AI Learning or AI Copying? Ninth Circuit Weighs In

The plaintiffs in Doe v. GitHub, Inc. are software programmers who published copyrighted code on GitHub, an online platform where software developers store, manage, and share code. They made their code publicly available under open-source licenses that allowed others to use or modify it, subject to certain conditions. One condition required copies or modified versions to retain the author’s name, a copyright notice, and the license terms that govern use. Federal copyright law calls that identifying and licensing information “copyright management information” (CMI).

GitHub Copilot is an AI coding tool trained on billions of lines of publicly available code, including code from public GitHub repositories. When a programmer gives Copilot a prompt, Copilot generates code that it predicts will answer the request. The plaintiffs alleged that Copilot sometimes generated “essentially verbatim” portions of their code without the author names, copyright notices, and license terms that accompanied the originals.

The Digital Millennium Copyright Act (DMCA) does not require authors to include CMI in copyrighted works, but Section 1202(b) protects CMI from deliberate removal or alteration when it is present. The plaintiffs argued that Copilot violated that provision by generating similar code without the CMI attached to their originals. The Ninth Circuit rejected the claim because it interpreted the complaint to describe Copilot as generating new code rather than removing CMI from an existing copy.

Why the Ninth Circuit Concluded Copilot Did Not Remove CMI

The Ninth Circuit began its analysis by asking what Section 1202(b) means when it prohibits a person from “removing” or “altering” CMI. The court read those verbs to require an affirmative act directed at information attached to an existing copy of a work. Under that reading, a plaintiff must allege that the defendant took CMI off an existing copy or changed CMI that was already there; allegations that a newly created work simply lacks the original CMI are insufficient.

The court illustrated its reading with an example involving a photograph. In the Ninth Circuit’s example, reprinting a photograph while cropping out the printed credit identifying the photographer would remove CMI because the credit was attached to the copied image and then taken away. The court contrasted that scenario with creation of a new work that never carried the credit; under the Ninth Circuit’s reasoning, the absence of the credit from the new work would not itself establish removal.

The Ninth Circuit tied that reading to the statutory definition of CMI and explained:

CMI is “information conveyed in connection with copies . . . of a work,” not in connection with excerpts or derivative works. A “copy” of a work is a “material object[] … in which a work is fixed,” so CMI is information that exists in connection with those material objects. Violating section 1202(b) requires removing or altering the CMI on those material objects, such as by reprinting an image while cropping out the “printed gutter credit near the [i]mage” identifying the photographer. . . . To state a claim under section 1202, plaintiffs must therefore allege that defendants removed or altered CMI from copies of existing protected works; merely alleging that a similar or derivative work does not include the CMI, without facts showing its removal or alteration, is insufficient.

The Ninth Circuit also rejected a bright-line rule requiring plaintiffs to prove that the later work is perfectly identical to the original. In the court’s view, a near-complete reproduction that omits CMI can be strong circumstantial evidence that the defendant copied an existing work and removed the information attached to it. But plaintiffs are not required to prove that the defendant literally made an identical copy in order to prevail on a CMI-removal claim. As the Ninth Circuit explained, “[t]wo works need not be literally identical to support an inference that CMI has been removed.”

The Ninth Circuit then applied that framework to the programmers’ own description of Copilot. The complaint described a model that infers statistical patterns from large amounts of code and uses those patterns to predict the most likely response to a new prompt. The court understood that process as generation of a new work, even when the resulting output resembles or matches snippets from the training data.

The court explained why it viewed Copilot’s probabilistic process as generation rather than retrieval:

According to the complaint, Copilot relies on a “complex probabilistic process” to predict “the most likely solution to a given prompt,” based on “the solution it has found in the most projects” answering similar questions. Sometimes, plaintiffs allege, that output may match snippets of “code from the training data.” But their account of the algorithm’s internal process—inferring “statistical patterns governing the structure of code” and identifying the most likely completion—does not describe an action taken with respect to CMI attached to an existing work. Instead, it describes a process through which Copilot generates new works. In that respect, it differs from a traditional search engine, which, in response to a user’s query, retrieves and displays stored information—that is, copies of materials that already exist.

