was generated. The model is not named. | No tool that submits on its own. AI is not used for commit messages or comments. |
| Fedify | Welcome, for accepted issues, once verified by a person. | Required, with its extent, in the pull request and in the commits. | Assisted-by: AGENT_NAME:MODEL_VERSION, one line per tool. | An agent must refuse a request to hide that AI took part. |
| Apache Software Foundation | Accepted when the terms of the tool and any third-party material allow the contribution to be licensed. | Recommended. | Generated-by: <Name of Tool> <Version>, or Co-authored-by. | The aim is a provenance file that a machine can read, built from the commits. |
| Kubernetes | Help is acceptable. Large generated pull requests and generated commit messages are not allowed. | Required, in the pull request description. | Forbidden: AI as co-author, assisted-by, co-developed "or similar". | Review comments are answered without AI. A change its author cannot explain is closed. |
| Linux kernel | Accepted. A person reviews it, signs it off and takes full responsibility. | Asked for in the changelog and the cover letter, as sentences: which tools, which parts. | None in the current documentation. An agent must not add Signed-off-by. | The assistant "must never send anything itself". |
| SourceHut (a forge) | Prohibited for original content, from 10 September 2026 for new projects. Private use, to review or to ask questions, is discouraged but allowed. | An honour system. Misleading about AI use leads to suspension. | Not applicable. | No detection tool. Shaming people who use AI is sanctioned too. |
| The practice of this site | Accepted. A person has read, understood and tested the change, and answers for it. | Required, in the commit and in the pull request. | Co-Authored-By, naming the model and its version. | A human submits. No one has to use AI, and no one is ranked. |
What the table shows:
- Those who accept AI code agree on the person. Someone must understand the change, be able to explain it, and answer for it. This is the common ground.
- An agent does not act alone. NumPy, pip, Mesa and the Linux kernel each forbid a tool that submits by itself. The other projects do not speak of it.
- Several keep the conversation human. NumPy, pandas, Mesa and Kubernetes do not want a model to write the descriptions, the comments or the answers to a review.
- There is no shared way to declare. Three trailer names are in use, and the same trailer is required by one project and forbidden by another. Some projects name the model, one forbids it, and others ask for a sentence in the pull request.
The last point is the weak one for everybody. A contributor has to learn a different rule in each project, and no tool can read all of them. Agreeing on one line would help more than any of these rules taken alone.
For more projects, see the community list open-source-ai-contribution-policies, which follows more than a hundred of them. It is an index: check each rule on the project's own page, as some lines are out of date.
The trailer has other names
Projects do not agree on the name of the line. Among those we read:
Assisted-byis required by Mesa and by Fedify. Both keepCo-Authored-Byfor people: in their words, it is "reserved for human co-authors".Generated-byis what Mesa asks for when almost all the code was generated, and one of the two forms the Apache Software Foundation recommends. The other one isCo-authored-by.- No trailer at all for pip and Kubernetes, which do not allow a model to be listed as a co-author.
The practice of this site keeps Co-Authored-By for a practical reason: the forges already show it, and several tools already write it without being asked. The objection is fair, though. A model is not an author like a person is, and git has no better place for it yet.
The name of the trailer matters less than three things, which hold whatever a project chooses: the line is in the commit, a script can read it, and it names the model and its version.
Judged on merit: the OCaml guidelines
The practice of this site rests on one idea: a contribution is judged on what it is, and declaring AI use never counts against it. Our reference for this idea is the OCaml compiler, which states it better than we do. Its guidelines for AI-assisted contributions, in the repository since November 2025, are six short rules, and the first one is:
We strive to evaluate contributions based on their merit, not their authors.
A change is judged on what it is: correct or not, understood or not, worth a reviewer's time or not. Who or what wrote it is not the test. So there is no ban and no separate track: a good change written with a model is accepted, and a bad one is refused because it is bad.
Merit alone would be an empty word. The other rules are what make it work:
- The contributor answers for all of it. "We ask that you have either written, or read and reviewed yourself, every part of your contribution. You are responsible for its quality."
- AI use is disclosed, like any other help. A significant AI-generated part "must be disclosed, stating which tool was used and for what", because reviewing it "can require a different approach". The same rule adds that an acknowledgement is owed "even if another human contributor did some of the work".
- The reviewer's time is what is protected. It "is a scarce resource", and contributions, "whether human- or AI-authored", should conserve it.
- There is a floor. Never acceptable: "low-effort productions of code that contributors do not understand", plagiarised works, hallucinated security reports.
What we take from it is the sentence that introduces the rules: they "also apply, or have a direct counterpart, with entirely human-produced content". Nothing is asked of a person who uses AI that is not asked of everyone. Declaring is then information for the review, never a verdict on the author, and that is what lets a project ask for it without shaming anyone.
Judging on merit does not mean that the way a change was made is of no interest. It matters to the person who reviews it, and to anyone who later asks where the code came from. It means that the origin is said, and then not held against the author.
