The Ninth Circuit has affirmed, in a published interlocutory opinion filed 16 September 2026, the dismissal of the DMCA section 1202(b) claims in Doe v. GitHub, holding that on the complaint as pleaded Copilot creates new works rather than removing or altering copyright management information from copies of existing works, while the contract claims remain pending

What the court decided, and what it left alone

The panel affirmed the Northern District of California's dismissal of the plaintiffs' claim under 17 U.S.C. section 1202(b). The opinion's own statement of the holding, on page 6, is that the allegations show Copilot and Codex do not remove or alter copyright management information from a copy of an existing protected work but instead create new works that never contained that information. The word AFFIRMED closes the opinion on page 18.

The posture matters as much as the result. This is an interlocutory appeal under 28 U.S.C. section 1292(b). Page 9 records that the district court certified one controlling question: whether sections 1202(b)(1) and (b)(3) of the DMCA impose an identicality requirement. Page 7 records what is left of the case after two rounds of dismissals and amendments: one DMCA claim and two claims for breach of contract. Page 9 states that the district court denied a motion to dismiss the contract claims, which remain pending.

So the lawsuit has not ended. What ended, on this opinion, is the section 1202(b) output theory as pleaded in the operative complaint against GitHub, Microsoft and the OpenAI entities named on pages 1 and 2. The opinion does not adjudicate the liability of any other defendant, any other tool, or any user of Copilot.

The input theory was forfeited, not decided

Page 9 sets out two theories. The input theory is that defendants violated section 1202(b)(1) at the training stage by removing CMI from class members' code before feeding the stripped code into Copilot as training data. The output theory is that Copilot sometimes returns memorized training data to users without the source code's CMI.

On the input theory the court did not reach the merits. Pages 10 and 11 hold that plaintiffs forfeited it. The opinion quotes the district court's observation at the first motion to dismiss hearing that the complaint is not about training, and the district court's later written statement that plaintiffs do not allege they were injured by defendants' use of licensed code as training data. It then finds that plaintiffs identified nothing in their later district court briefing that would have put the court on notice that they wished to pursue an input theory.

That is a procedural ruling. Whether a properly preserved training-stage section 1202(b) claim could succeed is not answered anywhere in the opinion, and nothing in it approves or disapproves the use of licensed code as training data.

Standing survives on a substantial-risk allegation, at this stage

Defendants argued that plaintiffs lack standing because the risk that Copilot will reproduce their code, and not some other contributor's, is conjecture. Page 12 says: We disagree. The court found the complaint alleges a substantial risk of injury, citing its allegations that Copilot reproduces the named plaintiffs' code without attribution and that verbatim, near-verbatim or modified copies of each plaintiff's code have likely already been emitted.

Page 13 lists what made those allegations plausible: academic research cited in the complaint that large language models sometimes emit memorized training data verbatim; GitHub's duplicate-detection feature, which the complaint says lets users block output matching public code in verbatim snippets of 150 characters or more; and examples in the complaint of Copilot reproducing portions of the named plaintiffs' code verbatim. All of these are allegations taken as true at the pleading stage, not findings of fact.

The court set the boundary itself. Page 13 states that if the case were to proceed to summary judgment, plaintiffs would need to present evidence sufficient to create a genuine factual issue on substantial risk, and that the court need not determine whether the evidence referred to in the complaint would meet that standard.

Identicality is a gloss, not an element

The district court had reasoned that section 1202(b) claims require copies to be identical, and dismissed because plaintiffs' code and Copilot's output were not identical. Pages 8 and 9 quote that reasoning. The Ninth Circuit did not adopt it as stated.

Page 15 calls identicality something of a misnomer, because the DMCA does not require literal identicality between the plaintiff's work and the allegedly infringing work. The court reads identicality as a gloss on the statutory terms remove, alter and copies rather than an independent and atextual element. What the statute does require, on the court's reading of section 1202(b) and the section 1202(c) definition of CMI on page 14, is an affirmative act on CMI connected to a work that already exists. Creating a new work and failing to include CMI is not removal or alteration.

Comparison still has a role. Pages 15 and 16 explain that where two works are otherwise identical but the accused version omits the CMI, a factfinder may reasonably infer removal, as in the court's 2016 decision in Friedman v. Live Nation Merchandise. Page 16 adds that substantial reproduction without CMI will often be strong circumstantial evidence of removal, that defendants conceded the works need not be literally identical, and that minor cosmetic changes will not necessarily protect a defendant who substantially or entirely reproduces the protected work and removes CMI. A pleader who can show that pattern is not shut out by this opinion.

