If a Machine Wrote the Code, There Is No One to License It

In short
Key takeaways
- The US Copyright Office's Part 2 report, published 29 January 2025, holds that purely AI-generated material is not copyrightable and that prompts alone do not give a user enough control to be an author.
- Thaler v. Perlmutter, decided by the D.C. Circuit on 18 March 2025, held that only natural persons can be authors under the Copyright Act.
- Doe v. GitHub, decided by the Ninth Circuit on 16 September 2026, found Copilot produces new works rather than stripping copyright management information, so the DMCA 1202(b) claims were dismissed; the breach of contract claims remain pending.
- The Apache Software Foundation's contributor agreement warrants that a contribution is the signer's original creation, which generated code cannot support.
- Codeberg voted 358 to 144 to stop hosting projects that are mostly AI-written, per It's FOSS on 14 November 2026.
- No court has ruled on whether an open source licence applied to purely AI-generated code is void, so that conclusion is an inference rather than a holding.
The Apache Software Foundation's contributor agreement asks every signer to warrant that a contribution is their own original creation. The foundation's generative tooling guidance then tells its projects that code written by an AI assistant is hard to accept under that warranty, because there is no settled answer to who the author is. Apache reached that position through paperwork, before the current argument started.
Jakub T. Jankiewicz published a post on 15 November 2026 titled "All AI-generated code should be Open Source". It drew 1,401 points and 704 comments on Hacker News. The claim is larger than the title suggests. Every AI-generated program should be released under an open source licence, or dedicated to the public under CC0, because the models were trained on scraped open source code and because the output has no author who can hold rights in it.
Both premises are sound. The conclusion does not follow, and the reason sits in a part of licensing the debate keeps skipping.
What Jankiewicz argued
His first argument is about inputs. Models were trained on open source repositories pulled from the internet, much of it under licences that require attribution or copyleft, and the output comes back carrying neither. He calls this licence laundering. The description is fair as far as it goes.
His second argument is about outputs. Purely AI-generated material is not copyrightable, he writes, citing the US Copyright Office and the monkey selfie case, so nobody can claim ownership of it. From there he concludes the code should be dedicated to the public or released under an open source licence.
He also notes that the United Kingdom grants a limited copyright to computer-generated works, which would change the picture there. This post does not build on that point, because the research behind it did not verify the current statute. Treat it as his observation rather than a settled comparison.
What the post does not offer is a mechanism. No disclosure rule, no named licence, no answer to what happens when a maintainer declines. The proposal is addressed to developers as an obligation of fairness, and fairness is not enforceable.
A licence needs a right to license
Open source licences are copyright licences. They do not exist as independent objects that attach themselves to files. Each is a permission granted by a rights holder, conditional on things the licensee must do.
MIT asks you to keep the copyright notice and the permission text. Apache 2.0 adds a NOTICE file, a requirement to mark modified files, and a patent grant. GPL requires that derivative works go out under the GPL. Every one of those obligations is triggered by a copyright and by a holder entitled to enforce it.
Remove the holder and the structure has nothing to rest on. If a file is genuinely uncopyrightable, there is no right for MIT to license, no notice for a licensee to preserve, and nobody with standing to complain when the terms are ignored. Copyleft cannot reach a work that was never protected, because it is a condition on copying a protected work rather than a property that floats free of one.
That is a logical inference from what a copyright licence is. No court has held that an MIT or GPL notice on purely AI-generated code is void, which is the point: everyone writing licence headers on generated code is relying on an untested theory.
Two consequences follow, and both cut against the proposal.
Public domain is already more permissive than any open source licence. If the code has no owner, anyone may use it, modify it, and ship it closed. Adding an MIT header grants nothing the law was not already granting, and it may grant less, since a licensee who relies on the header inherits conditions that may be unenforceable.
And no copyright does not mean no secrecy. An uncopyrightable file can sit in a private repository indefinitely. Nothing compels publication, and a licence cannot be imposed on code nobody received. Jankiewicz asks developers to do something they are already free to do, and offers them no reason that binds.
There is a third problem, in the training argument itself. The licence on the training data governs the copying of that data. It does not automatically govern the model's output, and treating the two as one step assumes the conclusion. The Ninth Circuit had something to say about that, and it did not say what either camp wanted.
