Ninth Circuit Closes Off DMCA Claims Against GitHub Copilot and OpenAI Codex

A federal appeals court has shut off one legal path that rights holders were using against generative AI tools, ruling on September 16 that an AI model reproducing code or other material without the original author's credit does not, by itself, violate the Digital Millennium Copyright Act's ban on stripping copyright notices.

The Ninth Circuit's opinion in Doe v. GitHub, Inc., No. 24-7700, affirmed the dismissal of DMCA claims brought by anonymous open-source programmers against GitHub, Microsoft and OpenAI over the Copilot and Codex coding assistants. Judge Eric Miller wrote for a unanimous panel that also included Senior Circuit Judge Sidney Thomas and, sitting by designation, U.S. District Judge Stanley Blumenfeld Jr. of the Central District of California. The case reached the Ninth Circuit as an interlocutory appeal under 28 U.S.C. § 1292(b), after U.S. District Judge Jon Tigar of the Northern District of California certified the question for immediate review, according to the opinion.

Section 1202(b) of the DMCA bars intentionally removing or altering "copyright management information," or CMI — an author's name, a copyright notice, license terms — from a work. The programmers argued Copilot and Codex violate that section because their outputs reproduce open-source code without carrying its CMI along. The panel disagreed, holding that Copilot and Codex "create new works that never contained that information" rather than removing CMI from a copy of an existing work, so there is nothing existing from which CMI could have been "removed." As Judge Miller put it, according to TheNextWeb's account of the opinion: "One who creates a new work and fails to include CMI cannot be said to have 'removed' or 'altered' anything." The court also rejected the argument that outputs must be identical to the original code to trigger the statute, calling "identicality" a "misnomer" for the actual statutory test — but that did not save the plaintiffs' claim here.

The ruling turned on which of two theories the programmers actually pursued. According to the opinion, plaintiffs could have argued an "input theory" — that CMI was stripped from their code before it was fed into training — or an "output theory," that Copilot's outputs reproduce their code without its CMI. The panel held the programmers forfeited the input theory by never clearly asserting it in the district court, so it "decline[d] to consider" that question at all, per the opinion. It reached and rejected only the output theory, finding the complaint's own allegations describe Copilot as learning from code and generating new code, "not as making copies of existing" works, in the court's words.

That distinction carries financial weight. DMCA claims under 17 U.S.C. § 1202(b) can draw statutory damages of $2,500 to $25,000 per violation, assessed per act — a figure that scales fast against a model trained on millions of files. Ordinary copyright infringement claims, by contrast, cap statutory damages at $30,000 per work under 17 U.S.C. § 504(c)(1), aggregated regardless of how many times that work was copied. Cutting off the DMCA route removes the larger damages exposure while leaving the harder-to-win infringement claims, which require plaintiffs to show a defendant's output is "substantially similar" to the copyrighted work — a standard the court expressly declined to apply here, as Reuters reported.

None of that resolves whether training on copyrighted material without permission is itself lawful. The opinion says the court "expresses no view" on whether the similarity between Copilot's outputs and the plaintiffs' code could support an infringement claim, as Reuters reported, and Gibson Dunn's client alert stresses that the court did not decide whether removing CMI before training violates the DMCA either - it never reached the input theory that would have tested the training pipeline itself. Attorneys for the developers and spokespeople for OpenAI and Microsoft did not immediately respond to requests for comment, according to Reuters. Tigar's earlier order still lets the plaintiffs press breach-of-contract claims that Copilot and Codex violated the open-source licenses attached to their code, and those claims remain pending in the Northern District of California.

The court flagged the awkward fit between a 1998 statute written for stripped title pages and defaced photo captions and a technology Congress never anticipated, citing print-media precedent that "do not map neatly onto the emerging digital technologies like artificial intelligence." As Bloomberg Law framed it in the unpaywalled portion of its report, this is the first AI copyright ruling from a federal appeals court, and it narrows one claim without resolving the industry's central uncertainty over training data.

For anyone building or defending a generative AI model facing a similar suit, the lesson is pleading precision: alleging a model's output merely lacks a watermark or credit line is not the same claim as alleging it memorized and reproduces a specific copy of a specific work. Courts following this reasoning will likely require the latter — closer to a traditional infringement case — rather than accept the DMCA's CMI provision as a shortcut around it.

The programmers' contract claims proceed before Judge Tigar in the Northern District of California; no trial date has been set, and any appeal of this ruling to the full Ninth Circuit or the Supreme Court would be the next procedural step to watch.