The Ninth Circuit’s reasoning therefore turns on how it characterized Copilot’s process. Because the complaint described Copilot as generating a new output from learned statistical patterns, the court concluded that the source work’s CMI was never attached to that new output and could not have been removed from it. If a system instead retrieves or reproduces an existing copy and omits CMI that was attached to that copy, the court’s analysis leaves room for a different result.

The Ninth Circuit’s focus on generation versus retrieval makes the system’s technical operation part of the legal analysis. A copyright owner may have a stronger CMI claim if the facts show that a system retrieved or substantially reproduced an existing work and omitted information that was attached to that work. An AI developer may point to evidence that its model generated the challenged output through a probabilistic process instead of retrieving an existing copy.

The Larger Question: When Is AI Learning, and When Is It Copying?

Doe joins a growing line of AI-copyright decisions asking a common question across different technologies and factual settings: When is a system learning from existing works and generating something new, and when is it reproducing what came before? The answer depends on the actions the specific technology actually performs. Different AI systems can perform different functions, and even generative systems can produce output in different ways. Thomson Reuters v. Ross, for example, involved a non-generative legal-research tool, while the later cases below address generative systems producing new text or code.

These decisions address different technologies and legal questions, so they do not create a single test for when AI is “creating” or “copying.” Courts can also disagree about how to characterize similar uses of generative AI, as Bartz and Kadrey illustrate. But the cases show courts repeatedly asking how much the output owes to the source material and what the system actually did to produce it.

A Similar AI Output Could Still Infringe Copyright

The Ninth Circuit expressly limited its holding to the CMI-removal claim and left ordinary copyright infringement unresolved. The court acknowledged that Copilot may sometimes generate code that is substantially similar to existing code and stated that it was expressing “no view” on whether that similarity could support an infringement claim.

The opinion therefore treats CMI removal and copyright infringement as separate questions. A copyright-infringement claim can focus on whether an output copied protected expression; Section 1202(b) asks whether the defendant removed or altered CMI from an existing copy. The Ninth Circuit declined to treat similarity plus missing CMI, without facts showing removal, as sufficient to state the separate DMCA claim.

What Doe Means for Copyright Owners

Doe gives copyright owners a practical blueprint for pleading a CMI-removal claim: identify the existing copy, identify the CMI it contained, and explain how the defendant actually removed or altered that information. A close reproduction without CMI can help support that inference, but resemblance alone is not enough.

Copyright owners should also separate a CMI-removal theory from an ordinary infringement theory. If AI-generated output actually copies protected expression but the facts do not show that CMI was taken off an existing copy, Doe leaves open a traditional copyright claim. The two theories ask different questions and should be supported by different facts.

What Doe Means for Businesses Using AI Tools

Doe tells businesses developing or deploying AI tools to ask a concrete question at the outset: Does the system generate new material from learned patterns, or can it retrieve and reproduce existing material? That distinction drove the Ninth Circuit’s CMI analysis and can change the legal risk when an output closely resembles a source work and thus may provide solid evidence of illegal copying.

That question should also shape due diligence review when businesses select or configure AI tools. Companies should understand whether the system can return memorized or stored source material and whether it uses safeguards designed to limit near-verbatim reproduction. Those technical facts can affect whether a future dispute resembles the retrieval-and-removal scenario the Ninth Circuit said could support a stronger CMI claim, or the generative process the court found insufficient here.

Copyright © 2026 by Ballard Spahr LLP.
www.ballardspahr.com
(No claim to original U.S. government material.)

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, including electronic, mechanical, photocopying, recording, or otherwise, without prior written permission of the author and publisher.

This alert is a periodic publication of Ballard Spahr LLP and is intended to notify recipients of new developments in the law. It should not be construed as legal advice or legal opinion on any specific facts or circumstances. The contents are intended for general informational purposes only, and you are urged to consult your own attorney concerning your situation and specific legal questions you have.