Generative AI: Copying or Creating? The Ninth Circuit Limits DMCA Attribution Claims in Doe v. GitHub

Doe v. GitHub, Inc., No. 24-7700 (9th Cir. Sept. 16, 2026)[1]

Introduction

On September 16, 2026, the Ninth Circuit affirmed the dismissal of the DMCA claims that open-source programmers had brought against GitHub, Microsoft, and OpenAI over GitHub Copilot and Codex. The court held that even if these tools sometimes output code resembling their training data, they do not "remove or alter" copyright management information (CMI) from a copy of an existing work under 17 U.S.C. § 1202(b), because they “create new works that never contained that information.” The ruling is one of the first appellate decisions to address how DMCA attribution protections apply to generative AI, and it is likely to shape the many AI copyright cases now pending.

Background

The plaintiffs published code on GitHub under open-source licenses. Many of those licenses require that the author's name, copyright notice, and license terms travel with any copy or derivative. The plaintiffs alleged that Copilot sometimes emits verbatim or near-verbatim copies of their code without that attribution, in violation of 17 U.S.C. § 1202(b), part of the DMCA.[2] After the district court dismissed the claim for lack of "identicality," it certified the question for interlocutory appeal. The contract claims remain pending below.

Key Rulings

1. The "input" theory was forfeited. The plaintiffs argued that CMI was stripped at the training stage, but the court declined to consider this theory. Although the complaint alleged that CMI was removed when code was incorporated into Copilot, plaintiffs' counsel conceded that training "[p]erhaps" did not violate the licenses, and plaintiffs later pointed only to a single remark in their briefing and an OpenAI filing, neither of which put the district court on notice. The court found that plaintiffs never corrected the district court's repeated view that "the complaint is not about training.” As a result, the training-stage question remains open.

2. The plaintiffs have standing. The court found a plausible "substantial risk" of injury based on the following three things: (1) research showing that large language models memorize training data; (2) GitHub's own filter for duplicate output; and (3) concrete examples of Copilot reproducing the named plaintiffs' code.

3. The "output" theory fails on the merits. The court read "remove" and "alter" in DMCA § 1202(b) as requiring an affirmative act on CMI attached to a work that already exists. On that reading, someone who creates a new work and simply leaves CMI out has not removed or altered anything. The court said "identicality" is a misleading label rather than a separate element of the claim. Near-identity between two works is just circumstantial evidence that CMI was removed, so minor cosmetic changes will not shield a defendant who copies wholesale.[3] Based on the plaintiffs' own description, though, Copilot infers statistical patterns and generates new code. It is not like a search engine that retrieves stored copies.

4. Infringement remains open. The court expressly declined to decide whether Copilot's output infringes copyright. It also warned that treating every substantially similar work without attribution as a DMCA violation would let § 1202 "supplant traditional copyright protections." That would expose defendants to statutory damages of up to $25,000 per violation.

US Implications

For AI developers, the decision is a significant but narrow win. It removes a route to large per-violation DMCA damages for output that merely resembles training data. However, it does not resolve infringement or fair use. The decision leaves several important questions for future litigation: (1) whether retrieval-based systems that function more like search engines and return stored copies without CMI violate the DMCA; (2) whether DMCA liability may arise under the input theory based on training; (3) whether AI outputs infringe copyright under traditional substantial-similarity analysis; (4) whether fair use protects training and/or outputs; and (5) the scope of the plaintiffs' remaining contract claims.

Comparative Perspective

The question before the Ninth Circuit differs fundamentally from the central issue currently debated in Europe. In the EU, the central question is whether using works for training is lawful under the text-and-data-mining exceptions in Articles 3 and 4 of the DSM Directive (2019/790), including a rightsholder's machine-readable opt-out, together with the AI Act's copyright and transparency duties for general-purpose AI models (Art. 53). In Germany, GEMA v. OpenAI (LG München I) [4] held that memorizing works in a model during training, and reproducing them in outputs, falls outside the TDM exception. Doe v. GitHub, by contrast, is a textual reading of a single conduct-focused provision that asks whether CMI was "removed.” In short, the US court focused on a specific act of removing CMI, while the EU focuses on whether the use of training data was lawful.

The two systems also treat attribution differently. European law grounds attribution in the author's moral rights, which are not limited to stripping a notice from an existing copy. Under that approach, AI output that reproduces recognizable expression without credit may face claims that the Ninth Circuit's reading of § 1202 would not reach.

Japan points the same way on outputs and attribution, although it takes a more permissive approach to training. Article 30-4 of the Japanese Copyright Act generally permits use of works for training, subject to a proviso excluding uses that would unreasonably prejudice the copyright owner’s interests, but output that reproduces creative expression falls outside that exception.[5] Japan's moral rights, particularly the rights of attribution and integrity, may provide broader protection than the DMCA because they are not limited to the removal of notices attached to an existing copy.[6] Japan's non-binding, comply-or-explain Principle Code on generative AI goes further on the input side: it applies to foreign providers serving Japan, encourages disclosure of training-data sources, calls for technical measures to reduce infringing outputs, and allows rightsholders preparing legal action to request information about whether their URLs were included in training data, information that may be relevant to claims concerning the training stage.[7] Japan's approach is noteworthy because it permits AI training under a broad “non-enjoyment” exception (art. 30-4) while leaving ordinary infringement remedies (injunctions and damages) available where AI outputs reproduce or adapt protected expression, making it a notable model for balancing innovation and authors' rights.

Practical Lessons for AI Developers from a Japanese Perspective

(1) Lawful training does not immunize infringing outputs;
(2) Technical safeguards should be used to reduce memorization and reproduction of copyrighted works;
(3) Transparency regarding training-data sources and risk-mitigation measures is increasingly expected under Japan’s Principle Code on generative AI.

---

[1] Doe v. GitHub, Inc., No. 24-7700 (9th Cir. Sept. 16, 2026)．

[2] Section 1202(b)(1) provides that no person shall “intentionally remove or alter any copyright management information”; § 1202(b)(3) prohibits distributing works knowing that CMI has been removed or altered.

[3] See Doe, slip op. at 16–17 (citing New York Times Co. v. Microsoft Corp., 777 F. Supp. 3d 283, 317 (S.D.N.Y. 2025)).

[4] LG München I, judgment of Nov. 11, 2025, Az. 42 O 14139/24, GRUR 2025, 1917 – GEMA/Open AI (not final).

[5] See Subdivision on Legal Systems, Copyright Subcommittee of the Council for Cultural Affairs, General Understanding on AI and Copyright in Japan (Mar. 15, 2024) (art. 30-4 does not apply where an "enjoyment" purpose coexists with analysis, e.g., training intended to output the creative expression of training data; where AI output directly conveys the creative expression of an existing work, its generation and use may infringe).

[6] Copyright Act of Japan (Act No. 48 of 1970), arts. 19 and 20. Saikō Saibansho [Sup. Ct.] July 21, 2020, Hei 30 (ju) no. 1412, 74 Minshū 1407 (Retweet case) (attribution right only; the attribution right under art. 19(1) is not limited to acts that themselves exploit the economic rights in arts. 21–27, and displaying a cropped image without the author's name infringes it even if users can click through to the original credit); see also id. (Hayashi, J., dissenting).

[7] Intellectual Property Strategy Headquarters, Cabinet Office, Principle Code on the Protection of Intellectual Property and Transparency for the Appropriate Use of Generative AI (2026), §§ 1(2)(d), 1(3); Principles 1–2.