The office drew the line at human control
The Copyright Office published Part 2 of its report on artificial intelligence on 29 January 2025, and it remains the most detailed statement of the US position. Its conclusions are narrower than either side usually admits.
Material generated purely by a machine is not copyrightable. Prompts alone do not give a user the kind of control over expression that authorship requires. Where a human selects, coordinates, and arranges generated material, the result can be protected, judged case by case.
That standard does real work. A repository is not one thing. A generated function with no human edits may carry no copyright at all. The selection of which functions to keep, the way they are wired together, an interface someone designed, and the comments explaining a choice can carry copyright in the arrangement even when no single line does.
It also makes the question unanswerable by inspection. A maintainer reading a pull request cannot tell whether the copyright line is accurate without knowing how the code was produced, and the developer who produced it often cannot say either. A commenter on the Hacker News thread put the practical version of this: you do not find out whether your code is protected until someone takes it to court.
Two rulings, two questions
Thaler v. Perlmutter, decided by the D.C. Circuit on 18 March 2025, settles the narrow question. The Copyright Act's use of "author" means a natural person. A machine cannot be one, so the registration refused in that case stays refused.
Doe v. GitHub, decided by the Ninth Circuit on 16 September 2026, gets cited for much more than it holds. The court found that Copilot produces new works rather than reproducing the plaintiffs' code with copyright management information stripped out. On that basis the claims under section 1202(b) of the DMCA were dismissed. The breach of contract claims were not, and remain pending.
| Case | Court and date | Question | What it decided |
|---|---|---|---|
| Thaler v. Perlmutter | D.C. Circuit, 18 March 2025 | Whether a machine can be an author | No. Only natural persons can be |
| Doe v. GitHub | Ninth Circuit, 16 September 2026 | Whether Copilot reproduced code with copyright management information removed | No. It generates new works, so the 1202(b) claims failed |
Read the second row closely. It says the output is new. It does not say the output is uncopyrightable, and it does not say that training on GPL code imposes no obligations downstream. The court was not asked to decide that and did not. Anyone citing Doe v. GitHub for the proposition that copyleft dies in training is quoting a holding that does not exist.
What the projects already decided
While the argument runs, the institutions holding the most copyrightable code have been making their own calls, and most have gone the other way.
Apache's problem is the contributor agreement. An ICLA signer warrants original creation, and generated code breaks that warranty in a way the signer cannot cure afterwards. The foundation's responsible AI material sets out the same problem for member projects.
Creative Commons Open Source went further on 1 December 2025 and said it would not accept AI-generated code contributions to its projects.
Codeberg, the German non-profit forge, voted 358 to 144 to stop hosting projects that are mostly AI-written, in a poll reported by It's FOSS on 14 November 2026. The same report says cURL ended its bug bounty after AI-generated submissions made the reports unusable.
Debian's general resolution proposals included one that would ban AI-generated code from the distribution, reported by XDA Developers on 25 July 2026, with copyright, quality, community, and ethics given as the reasons.
Red Hat's engineering write-up on bots in open source, published in March 2025, made a related point: the review load and the provenance problem arrive before any legal question does.
None of these bodies argues that AI code is unlicensed. They argue that they cannot establish where it came from, and provenance is what every licence warranty depends on. Apache cannot accept a contribution whose author is unknown, because the ICLA is a statement about authorship and there would be nothing behind the signature.
That is the practical shape of the problem Jankiewicz names. He is right that the arrangement is unfair to the commons. His remedy asks institutions to accept exactly the provenance they have spent two years rejecting.
Where a licence still does work
Ask instead where a licence does anything at all, and the answer is narrower than either camp wants.
Where a developer directed the output, edited it, chose among alternatives, and structured the result, there is a human contribution the Copyright Office says can be protected, and a licence covers that contribution. The protection may be thin. It may reach the arrangement rather than the functions. It is still a right, and a right is what a licence needs.
Practically, that puts weight on records. A commit showing a person writing, rejecting and rewriting generated code is evidence of authorship. A commit that pastes a generated file with no edits is evidence of the opposite. Almost nobody keeps the difference, and no rule requires them to, which is the gap in the proposal that no licence can close.
Where this could be wrong
The central legal claim here, that open source licences are nullities when applied to uncopyrightable output, is an inference and not a holding. No court has ruled on whether an MIT or GPL licence can attach to purely AI-generated code. If one does, and finds some contractual effect in the notice even without copyright, part of this argument weakens.
