Facts
GitHub Copilot and OpenAI’s Codex are code-writing tools built on large language models trained on billions of lines of public code, including code in public GitHub repositories. The plaintiffs are programmers who published copyrighted code on GitHub under open-source licenses that require attribution. In a putative class action against GitHub, Microsoft and OpenAI, they alleged that Copilot sometimes reproduces their code verbatim or nearly verbatim without the copyright notices and license terms, in violation of 17 U.S.C. § 1202(b). The district court dismissed that claim with prejudice, reasoning that the provision requires identical copies, and certified the question for interlocutory appeal. Two breach-of-contract claims remain pending below.
Question
The certified question was whether § 1202(b)(1) and (3) of the Digital Millennium Copyright Act require the defendant’s output to be identical to the plaintiff’s work. The panel also had to decide whether the plaintiffs had Article III standing, given that they could only allege a risk that Copilot would emit their particular code, and whether they had preserved an “input” theory that copyright management information (CMI) was stripped from code when it was used as training data. At the core was whether an AI tool that generates code without attribution “removes or alters” CMI.
Decision
In an opinion by Judge Miller, the panel held that the plaintiffs had standing: the complaint plausibly alleged a substantial risk that Copilot would reproduce their code, citing research on memorization and GitHub’s own duplicate-detection filter. It treated the input theory as forfeited and affirmed on the merits. “Remove” and “alter” imply an affirmative act on CMI attached to an existing copy, and someone who creates a new work and omits CMI has removed nothing. On the plaintiffs’ own account, Copilot learns statistical patterns and generates new code. Identicality is not an independent element, though substantial reproduction without CMI will often be strong circumstantial evidence of removal. The court added that a retrieval tool such as a search engine might be treated differently.
Why it matters
The ruling addresses at federal appellate level whether generative AI output can create liability under § 1202(b), which carries statutory damages of up to $25,000 per violation and has been invoked in other AI cases; the panel itself cited the district court ruling in New York Times v. Microsoft. The court expressed no view on whether such output infringes copyright and declined to turn ordinary infringement claims into DMCA claims. It confined its reasoning to the allegations before it, noting that systems that retrieve and display stored copies could be assessed differently.
Related stages
The two breach-of-contract claims continue in the Northern District of California. The panel also relied on the Fifth Circuit’s August 2026 decision in Kipp Flores Architects v. AMH Creekside Development. As of October 2026 no petition for rehearing en banc or for certiorari had been reported; the certiorari deadline had not yet passed.