Ninth Circuit Holds Copilot Creates New Code, Not Stripped Copies
The Ninth Circuit issued its opinion in Doe v. GitHub, Inc., No. 24-7700, on September 16, 2026, affirming the district court’s dismissal of claims that Copilot and OpenAI’s Codex violated § 1202(b) by reproducing plain…
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
The Ninth Circuit issued its opinion in Doe v. GitHub, Inc., No. 24-7700, on September 16, 2026, affirming the district court’s dismissal of claims that Copilot and OpenAI’s Codex violated § 1202(b) by reproducing plaintiffs’ code without its author attributions, copyright notices, and license terms. The Ninth Circuit held that plaintiffs’ own description of how GitHub Copilot works defeated their claim. A tool “that does not look up and reproduce stored work but rather creates new work” cannot strip or remove CMI from a copy that never existed. The decision is among the first federal appellate rulings applying the CMI rules of § 1202(b) to generative AI output.
Plaintiffs are anonymous programmers who published code in public GitHub repositories under open-source licenses that require attribution. Copilot and OpenAI’s Codex are large language models trained on billions of lines of publicly available code, including public GitHub repositories. Plaintiffs alleged that Copilot sometimes emits “essentially verbatim” copies of their code stripped of attribution and license terms, in violation of the DMCA (17 U.S.C. § 1202(b)(1) and (b)(3)).
After two rounds of dismissals and amendments, the district court dismissed the § 1202(b) claims with prejudice, reasoning that § 1202(b) requires “identical” copies and that plaintiffs’ own examples described a “modified format,” “variation,” or “functional equivalent” of the licensed code.
On appeal, the Ninth Circuit started with ordinary dictionary meanings. To “remove” is to get rid of something; to “alter” is to change it. Both require doing something to credit information that is already attached to an existing work. One who creates a new work and fails to include CMI has not removed anything.
The statute is worded similarly, protecting author information attached to “copies” of a work, not to excerpts or works merely inspired by it. That distinction matters for generative AI, whose output is a new work rather than a copy.
The panel then declared “identicality” “something of a misnomer” and narrowed its role:
Plaintiffs’ own allegations sank the output theory. The complaint did not allege facts showing that CMI was removed from, or altered in, copies of plaintiffs’ code. It described a complex probabilistic process that infers statistical patterns and predicts the most likely completion, which generates new works, rather than removing CMI attached to an existing one.
The panel also drew an express line at search engines, which retrieve and display stored copies. Had Copilot worked like a search engine and produced CMI-free identical output, the claim may have been stronger.
The closing passage of the court’s decision is one future defendants may quote. Many copyright cases involve creating a substantially similar work without attribution. If that alone violated § 1202(b), the DMCA “would supplant traditional copyright protections and subject defendants to potentially ruinous liability” — up to $25,000 per violation under § 1203(c)(3)(B), compared with $30,000 per work under § 504(c)(1). The court expressed no view on whether Copilot’s output infringes copyright but declined to convert run-of-the-mill infringement claims into DMCA claims.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。