Whether training on GPL-licensed code makes model output a derivative work is untested. The research behind this post records Finnegan and other firms flagging it as open. Doe v. GitHub did not reach it, and nothing here should be read as a finding either way.
The Copyright Office analysis is US law. Jankiewicz's point about UK protection for computer-generated works was not checked against the current statute, and no other jurisdiction was surveyed, so the conclusion may not travel.
The institutional evidence is uneven. Apache and Creative Commons are documented in their own published policies. The Codeberg vote, the cURL decision, and the Debian proposal come through It's FOSS and XDA Developers rather than the primary minutes, so the numbers and the framing are those outlets' characterisation.
The Hacker News figures and quotations are second-hand in part. The point count is from the thread itself, but the commenter's line about finding out in court is relayed rather than independently verified against the original page.
The case-by-case standard means this post cannot tell any developer whether their own code is protected. That is a limit of the standard rather than a gap in the reporting, and it is the reason the proposal is hard to implement even for someone who agrees with it.
On this page
Sources
- All AI-generated code should be Open SourceJakub T. Jankiewicz, 2026
- Hacker News discussion of the Jankiewicz postHacker News, 2026
- Copyright and Artificial Intelligence, Part 2: CopyrightabilityU.S. Copyright Office, 2025
- Thaler v. Perlmutter, No. 23-5233U.S. Court of Appeals for the D.C. Circuit, 2025
- Doe v. GitHub, Inc., No. 24-7700U.S. Court of Appeals for the Ninth Circuit, 2026
- ASF Generative Tooling GuidanceApache Software Foundation, 2026
- Avoiding generative AI development toolsCreative Commons, 2025
- Open Source's Anti-AI Sentiment Has Real ReasonsIt's FOSS, 2026
- As Ubuntu embraces AI, Debian discusses banning all AI-generated codeXDA Developers, 2026
- IP Issues With AI Code GeneratorsFinnegan, 2023
- When bots commit: AI-generated code in open source projectsRed Hat, 2025
- Ethics, Copyright, and Licensing in AI-Assisted DevelopmentCarnegie Mellon University, 2026
Frequently asked questions
Is AI-generated code copyrightable?
Purely AI-generated material is not copyrightable in the United States. The Copyright Office's Part 2 report of 29 January 2025 says prompts alone do not amount to the control over expression that authorship requires, and that material a human selects, coordinates or arranges can be protected case by case. A file with human edits may carry protection for those edits while the generated portions carry none.
Can you put an open source licence on AI-generated code?
You can attach the text, but it may do nothing. MIT, Apache 2.0 and GPL all work by a rights holder granting permissions under conditions. If no one holds copyright in the code, there is no permission to grant and no one with standing to enforce the conditions. That reasoning is an inference from how copyright licences work. No court has ruled on it.
What did Doe v. GitHub decide about Copilot and GPL code?
The Ninth Circuit decided on 16 September 2026 that Copilot produces new works rather than reproducing the plaintiffs' code with copyright management information removed, so the claims under section 1202(b) of the DMCA failed. The court did not decide whether training on GPL-licensed code makes output a derivative work, and it did not decide that copyleft imposes no obligations. Breach of contract claims are still pending.
Why does Apache reject AI-generated contributions?
The Apache Individual Contributor License Agreement requires a signer to warrant that a contribution is their own original creation. Generated code breaks that warranty, and the foundation's generative tooling guidance tells projects to treat AI-assisted contributions as a case to escalate. Apache is not saying the code is unlicensed. It is saying it cannot establish who the author is.
Does public domain code have to be released?
No. Code without copyright protection can stay in a private repository indefinitely. Copyright law grants rights and limits them, but it does not compel distribution. That is the gap in the argument that AI-generated code should be open source: the absence of protection does not create an obligation to publish, and a licence cannot be imposed on code nobody received.
What counts as human authorship in AI-assisted code?
The Copyright Office looks for selection, coordination and arrangement that a person controlled. Directing the output, choosing among generated alternatives, editing the result and structuring how parts fit together can qualify. The test runs per work and per case, so it cannot be applied to a repository as a whole, and a maintainer reading a pull request usually cannot apply it at all.