IP Law Daily, COPYRIGHT—N.D. Cal.: Portion of copyright claim against Adobe over AI training data dismissed, (Aug 21, 2026)

By Wendy Biddle, J.D.

An author's allegations about Adobe's use of pirated books to train NVIDIA's Nemotron AI models were too conclusory to satisfy the standing requirements.

A federal district court in the Northern District of California granted Adobe Inc.'s motion to dismiss the portion of a copyright infringement complaint alleging the company used pirated books from the shadow library Anna's Archive to train NVIDIA's Nemotron large language models, finding that the plaintiff's allegations amounted to bare legal conclusions unsupported by facts sufficient to raise a plausible inference of infringement (Tanzer v. Adobe Inc., No. 3:26-cv-04712-JSC (N.D. Cal. Aug. 20, 2026)).

Background. Plaintiff E. Molly Tanzer, an author who holds registered copyrights in at least three books, brought a single-count copyright infringement claim against Adobe, alleging the company unlawfully used her copyrighted works in developing two distinct generative AI products. Adobe describes itself as a company providing services and tools for image, audio, video, and text editing and viewing, and has in recent years added AI capabilities to its products.

Central to the author's theory is the role of so-called shadow libraries, illicit online repositories of copyrighted content. The author identified six examples, including Library Genesis, Z-Library, Sci-Hub, Internet Archive, Bibliotok, and Anna's Archive. She described Anna's Archive as the most active and notorious of these repositories, one that in 2022 made available the content of all other major shadow libraries and now makes millions of pirated books available for public download.

The Nemotron series of large language models, the subject of Adobe's motion, originated with NVIDIA. NVIDIA released the Nemotron-4 15B model in February 2024. Although NVIDIA did not disclose the specific datasets used to train the model, it indicated generally that the model was trained with books. The training corpus totaled 8 trillion tokens, of which 240.6 billion were English-language books. Using average novel length estimates, Tanzer calculated the model was trained on approximately 1,881,250 English-language books. Tanzer's copyrighted books are among the works available on Anna's Archive.

Adobe's connection to Nemotron arose from a publicly announced partnership with NVIDIA. In March 2023, nearly a year before NVIDIA released Nemotron-4 15B, Adobe announced a collaboration with NVIDIA to develop next-generation generative AI models. Then in March 2024, a month after that release, Adobe announced it was working with NVIDIA to train new NVIDIA LLMs and was building datasets to train and tune next-generation NVIDIA AI foundation models, including Nemotron LLMs. Adobe also disclosed it was using its PDF Extract tool, an AI-powered utility for extracting content from PDFs into a format suitable for LLM training, in that process. Based on these announcements, Tanzer alleged Adobe admitted to training and fine-tuning Nemotron models and building the datasets used to train them.

Tanzer's theory connecting Adobe to Anna's Archive rested on a chain of inferences. Because the Nemotron model was trained on nearly 1.9 million English-language books, she alleged that acquiring such a volume could only have been possible by obtaining them from pirated sources such as Anna's Archive.

Procedural ruling on partial dismissal. Before reaching the merits, the court addressed whether it was procedurally proper for Adobe to seek dismissal of only the Nemotron portion of Tanzer's single-count claim. The court held that it was. Because the SlimLM and Nemotron products involved separate transactions or occurrences, the court reasoned they should have been pleaded as separate counts under Federal Rule of Civil Procedure 10(b). Relying on Ninth Circuit authority, the court reaffirmed that partial dismissal of a claim is a recognized procedural device in the Ninth Circuit, and that a plaintiff cannot defeat a motion to dismiss simply by bundling distinct theories into a single count.

Motion to dismiss. Adobe argued that the author did not have standing to pursue the claim because she failed to sufficiently allege a copyright infringement claim or injury in fact. Article III standing requires a plaintiff to allege an injury in fact that is actual or imminent, fairly traceable to the defendant, and redressable by a favorable decision. A direct copyright infringement claim separately requires the plaintiff to allege ownership of the copyright and that the defendant itself violated her rights. The court noted that the Article III standing inquiry substantially overlapped with the merits analysis and addressed both together.

The court distinguished between allegations that are merely possible and those that are plausible, as required under Twombly and Iqbal. The complaint's assertion that Adobe used Anna's Archive to “acquire” the author’s copyrighted work was, in the court's assessment, a bare conclusion. At oral argument, when asked to identify which complaint allegations supported the inference that Anna's Archive was used to train Nemotron, plaintiff's counsel pointed to two paragraphs. The first alleged that NVIDIA acquired millions of unauthorized copies of copyrighted books from Anna's Archive and that, upon information and belief, Adobe used those books to train and fine-tune Nemotron models. The second alleged that major AI companies including OpenAI, Meta, Anthropic, and NVIDIA pirated books from Anna's Archive and other shadow libraries. The court found both paragraphs themselves conclusory as to NVIDIA's use of Anna's Archive, rendering the downstream inference about Adobe's use doubly unsupported.

The court was equally unpersuaded by the volume-based argument embedded in the complaint; that the sheer scale of the Nemotron training corpus logically entailed resorting to shadow libraries. That numerical inference, however intuitive, did not constitute a factual allegation capable of filling the pleading gap. As the court observed, "repeating a conclusory allegation several times in a complaint does not make it any less conclusory." Without a plausible factual basis for concluding that the plaintiff's works were among those used to train Nemotron, she could not establish either an injury in fact traceable to Adobe or the direct copying element of her infringement claim. The court therefore did not reach Adobe's separate argument that even if the dataset allegations were sufficient, the complaint failed to support an inference that Adobe, as opposed to NVIDIA, was the party that actually acquired the works.

The court granted Adobe’s motion to dismiss with leave to amend. The SlimLM portion of the plaintiff’s claim, which rests on a more particularized chain of dataset attribution, remains pending.

The Case is No. 3:26-cv-04712-JSC.

Judge: Corley, J.

Attorneys: Christopher K.L. Young (Saveri Law Firm, LLP) for E. Molly Tanzer. Douglas Andrew Winthrop (Arnold & Porter Kaye Scholer LLP) for Adobe Inc.

Companies: Adobe Inc.

Cases: Copyright CaliforniaNews