Ninth Circuit: Copilot Writes New Code, So It Cannot Strip Copyright Notices
Photo: markusspiske / Pixabay
One of the claims most often added to AI copyright complaints in the United States has just become harder to win. In Doe v. GitHub, Inc., No. 24-7700, the US Court of Appeals for the Ninth Circuit affirmed the dismissal of claims that GitHub Copilot and OpenAI's Codex broke the Digital Millennium Copyright Act (DMCA) by producing code without the author names, copyright notices and licence terms that accompanied the original. The opinion, written by Judge Eric Miller for a panel with Judge Sidney Thomas and District Judge Stanley Blumenfeld sitting by designation, was filed on 16 September 2026. Fenwick published a client alert on the ruling on 16 October.
The claim and the district court
The plaintiffs are developers whose open source projects sit on GitHub under licences that require attribution. Their putative class action alleged that Copilot and Codex, both trained on large volumes of public code, sometimes emit their code, verbatim or with changes, without the attribution and licence information. They framed this as a breach of 17 U.S.C. § 1202(b), which prohibits intentionally removing or altering copyright management information (CMI) and distributing copies knowing the CMI has been removed.
District Judge Jon Tigar dismissed the DMCA claims on the basis that section 1202(b) requires the output to be identical to the original work, and the code examples offered were not. He certified the question for interlocutory appeal, which is how it reached the Ninth Circuit.
Standing survives
The plaintiffs won the first point. The defendants argued that the programmers had not shown any real risk that their own code would be reproduced. The panel disagreed. At the pleading stage, it found enough in the complaint to make a substantial risk of injury plausible, including research cited in the complaint on memorisation in large models, examples of named plaintiffs' code appearing in outputs, and GitHub's own filter that lets users block suggestions matching public code of 150 characters or more. The panel noted that the plaintiffs would need evidence, not allegations, at summary judgment.
The merits: no copy, no removal
The plaintiffs lost on the main question. The court read the statutory verbs "remove" and "alter" as requiring an act directed at CMI attached to a work that already exists. A plaintiff must therefore show that the defendant took CMI off a copy of the protected work. Pointing to a similar or derivative work that lacks the CMI is not enough.
The panel then took a different route from the district court on so called identicality. It said the term is misleading because the DMCA does not demand a literal match. Identity between the works is best treated as evidence: if a copy matches the original in everything except the missing notice, a factfinder can infer removal, as in the court's earlier Friedman v. Live Nation photograph case. Minor cosmetic edits will not save a defendant who copies a work almost entirely and drops the CMI. Material differences, on the other hand, point towards a new work that never carried the plaintiff's CMI.
Applying that to the complaint, the court relied on the plaintiffs' own description of Copilot as a probabilistic system that predicts likely code from patterns learned in training. On that description, Copilot generates new works instead of retrieving stored copies, so there is no copy from which CMI could have been removed. The court contrasted this with a search engine that returns stored material. If Copilot worked that way and returned identical code without the notices, the claim might have been stronger. The output may still infringe copyright, the court noted, but that is a separate question.
The plaintiffs also argued an "input" theory, that CMI was stripped from code when it was gathered for training. The panel held that this theory had been forfeited because it was not properly raised below, so the decision says nothing about it.
Why it matters
Section 1202 has been attractive to AI plaintiffs because it carries statutory damages per violation and does not require registration of the work. It appears in many pending suits against AI developers, from news publishers to music publishers. This ruling is binding in the Ninth Circuit, which covers California, where many of those cases are filed. Output based CMI claims will now need concrete examples of near verbatim reproduction without the notices. Claims about what happened during data collection remain open, and plaintiffs can be expected to plead them with more care.
For AI developers, the ruling supports technical measures that reduce verbatim output, since the closer an output is to an existing work, the easier it is to infer removal. Duplicate detection filters and attribution features look more valuable after this decision, not less.
The view from Indonesia
Indonesia's Copyright Law, Law No. 28 of 2014, also protects rights management information and prohibits removing or changing it without permission. No Indonesian court has applied those provisions to AI output. The Ninth Circuit's distinction between removing information from a copy and creating a new work that never carried it offers a useful framework if such a case arises here, although Indonesian courts are not bound by it and could take a different view of the statutory text.
Sources
Do you or your clients need trademark, patent, annuity or industrial design work in Indonesia? We handle filings, prosecution and renewals, with updates in English. Email us for a quote.
Contact UsIf you enjoyed this article on our portal, you can buy me a coffee via the PayPal link below.
PayPal
