Ninth Circuit Rules Generative AI Outputs Are Not “Stripped Copies” Under the DMCA

Introduction

On September 16, 2026, the Ninth Circuit held in Doe v. GitHub, Inc. that outputs produced by a generative AI coding tool are newly generated works, not copies of existing works, and therefore fall outside the Digital Millennium Copyright Act’s (DMCA) prohibition on removing or altering copyright management information (CMI).¹ The court limited that characterization to the § 1202(b) removal and alteration analysis, and declined to reach whether AI outputs that reproduce copyrighted material could constitute copyright infringement.² The ruling turned on the generative architecture of the accused AI systems, distinguishing generative AI tools that generate outputs by predicting sequences from statistical patterns from retrieval-based systems that retrieve and disseminate existing copyrighted material, with the court suggesting the latter could face different legal treatment. The Ninth Circuit also declined to reach, on procedural grounds, the question of whether training on copyright-protected materials from which CMI has been stripped violates § 1202(b).

Background

Section 1202(b) of the DMCA protects CMI from deliberate removal or alteration. CMI includes data identifying a work, its author or copyright owner, and the terms and conditions for use. The statute prohibits (1) intentionally removing or altering CMI, (2) distributing CMI knowing it was removed or altered without authority, and (3) distributing works or copies of works knowing that CMI was removed or altered without authority.³ Each prohibition also requires a showing that the defendant knew, or had reasonable grounds to know, that the challenged conduct would “induce, enable, facilitate, or conceal” copyright infringement.⁴ The DMCA does not require works to carry CMI in the first place; but where CMI is present, § 1202(b) shields it.

Although § 1202(b) has been applied for decades in traditional media contexts, such as cropping a photographer’s credit from a reprinted image, courts have only recently confronted how the statute applies to generative AI systems that learn from copyrighted material and produce new outputs. As plaintiffs have increasingly turned to § 1202(b) as a vehicle for challenging AI-generated content, Doe v. GitHub marks the first appellate-level decision to squarely address the question.

GitHub Copilot, which uses a modified version of OpenAI’s Codex model, was trained on billions of lines of publicly available code, including repositories carrying author names, copyright notices, and license terms as CMI.⁵ In this dispute, plaintiffs are anonymous programmers who alleged that Copilot violated § 1202(b) by generating code without such information. They advanced two theories of liability: (1) the “output theory,” that Copilot’s outputs are reproductions of copyrighted code from which CMI has been removed; and (2) the “input theory,” that ingesting copyrighted code into the training pipeline, while stripping CMI in the process, independently violates § 1202(b).⁶

The district court dismissed the output theory claims, holding that the complaint’s examples did not meet § 1202(b)’s requirement that copies be “identical.” Further, the district court understood plaintiffs not to have asserted the input theory, and did not address it.⁷ The Ninth Circuit took up the case on interlocutory appeal.

The Court’s Analysis

1. The Output Theory: Generative AI Creates “New Works,” Not Stripped Copies

The central holding of Doe v. GitHub turns on the court’s interpretation of the operative terms in § 1202(b). The statute prohibits intentionally “remov[ing] or alter[ing]” CMI and distributing works knowing that CMI has been removed or altered.⁸ The court read those terms as affirmative actions directed at CMI associated with an existing work.⁹

The court grounded its analysis in the statutory definition of CMI as “information conveyed in connection with copies of a work,” where “copy” means a “material object in which a work is fixed.”¹⁰ A system that generates a new work without including CMI has not altered CMI on an existing copy. As the court explained: “One who creates a new work and fails to include CMI cannot be said to have ‘removed’ or ‘altered’ anything.”¹¹ The panel further found that Copilot generates new code by predicting sequences from statistical patterns rather than retrieving and distributing a stored file, and that this generation-versus-retrieval distinction makes the omission of CMI a failure to add information to a new output, not removal from an existing work.¹²

Critically, the result is architecture-dependent: a search engine or system that reproduces stored passages may present a different § 1202(b) question.

2. Identicality: An Evidentiary Inference, Not a Freestanding Requirement

The degree of similarity between an AI system’s outputs and its training data is poised to become a central factual battleground in future § 1202(b) cases, and the court’s treatment of “identicality” helps explain why.

The Ninth Circuit rejected the district court’s conclusion that § 1202(b) imposes a freestanding “identicality” requirement (that is, that a plaintiff must show the AI output is a literal, character-for-character reproduction of a copyrighted work).¹³ The panel held that the concept of identicality is best understood as a “gloss” on the statutory terms remove, alter, and copies, rather than an independent (and atextual) element of a § 1202(b) claim.¹⁴ As the court explained, where two works are otherwise identical but the allegedly infringing version omits CMI, “a factfinder may reasonably infer that the defendant removed CMI from the copied work.”¹⁵ Conversely, material differences between the works could indicate that a defendant instead created a “new, derivative work to which the plaintiff’s CMI was never attached.”¹⁶

The practical takeaway cuts both ways. Substantial reproduction of a work without its CMI can still support an inference that CMI was removed, and minor cosmetic changes will not necessarily insulate a defendant. But where the output differs materially from any identifiable training work, those differences may undercut the inference of removal and may instead point toward generation of a new work rather than copying of an existing one.

