What the court actually decided
On September 16, 2026, the Ninth Circuit affirmed dismissal of the lead copyright claim in Doe v. GitHub, Inc., the long-running suit brought by anonymous programmers against GitHub, Microsoft, and OpenAI over Copilot and Codex. The plaintiffs' theory rested on Section 1202(b) of the DMCA, which bars intentionally removing or altering "copyright management information" (CMI) — an author's name, a license notice — from a copyrighted work. Their argument: when Copilot generates code resembling a plaintiff's snippet but without the original attribution or license terms, that omission is functionally a removal.
A three-judge panel — Circuit Judges Sidney R. Thomas and Eric D. Miller, joined by District Judge Stanley Blumenfeld Jr. sitting by designation, with Judge Miller writing — rejected that framing. As the Electronic Frontier Foundation summarized in its analysis of the ruling, the court held that "removing copyright information from a copyrighted work is fundamentally different from creating a new work that didn't have CMI in the first place" (EFF Deeplinks). Copilot's output, the panel found, is newly generated text — not a retrieved copy of an existing file from which CMI was stripped. Section 1202(b) reaches the latter, not the former.
Steelmanning the plaintiffs' theory
The programmers' argument was not frivolous. Attribution and license terms are how open-source developers get credit and how copyleft licenses (GPL, MIT, Apache) propagate their obligations forward. If a tool can reproduce someone's code near-verbatim while silently erasing the license that would otherwise bind the recipient, that's a real harm to the open-source commons — and one that ordinary copyright infringement law, keyed to "substantial similarity" of a whole work, may not squarely capture when outputs are short snippets mixed with other training data. DMCA Section 1202 also carries statutory damages of $2,500 to $25,000 per violation (Pub. L. 105-304) — far more punitive, on a per-instance basis, than the $30,000-per-work cap under ordinary copyright law, because a model that emits the same uncredited snippet thousands of times could in theory generate thousands of separate violations rather than one. Reporting on the case pegged the plaintiffs' theoretical exposure calculation at roughly $9 billion (ppc.land) — a number that illustrates why this statutory hook, if accepted, would have dwarfed ordinary infringement damages and made Section 1202 the preferred weapon in nearly every future AI-training dispute.
Why the court was right to draw this line
That is precisely the problem with the plaintiffs' theory: it would have converted a narrow anti-tampering provision, aimed at people who deliberately strip credit off an existing file to launder it, into a strict-liability regime triggered any time a generative model's output happens to lack the metadata a human editor would have included. Under that reading, any tool that transforms, summarizes, remixes, or paraphrases copyrighted material — not just LLMs, but search indexes, code formatters, or even a student quoting a paper from memory — could face $25,000-per-instance exposure merely for producing something that doesn't carry the original's byline. EFF's brief in the case flagged exactly this: accepting the theory would have created "a brand-new source of liability for otherwise perfectly lawful activities," threatening remixes, classroom use, and reverse engineering far beyond the AI context. Congress wrote Section 1202 to punish affirmative tampering, not to serve as a backdoor attribution mandate for every transformative use. As Gibson Dunn's analysis of the ruling put it, the panel's distinction turns on whether CMI was "attached to a work that already exists" versus never present in a newly created one (Gibson Dunn) — a textually faithful reading of a statute that Congress could always amend if it wants a different rule for AI outputs specifically.
What this does — and doesn't — settle
The ruling forecloses one specific statutory-damages theory in the circuit that hosts most generative-AI litigation. It does not touch whether Copilot's output infringes anyone's copyright on ordinary substantial-similarity grounds, and it does not resolve the separate "input theory" — whether GitHub stripped CMI during training itself — because the panel found plaintiffs had forfeited that argument on appeal rather than deciding it favors either side. Most importantly, the district court's separate breach-of-license claims, built on the theory that Copilot reproduces GPL- and MIT-licensed code without honoring the attribution and share-alike terms those licenses require, continue in the Northern District of California under Judge Jon Tigar. Open-source contributors angry about uncredited reproduction still have a live vehicle for that grievance — just a contract one, not a copyright-statutory one.
The bigger pattern
This fits a broader judicial instinct in 2026: courts are willing to let AI copyright disputes proceed on conventional infringement and contract theories, while resisting invitations to stretch adjacent statutes — CFAA, DMCA anti-circumvention, now 1202 CMI — into novel liability engines calibrated to AI's speed and statutory damages' size rather than to what Congress actually wrote. That's the right instinct. Proportionate regulation of AI training and output should come from copyright's existing substantial-similarity and licensing doctrines, or from Congress writing a purpose-built rule, not from courts improvising strict liability out of a narrow anti-tampering clause. The programmers behind Doe v. GitHub still have their contract claims. What they don't have anymore is a shortcut to nine-figure statutory damages for building a tool that writes new code.