What the programmers said Copilot does
Microsoft owns GitHub, the largest host of public source code. Copilot is a paid service GitHub built with OpenAI on a modified Codex model, trained, the complaint says, on billions of lines of public repositories. Much of that code is published under licenses that allow reuse only if the copy keeps the author's name and the license text. The named plaintiffs are anonymous programmers who posted copyrighted work on public repos. Their complaint includes examples in which Copilot produced "essentially verbatim" snippets of their code and omitted the copyright management information, or CMI, that the DMCA protects: titles, author names, copyright notices, and terms of use.
Section 1202(b) makes it unlawful to intentionally remove or alter CMI, or to distribute copies knowing the CMI was stripped. Congress passed it in 1998 to keep digital files from traveling without their labels. The programmers argued that a Copilot suggestion which matches training data and arrives naked is the same crime as reprinting a book after tearing out the title page.
Standing without a remedy
Microsoft and OpenAI told the panel the risk that Copilot would emit these particular plaintiffs' code was conjecture. Miller disagreed, at least at the pleading stage. The complaint cites research on models that memorize training data, GitHub's own filter that can block verbatim snippets of 150 characters or more, and examples of the named plaintiffs' code coming back out. That was enough for Article III. A substantial risk of injury got them in the door. It did not get them a DMCA claim.
The merits holding is the sentence that will travel. Copilot, Miller wrote, "does not look up and reproduce stored work but rather creates new work." That new work "may or may not infringe plaintiffs' copyrights in their code," yet it "cannot reasonably be described as a copy of that code from which CMI has been removed or altered." Removal and alteration, on this reading, are affirmative acts against information already attached to an existing copy. A system that generates a fresh string never carried the notice, so it cannot have taken the notice off.
The district court had called this an "identicality" requirement. Miller treated the word as a gloss, not a separate element. Exact copies with the credit missing can be strong evidence that someone stripped CMI. Cosmetic changes will not save a defendant who reproduced the work and deleted the label. The Copilot complaint, the panel said, alleged something else: a probabilistic machine that predicts the next tokens from patterns in a corpus. Sometimes those tokens match a training snippet. That match, by itself, is not the defacing of a title page.
The training theory they left on the table
Plaintiffs had a second story, which the panel called the input theory: that defendants stripped CMI from class members' code before feeding it into training. Miller refused to hear it. Judge Tigar had said, on the record, that the complaint was not about training. Plaintiffs' counsel, asked whether copying into the model violated attribution clauses, answered, "Perhaps it doesn't." The panel held they forfeited the argument by failing to correct the court.
That forfeiture is the quiet disaster in the opinion. The output theory asked the DMCA to police a chatbot's completions. The input theory would have asked it to police the construction of the corpus. One is a product claim. The other is an industrial claim. They argued the first with enough force to get an interlocutory appeal, then watched the second get treated as a misunderstanding they never cleaned up.
Blake Brittain reported the decision for Reuters the same day. Attorneys for the developers, and spokespeople for OpenAI and Microsoft, did not immediately comment.
What the DMCA will not become
Miller's policy paragraph is the one corporate counsel will quote. Many copyright cases already involve a new work that is substantially similar to an old one and arrives without credit. If that were enough for section 1202(b), the DMCA would "supplant traditional copyright protections" and expose defendants to statutory damages of up to $25,000 per violation, against a $30,000 cap per work under ordinary copyright. The panel declined the invitation.
| Claim | Status after September 16 |
|---|
| DMCA § 1202(b), output theory | Dismissed; affirmed |
| DMCA § 1202(b), input / training theory | Forfeited on appeal |
| Open-source license / contract | Still live before Judge Tigar |
| Copyright infringement (substantial similarity) | Expressly undecided |
The remaining contract claims are the original bargain: take the code, keep the name. Copilot's design takes statistical structure and returns a suggestion. Whether that suggestion, on a given prompt, is a derivative that still owes the license is a different lawsuit than the one the panel just ended. Traditional infringement is a third lawsuit, and the court said it was not deciding it.
The briefs that split the library
The caption is a programmers' case. The amicus list is a map of every other industry that wants this holding, or fears it. The Authors Guild, the Association of American Publishers, the News/Media Alliance, and a cluster of scientific publishers told the court they had an interest in how section 1202 treats generated output. The Electronic Frontier Foundation and Public Knowledge filed on the other side. So did the Authors Alliance and a group of intellectual-property professors from Berkeley. ACT The App Association, Chamber of Progress, and the Computer & Communications Industry Association showed up for the platforms.
That split is the story under the story. Publishers want the DMCA's per-violation damages because proving that a model memorized a particular news article, then emitted it, is expensive. Civil-liberties groups want the statute to stay a law about stripping labels from copies, because stretching "remove" to cover generation would turn every autocomplete into a potential $25,000 event. Miller chose the narrower verb. He also left the publishers a breadcrumb: if a generator functions like a search engine that retrieves a stored file and peels the credit off, the claim looks stronger. Copilot, as pled, was not alleged to work that way.
Matthew Butterick, the type designer and lawyer who helped bring the case, has spent four years arguing that GitHub's own licenses are the point. The panel's opinion is almost allergic to that point. Licenses live in the contract counts. The DMCA, Miller said, is about CMI on copies. GitHub's 150-character duplicate filter is an admission that verbatim emission happens. The panel used that admission for standing. It refused to treat the same facts as a 1202(b) violation. The company built a switch that can hide the copies. The court still refused to treat the copies as copies for DMCA purposes.
A social contract the statute will not police
Open source was never a gift with no string. The string was attribution. GitHub made that string cheap to attach and easy to scrape. Copilot made it possible to consume the corpus as probability rather than as a stack of files. The programmers tried to make the DMCA do the work of the licenses because the DMCA's damages are scarier and because proving infringement, snippet by snippet, against a model that rarely emits a whole file is miserable.
The Ninth Circuit told them the 1998 statute is still a law about labels on copies. A completion that looks like your function is, in this court's mouth, a new object that happens to resemble you. Resemblance goes to copyright. After Wednesday, the missing byline is no longer a CMI problem.
That holding will leak into every other training fight that tried to ride the DMCA past the harder questions of fair use and substantial similarity. Authors, newsrooms, and record labels have already been pointing at the same section. Doe does not bind those cases on the facts. It does tell district judges in the West what "remove" means when the defendant is a generator.
The programmers can still try to enforce the licenses as contracts. They can still try to prove that particular outputs are copies in the ordinary sense. What they cannot do, in this circuit, is treat the missing byline on a Copilot suggestion as the destruction of a label that the suggestion never wore. The names are still in the training set. The statute, Miller held, does not follow them out.
Letters
0
No letters yet.