3. Open Questions and Remaining Exposure

The decision resolves output-side DMCA liability as applied to the generative architecture at issue, but leaves several questions open and other avenues of potential liability intact.

Input-theory DMCA claims. The plaintiffs’ “input theory” (that ingesting copyrighted code and stripping CMI during the training pipeline independently violates § 1202(b)) was deemed forfeited for failure to preserve the argument in the district court.¹⁷ Because the Ninth Circuit did not reach its merits, the theory remains available to future plaintiffs. Training-stage CMI handling therefore remains an unresolved area of DMCA exposure.

Traditional copyright infringement. While the court acknowledged that Copilot’s outputs may in some cases be substantially similar to existing code, it reserved judgment on whether such similarity would support a claim for copyright infringement, stating it would “express no view on whether that similarity would allow plaintiffs to assert a claim for copyright infringement.”¹⁸ A plaintiff who can demonstrate that an AI system’s output is substantially similar to protectable expression in a copyrighted work retains a viable infringement theory. The evidentiary challenges are significant (tracing a specific output to a specific training input at scale), but the court’s refusal to address them confirms they remain live.

Open-source license breach. Contract and license theories remain independently viable where training, attribution, notice, or source-availability obligations attach to the underlying code. Many open-source licenses impose attribution or copyleft conditions that operate by contract regardless of DMCA or Copyright Act remedies. Plaintiffs’ breach of contract claims against GitHub remain pending, providing a vehicle through which license-compliance obligations may be tested.

Practical Implications

Doe v. GitHub carries both governance implications for industry participants and diligence implications for investors, boards, and sponsors.

For industry participants, such as AI developers and data-pipeline operators:

• The ruling is tied to Copilot’s generative architecture. Systems that retrieve source passages while generating surrounding text (such as retrieval-augmented generation (RAG) tools) may face a fundamentally different § 1202(b) analysis, because the retrieval step could be characterized as distributing a copy from which CMI was removed.
• AI developers and data-pipeline operators should audit training-data workflows to identify what CMI accompanies source materials, document whether and why collection or processing steps remove it, and retain records of how resulting copies are managed, as CMI handling will likely face scrutiny in future litigation.
• Enterprise deployers of third-party AI tools should consider their own exposure if an AI vendor’s outputs give rise to claims. Contracts with AI providers should address intellectual property representations, indemnification scope (including whether it covers CMI and license-compliance claims, not just copyright infringement), and allocation of defense costs.

• For rights holders, including authors, software developers, and content owners, the decision narrows one litigation pathway but clarifies others. Rights holders evaluating enforcement options should consider whether their claims are better framed under traditional copyright (substantial similarity), contract (license breach), or input-theory DMCA, and should ensure all viable theories are preserved early in any proceeding. The forfeiture of the input theory in this case illustrates the procedural stakes.
• The court’s focus on generation-versus-retrieval architecture reaffirms that systems with memorization safeguards or output filters designed to reduce verbatim reproduction may be better positioned to defend against both DMCA and copyright claims. Implementing and documenting such technical controls is emerging as a governance best practice.
• As AI systems increasingly incorporate third-party models, APIs, and datasets, the decision underscores the importance of understanding intellectual property indemnification throughout the AI supply chain. Companies developing or deploying generative AI should evaluate whether their vendor and licensing agreements address CMI handling, copyright exposure, and open-source compliance—and whether any indemnities are meaningful given counterparty creditworthiness and scope limitations.

Similar legal considerations also shape diligence priorities for investors, boards, and private equity sponsors:

• The decision resolves one vector of legal exposure (output-side DMCA liability) but leaves copyright infringement, training-stage DMCA claims, contract and license theories, and jurisdictional variation as distinct and open questions. Diligence should evaluate each separately, with particular attention to training-data sourcing and licensing history, the system’s generation-versus-retrieval architecture, controls for memorized or highly similar outputs, and the jurisdictions in which the product operates.
• Output-side DMCA exposure is now more defined, but the remaining open questions each carry their own risk profile. Investment and governance analyses should avoid treating the decision as a comprehensive clearance and instead map each unresolved vector against the portfolio company’s specific technology stack and data practices.
• Public companies developing or deploying generative AI should evaluate whether existing risk factor disclosures adequately address the legal uncertainties the decision leaves open. The unresolved status of input-theory DMCA claims, traditional copyright exposure, and open-source license compliance may warrant specific disclosure, particularly for companies whose business models depend heavily on AI-generated outputs or large-scale training data.

Conclusion

Doe v. GitHub is the first appellate decision to apply § 1202(b) to generative AI, and its emphasis on system architecture offers a framework likely to extend beyond this case. For now, the decision clarifies that omitting attribution from a generative output is not, without more, actionable under the DMCA’s CMI provisions. What remains is a familiar but unsettled landscape: copyright, contract, and the input theory each present distinct questions that future courts, and future litigants, will need to address on their own terms.

***

View Ropes & Gray’s AI Court Order Tracker for current standing orders/local rules regarding AI.