Confirmed in a primary source 16 Sept 2026 Courts

A federal appeals court threw out the first copyright case brought against an AI coding tool.

On September 16 a three-judge panel of the Ninth Circuit unanimously affirmed the dismissal of the claims at the centre of Doe v. GitHub, the 2022 class action brought by open-source programmers against GitHub, Microsoft and OpenAI over Copilot and Codex.

The programmers argued the tools stripped out the copyright notices attached to their code, which would have triggered a section of the Digital Millennium Copyright Act carrying damages of up to 25,000 dollars per violation; the court held that a model producing new text is not removing information from an existing copy. The caveat is that this kills one legal theory, not the case: the claim that the tools breached open-source licence terms is still alive in the district court, and the panel expressly left ordinary copyright infringement claims open to anyone who can prove them output by output.

Until now

Section 1202(b) of the DMCA was the open route to mass statutory damages against AI developers, contested only in district courts that had split on it.

Now

Treat the copyright-notice theory as closed at appellate level in the circuit where most AI companies sit, and expect plaintiffs to be pushed back onto ordinary infringement claims they must prove output by output.

Four years ago a group of anonymous programmers sued GitHub, Microsoft and OpenAI over Copilot, the tool that suggests code. Their strongest claim was not that the companies had copied their work. It was that the tools had removed the licence headers from it, which would have engaged a provision of the Digital Millennium Copyright Act that pays 2,500 to 25,000 dollars per violation whether or not anyone proves a loss. Across millions of repositories that arithmetic is why the case mattered. On September 16 the Ninth Circuit ended it. Judge Eric Miller, writing for a unanimous panel, held that a model producing a new suggestion is not removing information from an existing copy: the AI tools create new works that never contained that information.

The reasoning is narrow and the consequence is not. The panel said in terms that it would not let plaintiffs transform run-of-the-mill copyright infringement claims into DMCA claims, and named the reason: the liability that reading would create is potentially ruinous. It also found the programmers had standing, resting on research showing models do sometimes reproduce their training data word for word, and it left them free to bring the ordinary infringement claim instead. That claim is harder. It requires showing a particular output copied a particular work, which is a case about individual outputs rather than a class action about a corpus. The licence claims survive in the district court and the plaintiffs may seek rehearing. But this is the first appellate word on any of it, and it came from the circuit that contains almost every company being sued. Treat the copyright-notice theory as closed at appellate level in the circuit where most AI companies sit, and expect plaintiffs to be pushed back onto ordinary infringement claims they must prove output by output.

How it got here

Nov 2022
Anonymous open-source programmers sue GitHub, Microsoft and OpenAI over Copilot and Codex, the first copyright case against OpenAI over a generative product.
Sep 01
The Justice Department tells a federal court that training AI on copyrighted writing is fair use. Carded in Edition 009.
Sep 04
OpenAI, Microsoft and the news publishers all ask one judge to end the largest AI copyright case without a trial. Carded in Edition 013.
Sep 11
The magistrate judge running discovery in that case discloses she had owned stock in a defendant. Carded in Edition 019.
Sep 16
The Ninth Circuit affirms dismissal of the section 1202(b) claims in Doe v. GitHub, unanimously and in a published opinion.
Who has to react
Every plaintiff firm holding a DMCA notice-stripping claim against a model developer, including those in the consolidated news publisher case; the Joseph Saveri firm, which brought this one and now chooses between rehearing and the harder infringement route; GitHub, Microsoft and OpenAI, whose largest single exposure theory in this case is gone; the open-source maintainers whose licences depend on attribution and who have just been told attribution is not what section 1202 protects here; district judges in California with pending 1202(b) claims, who are now bound; and the Copyright Office, which has been asked for years to say whether this provision reaches model outputs.

The evidence

ObservedWednesday September 16, 2026. The United States Court of Appeals for the Ninth Circuit issued a published opinion in Doe v.