Why the Copilot allegations fail: generation, not retrieval

The court applied that framework to the complaint's own description of the tool. Page 17 summarises it: Copilot relies on a complex probabilistic process to predict the most likely solution to a given prompt, inferring statistical patterns governing the structure of code. That account, the court says, does not describe an action taken with respect to CMI attached to an existing work. It describes a process through which Copilot generates new works.

The contrast the court draws is with a search engine, which retrieves and displays stored copies. Page 17 states that if Copilot functioned like a search engine and produced outputs identical to plaintiffs' code without CMI, plaintiffs might have a stronger claim that Copilot had removed their CMI, but that is not what plaintiffs alleged. Page 17 also acknowledges that print-media examples such as defacing a title page do not map neatly onto emerging digital technologies like artificial intelligence, to which the DMCA's protections also apply.

Two things follow from how the court framed this. The holding turns on what these plaintiffs pleaded about this tool. And the opinion itself identifies an architecture, retrieval of stored identical copies, that would sit differently under section 1202(b). Neither point is an invitation we are adding; both are on page 17.

The damages logic and the infringement boundary

Page 18 explains why the court would not let attribution-free similarity alone carry a DMCA claim. Many copyright cases involve creating a work substantially similar to the plaintiff's without attribution. If that were all it took to violate section 1202(b), the court says, the DMCA would supplant traditional copyright protections and expose defendants to the DMCA's enhanced statutory damages. The opinion compares 17 U.S.C. section 1203(c)(3), permitting up to $25,000 per violation, with section 504(c)(1), which it describes as capping traditional copyright statutory damages at $30,000 per work. No damages of any kind were awarded in this opinion; the figures are cited as existing statutory provisions.

The final sentence before the disposition declines plaintiffs' invitation to transform run-of-the-mill copyright-infringement claims into DMCA claims. The court was equally explicit about what it was not deciding. Page 18 states that Copilot's output may in some cases be substantially similar to existing code, and that the court expresses no view on whether that similarity would allow plaintiffs to assert a claim for copyright infringement. Page 13 makes the same reservation: the new work may or may not infringe plaintiffs' copyrights in their code.

The opinion also records, on page 14, that the Fifth Circuit reached an accordant result on the failure-to-include point in Kipp Flores Architects v. AMH Creekside Development on 21 August 2026, and cites the Southern District of New York's 2025 decision in New York Times Co. v. Microsoft on page 16 for the cosmetic-changes point. We did not read either decision and report them only as the Ninth Circuit cites them.

What we did not verify

What we opened: the full 18-page slip opinion as served from the Ninth Circuit's datastore at the 2026/09/16 path, read end to end, including the caption on pages 1 and 2, the staff summary on pages 3 and 4, which states it constitutes no part of the opinion, the counsel and amici listing on pages 4 and 5, and the opinion of the court on pages 5 to 18. Every page reference in this piece is to the printed slip-opinion page.

What we did not open: the district court's dismissal orders and certification order, the operative complaint, the appellate briefs and amicus briefs, the Ninth Circuit docket, and the authorities the opinion cites, including Friedman v. Live Nation Merchandise, Kipp Flores Architects v. AMH Creekside Development and New York Times Co. v. Microsoft. We describe the complaint's allegations only as the opinion recounts them. We did not check whether a petition for rehearing has been or will be filed, or when the mandate issues.

What we refuse to claim: we do not say that AI-generated code is exempt from section 1202(b), because the holding is confined to these allegations and the court itself described a retrieval architecture that might support a stronger claim. We do not say the decision protects enterprise customers or users of Copilot: the opinion does not adjudicate enterprise-customer or Copilot-user liability. It mentions users and their prompts on pages 7, 9, 13 and 17, but no such party was before the court and no ruling on their liability appears. We do not say the court endorsed an identicality requirement, because page 15 rejects that formulation. We do not say the court approved training on licensed code or decided fair use, because the input theory was forfeited and no fair use question appears in the opinion. We do not say Copilot's output is original, copyrightable or non-infringing, because the opinion reserves the infringement question twice. We do not say the lawsuit is over, because the contract claims remain pending. We do not say the court found that copying occurred, because standing was decided on plausible allegations. We give no view on the likelihood of further review. Where the word must appears in this piece, it is inside the court's own quoted words or describes a pleading standard the court itself stated.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• Bartz v. Anthropic 1.5 billion dollar copyright settlement
• Ninth Circuit on Amazon v. Perplexity, CFAA and AI agent access
• AI court rules tracker
• North America AI regulation tracker
• AI Regulation News hub