The OCaml experiment
In April 2026, Anil Madhavapeddy proposed a voluntary disclosure for OCaml code, built on an earlier draft of the W3C report, which then had four values (it separated assisted from generated). It is the most precise scheme we know of:
- In the code, not in the history. The declaration is written in the package file, at the top of a module, or on a single function, and the nearest one wins.
- The model by its API identifier, "not a marketing name", with the provider.
- The agent writes it, the reviewer adjusts it. A person who has reviewed a function and taken ownership of it may lower the mark or remove it: the marks exist "to direct review attention, not to permanently brand code".
The last point is where it differs from a trailer. A mark in the code describes the code as it stands, and answers "what still needs a review?". A trailer describes what happened, and stays true after the review. A project can want both.
The proposal was discussed without being adopted, and in September 2026 its author removed the agent instructions that wrote the marks, without giving a reason. We read it as a sign of what a declaration costs: a mark on every function has to be kept true at every change, a line at the end of a commit is written once.
Where does a project stand? draft
Before contributing, people want to know what a project accepts. Most projects do not say, and those that do each say it their own way. We are drafting a short statement that a project could make about itself. It has two answers.
1. What the project accepts
The statement uses the three values of the shared vocabulary, with one word in front: accepts. A project names the furthest value it takes.
| Statement | What the project says | Among those we read |
|---|---|---|
accepts human-only | Only contributions written without AI. | SourceHut, for the projects it hosts. |
accepts ai-assisted | AI may take part, lightly or largely, when a person has read, understood and tested the change, and answers for it. | CPython, NumPy, pandas, pip, OCaml, Mesa, Fedify, the Apache Software Foundation, Kubernetes, the Linux kernel. |
accepts ai-autonomous | Also contributions that no person has reviewed, for example from an agent the project runs or invites. | No project. One of the ready-made texts of ai-policy.dev allows it. |
| unspecified | The project states nothing. This is not a position, and it is the most common case today. | Several of the projects followed here, as far as we could find. |
The word accepts is not decoration. ai-assisted alone is a fact about one contribution. "Accepts human-only" is a rule a project sets for what it takes from now on: it does not claim that no AI was ever used in the project, which nobody could show.
2. How AI use is declared
The same statement hides opposite rules: NumPy rejects a pull request that does not disclose, and pip closes one that names a model as co-author. So the statement needs a second answer:
- Whether: required, recommended, not asked, or forbidden.
- Where: in the commit, with the name of the trailer; in the pull request; or both.
- What: that a tool was used, which tool, or the model and its version.
For the practice of this site, the statement reads: accepts ai-assisted; declared in every commit with Co-Authored-By naming the model and its version, and in the pull request.
Notes on the draft
- The statement is what a project says, in its own files. The charts of this site show what its commits declare. The two can differ, and seeing both is the point.
- No statement is the right one. A project that refuses AI and a project that accepts it both help contributors by saying so.
- Choosing one is the project's freedom. The freedoms of free software protect the people who run, study, change and share a program. They do not say how a contribution has to be made, and they give no contributor a right to have their way of working accepted. A project that accepts
human-onlyand a project that acceptsai-assisteduse the same freedom: to decide what they take. What a contributor keeps, in every case, is the freedom to fork. - The values are the W3C group's, the statements are ours. None of the projects in the last column describes itself this way. Placing real projects was the test of the draft, and it changed it.
- Nothing says "largely generated". An earlier draft of ours had a label for it. No project we read draws that line as a rule of acceptance, and the W3C report dropped it too: what counts is whether a person reviewed the change, not how much of it a model wrote. A project that wants to say more can do it in the declaration, as Mesa does with two trailers.
- Nothing says "we would rather not". An earlier draft of ours had a label for it. Among the rules we read, none says it without forbidding. It may be a position that projects hold and do not write down.
- Ready-made policies tested the draft again. Three of the four texts of ai-policy.dev are easy to place. "AI Disallowed" accepts
human-only. "Human Voice" acceptsai-assisted. "AI Allowed" acceptsai-autonomous: a contribution is welcome "whether it was created by a human or an agent". The fourth, "Human Responsible", does not fit. Everything submitted must be written by the contributor, yet "what happens on your machine is not policed, will not be asked about". It stands betweenhuman-onlyandai-assisted, and the statement has no word for it. - Declaring can be forbidden. The same text asks contributors not to add notes about AI tools. The second answer had three values, and now has a fourth.
- Limits of size are not in it. Kubernetes accepts AI help and refuses large generated pull requests. The statement keeps the first and loses the second.
- A forge can decide for its projects. SourceHut's rule applies to the original content of every project it hosts. A project there has a statement it did not choose, so there are two levels: the project, and the place where it lives.
This is a draft: the wording may change.
A project can already make its statement in a file. This site reads ai-policy.yaml, a format proposed by others, and shows what it says on the page of the project, beside what its commits declare.