Your subcontract probably says the developer assigns you everything they write. It almost certainly does not say how much of it a person wrote. That gap matters more than it did two years ago, because the US Copyright Office has now said in writing that copyright does not extend to purely AI-generated material, and that prompts alone do not make the person typing them an author.
This post reads the Office's own report and turns it into one clause and one habit for agencies that pass work to subcontractors. It is general information, not legal advice. Have your own counsel finalize anything you sign.
What did the Copyright Office actually conclude?
On 29 January 2025 the Office published Part 2 of its report on copyright and artificial intelligence, titled Copyrightability. Its summary of conclusions includes three that matter here. Copyright "protects the original expression in a work created by a human author, even if the work also includes AI-generated material." Copyright "does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements." And whether human contributions are enough to count as authorship "must be analyzed on a case-by-case basis."
On prompts specifically, the report says: "given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output." The Office's reasoning is that prompts function as instructions that convey unprotectible ideas, and that even highly detailed ones do not control how the system generates the result.
Note what the Office did not say. It did not say AI-assisted work is unprotectable. It says using AI tools to assist rather than stand in for human creativity does not affect the availability of protection for the output.
Did a court say the same thing?
On human authorship, yes. In Thaler v. Perlmutter (No. 23-5233) the D.C. Circuit decided the case on 18 March 2025, affirming the refusal to register a work attributed to an AI system. The Supreme Court's docket for the petition (No. 25-449) shows certiorari denied on 2 March 2026.
A denial is not an endorsement of the reasoning, and Thaler tells you nothing about how much human involvement is enough for a mixed work. Treat it as confirmation that the floor is a human author, and the Office's report as the best current guide to what clears it.
Why does this land on the agency and not the subcontractor?
Picture a nine-person agency that sells a fixed-bid customer portal to a client and hands the build to a two-person subcontract firm. The subcontract assigns all code to the agency, and the master agreement assigns it on to the client with a warranty that the agency owns what it delivers. Nobody on that chain asked how the subcontractor wrote the code.
Say a good share of the data layer came out of a coding assistant with light editing. An assignment can only hand over rights that exist. If a module is purely AI-generated, there may be nothing in it to assign, and the agency has warranted ownership of something nobody owns. That does not make the code unusable. It means the client may have a weaker position against a copycat, and the agency's warranty may be broader than the facts.
The Office's registration guidance already points the same way. The report recounts that the 2023 guidance asks an applicant to disclose more than a de minimis amount of AI-generated material and to briefly describe the human author's contribution, and that the Office has registered hundreds of works that include AI material, with the registration covering the human contribution. A client's counsel who later wants to register or enforce will ask exactly that question. If your subcontractor never recorded the answer, you cannot give it.
Should you ban AI tools in the subcontract?
I would not. A ban is a rule your best subcontractors will quietly break, and then you have a breach nobody told you about. The report itself notes that programmers have used automated tools for refactoring and translating between languages for years, so the line between a tool and a substitute is not a bright one.
Disclosure is the better lever. You do not need to know that a model was used. You need to know where it was used and what a person did with the output. That is a short schedule, not a ban.
What should the clause and the habit be?
- 1.An AI-use disclosure with each delivery. The subcontractor lists which components, files or modules were generated in substantial part by an AI tool, and which tool. A line in the pull request template is enough.
- 2.A description of the human contribution. For anything generated, a sentence on what a person selected, rewrote or arranged. This mirrors what the Office asks of a registration applicant, so it is the record you would need anyway.
- 3.A warranty scoped to what is true. Warrant that the subcontractor has the right to assign what it delivers and has disclosed AI-generated material, rather than warranting that every line is independently copyrightable.
- 4.Review history you can read. Commits that show a person editing generated code are better evidence than a statement saying so. Ask for the repository, not a report about it.
- 5.Pass it down. If your master agreement with the client makes ownership promises, make sure the subcontract gives you what you need to keep them.
Copyright is not the only protection. Confidentiality terms, trade secret law and the contract itself still cover code that copyright may not. We covered the confidentiality side in the DTSA whistleblower clause check, and the ownership paperwork in why work made for hire is the wrong clause for a subcontractor. Whether the person doing the work is a real business or a disguised employee is a separate question, which the California B2B exemption checklist handles.
When this clause is the wrong tool
If the deliverable is a throwaway internal script, an AI-use schedule costs more than the exposure, so skip it. If the client's own contract already forbids AI-generated code, disclosure does not fix that, and you need an outright prohibition passed down. And if your concern is a specific licence or training-data claim, this is the wrong post: the Office's Part 3, on generative AI training, was released in pre-publication form on 9 May 2025 and says a final version will follow, so read the current version with counsel rather than relying on a summary.
If you want overflow capacity from a partner who will show you how the work was built and put its terms in writing, see how our agency partner arrangement works, and ask any partner you consider, including us, to show you the exact clause. If you have a scope in mind, tell us about it.