The whole observation, as recorded

GitHub, Inc., number 24-7700, posted to the court's own opinions store at cdn.ca9.uscourts.gov under the date September 16, 2026. The panel was Circuit Judge Eric Miller, Senior Circuit Judge Sidney Thomas and District Judge Stanley Blumenfeld Jr. sitting by designation, and it was unanimous. The claims at issue arose under section 1202(b) of the Digital Millennium Copyright Act, which forbids the removal of copyright management information, the author and licence notices attached to a work. Judge Miller wrote that the AI tools create new works that never contained that information, and that the court would decline plaintiffs' invitation to transform run-of-the-mill copyright infringement claims into DMCA claims, noting the exposure such a reading would create was potentially ruinous. The panel found the plaintiffs did have standing, resting in part on academic research showing that language models sometimes reproduce training data verbatim. The claims that Copilot and Codex breached the terms of open-source licences were not before the panel and remain in the Northern District of California. The case was filed in November 2022 and is the first copyright case brought against OpenAI over a generative product. The surrounding record in this brief: the Justice Department told a federal court on September 1 that training on copyrighted writing is fair use, carded in Edition 009; the parties in the consolidated news publisher case asked one judge to end it without trial on September 4, carded in Edition 013; the magistrate judge running discovery in that case disclosed a stock holding on September 11, carded in Edition 019; Suno replaced its models with licensed ones on September 9 and Universal licensed its catalogue to ElevenLabs on September 10, both carded. Fact discovery in the Suno case closes on September 30.

Established

The Ninth Circuit issued a published opinion in Doe v. GitHub, Inc., number 24-7700, on September 16, 2026, available in the court's own opinions store.
The panel unanimously affirmed dismissal of the section 1202(b) DMCA claims, holding that the tools create new works that never contained the copyright management information.
The panel found the plaintiffs had standing, relying in part on research that language models sometimes reproduce training data verbatim.
The open-source licence claims were not decided and remain pending in the Northern District of California.

Not established

That training on copyrighted work is lawful. The panel decided a notice-removal provision, not fair use, and said so.
That the case is over. The licence claims survive, rehearing en banc and certiorari remain available, and ordinary infringement claims were expressly left open.
That other circuits will agree. This binds the Ninth Circuit only, and the largest AI copyright case is being litigated in New York.
That damages exposure has fallen. It has moved: from fixed sums per file to proof-heavy claims about individual outputs, which is a different shape of risk, not the absence of one.

Why it weighs this much

Signal strength

CostHigh

CommitmentHigh

ReversibilityHard

PersistenceYears

Cost is high in the sense the method measures: a published appellate opinion takes a panel, a record, oral argument and a written rationale that binds the judges to it, and it was produced against the plaintiffs' interest in a case they had litigated for four years. Commitment is high because a published opinion is precedent in every district court in the circuit from the day it issues. Reversibility is hard: it takes rehearing en banc or the Supreme Court, and neither is common. Persistence is years, because appellate copyright rulings set the pleading standard for every case filed after them. These are the strongest bars on today's deck and they are earned by the instrument, not by the drama.

What to watch

strengthens A district court in the Ninth Circuit dismisses a section 1202(b) claim against a model developer citing this opinion, or the Supreme Court denies certiorari, by June 30, 2027.
weakens The opinion is vacated on rehearing en banc, or another circuit holds that section 1202(b) reaches model outputs, by June 30, 2027.

The words

Section 1202(b) of the DMCA: A provision making it unlawful to strip the author, title and licence information attached to a copyrighted work. It was attractive to plaintiffs because it carries fixed damages per violation without proof of loss.
Copyright management information: The notice attached to a work saying who made it and on what terms. In source code it is usually the licence header at the top of the file.
Statutory damages: A sum set by law rather than proved in court. Section 1202 allows 2,500 to 25,000 dollars per violation, which is why a class action over millions of files was worth bringing.
Published opinion: An appellate decision that binds every district court in the circuit. The Ninth Circuit covers California, where OpenAI, Anthropic, Google and Meta are headquartered.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.