-
Ninth Circuit Finds LLM Coding Agent Outputs Did Not Violate Digital Millenium Copyright Act by Not Including Attribution
09/29/2026In Doe v. GitHub, Inc., No. 24-7700 (9th Cir. Sept. 16, 2026) the United States Court of Appeals for the Ninth Circuit, on interlocutory appeal, affirmed a district court dismissal of claims under the DMCA related to LLM-based coding tools trained on open-source code.
GitHub Copilot and Codex are LLM-based tools that can write code. Among the materials on which these LLMs were trained were open-source software projects that were published on GitHub. The plaintiffs are a class of programmers who published copyrighted code on public GitHub repositories.
Section 1202 of the DMCA makes it unlawful to intentionally remove or alter CMI from copyrighted works, or to distribute works with removed or altered CMI. In this case, the plaintiffs asserted that the defendants violated the DMCA in two different ways. First, they argued that defendants removed the CMI from their copyrighted code prior to training the LLMs and that this violated the DMCA (the “input” theory). Second, they argued that defendants had violated the DMCA by distributing code generated by the LLMs that sometimes returned “memorized” training data to users without including the training data source code’s CMI, thereby removing or altering CMI on copies of their protected works (the “output” theory).
The district court had dismissed the “output” theory on the view that Section 1202 of the DCMA required the copies to be “identical.” The Ninth Circuit declined to adopt such a strict requirement, noting that minor changes would not protect a defendant from liability.
While it rejected the “identicality” argument, the Ninth Circuit affirmed the dismissal, determining that, as the AI tools generated new code, they did not distribute the copyrighted source code. The Ninth Circuit contrasted how LLMs operate from the manner in which search engines retrieve information. Because the generated code never contained the CMI, the court held that a newly created work that never contained CMI is not a copy from which CMI was altered or removed. “[P]laintiffs’ own allegations about Copilot show that it is best understood as learning from existing works and then creating new works based on that learning process, not as making copies of existing works.”
The Ninth Circuit declined to consider the “input” theory, holding that it had been forfeited at the district court level. Thus, the decision did not address whether their removal of such information from the training data violated the DMCA.
IP Litigation