---
type: "post"
id: "authors-sue-meta-alleging-muse-models-were-trained-1a05a0"
slug: "authors-sue-meta-alleging-muse-models-were-trained-1a05a0"
title: "Authors sue Meta, alleging Muse models were trained on scanned books via 'Project Vellum'"
description: "A group of 61 plaintiffs has sued Meta in federal court, alleging that its Llama and Muse models were trained on books obtained from shadow libraries and through a book-scanning effort called “Project Vellum.” For AI builders, the case is another reminder that model capability, access, and"
retrieval_nugget: "A group of 61 plaintiffs has sued Meta in federal court, alleging that its Llama and Muse models were trained on books obtained from shadow libraries and through a book-scanning effort called “Project Vellum.” For AI builders, the case is another reminder that model capability, access, and training-data provenance are separate operational risks."
published_at: "2026-10-28"
updated_at: "2026-10-28"
record_date: "2026-10-28"
date_kind: "published_at"
topics: ["models"]
entities: ["Meta Platforms","Llama","Muse","Muse Spark","Muse Glimmer","Project Vellum","Paul Levine","U.S. District Court for the Northern District of California"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Authors filed a separate lawsuit alleging Meta used scanned books and shadow-library downloads to train Llama and Muse models."
prior_refs: []
source_url: "https://copyrightnow.com/news/levine-v-meta-llama-muse-shadow-libraries-book-scanning-copyright-cmi-complaint-ndcal"
source_urls: ["https://copyrightnow.com/news/levine-v-meta-llama-muse-shadow-libraries-book-scanning-copyright-cmi-complaint-ndcal"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:authors-sue-meta-alleging-muse-models-were-trained-1a05a0"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/authors-sue-meta-alleging-muse-models-were-trained-1a05a0/","markdown":"https://newruntime.com/posts/authors-sue-meta-alleging-muse-models-were-trained-1a05a0.md","json":"https://newruntime.com/posts/authors-sue-meta-alleging-muse-models-were-trained-1a05a0.json"}
---

# Authors sue Meta, alleging Muse models were trained on scanned books via 'Project Vellum'

## Retrieval answer

A group of 61 plaintiffs has sued Meta in federal court, alleging that its Llama and Muse models were trained on books obtained from shadow libraries and through a book-scanning effort called “Project Vellum.” For AI builders, the case is another reminder that model capability, access, and training-data provenance are separate operational risks.

## Summary
A group of 61 plaintiffs has sued Meta in federal court, alleging that its Llama and Muse models were trained on books obtained from shadow libraries and through a book-scanning effort called “Project Vellum.” For AI builders, the case is another reminder that model capability, access, and training-data provenance are separate operational risks.

## What happened
Paul Levine, 58 other individual authors, and two companies filed a complaint against Meta Platforms, Inc. on October 19, 2026, in the US District Court for the Northern District of California. The case is Levine v. Meta Platforms, Inc., No. 3:26-cv-12137. The plaintiffs demand a jury trial.

The complaint contains four claims. It alleges direct copyright infringement based on copies obtained from shadow libraries; direct copyright infringement based on a book-scanning operation; contributory copyright infringement tied to alleged uploading through torrenting; and removal of copyright-management information. It seeks declarations, a permanent injunction, an accounting, damages, restitution or disgorgement, costs, attorneys’ fees, and interest. The complaint alleges willfulness for its two direct-infringement counts.

For Llama, the plaintiffs allege that Meta obtained books from Books3, Library Genesis, Z-Library, and Anna’s Archive. The complaint says Meta used those materials in training Llama 1 through Llama 4, drawing in part on interrogatory responses that Meta allegedly gave in another copyright case. It further alleges that Meta torrented data from shadow libraries, thereby re-uploading at least some material through leeching or seeding, and removed titles, author and copyright-owner information, and copyright notices from downloaded data. The pleading says each plaintiff’s works appeared in one or more of those libraries and asks the court to infer that Meta downloaded them.

The allegation concerning Muse is distinct. The complaint says that, beginning in or around mid-2025, Meta bought physical books in bulk, scanned them, ran the images through optical character recognition, and used the resulting text to train Muse models. It names Muse Spark and Muse Glimmer. According to the complaint, Meta internally called this operation “Project Vellum.” The plaintiffs contend that their works were commercially available in print while the operation was running and allege, on information and belief, that Meta acquired, scanned, and used copies of those works.

The plaintiffs’ theory is not merely that the books were used as training inputs. They argue that possessing a physical copy does not itself confer a right to reproduce the work, and present the alleged scanning as an independent basis of liability from the shadow-library claims. Meta is the sole defendant in this complaint.

Levine was filed the same day as Winspear v. Meta Platforms, Inc., No. 3:26-cv-12100, by the same counsel. The source says the cases share the same four counts, requested relief, jury demand, and factual allegations about Llama, torrenting, copyright-management information, and Project Vellum. Their principal difference is the plaintiffs and the works asserted: Levine names 59 individual authors and two companies, while Winspear names 60 individual authors.

## What the archive adds
This is a new legal fact on a known Muse story, rather than a restatement. Earlier coverage examined a different boundary: [Muse Wonder Full was restricted to paid Muse Code plans while its smaller sibling was available through an API](https://newruntime.com/posts/larger-sibling-muse-wonder-full-is-locked-to-e67a35/). That was an access and product-packaging question; this filing concerns the alleged provenance of training material for other Muse models.

## Trend signal
No supplied pattern genuinely fits this legal filing closely enough to support a trend claim.

## Why now
The immediate trigger is the October 19 filing. The source also identifies a parallel complaint filed on the same day, with matching factual allegations and legal counts but a different plaintiff group. That coordinated structure helps explain why Project Vellum appears here as a specific allegation rather than as a general dispute over AI training.

## Why it matters
Teams do not control a frontier model provider’s historical data-acquisition practices, but they do absorb some resulting uncertainty. A lawsuit does not establish that the alleged conduct occurred, nor does a complaint itself change a model’s API behavior. It can, however, affect procurement review, contractual conversations, public-sector or regulated deployments, and plans that depend on a particular provider or model family.

The practical distinction is important. Model availability answers whether a team can use a system; capability answers what it can do; provenance litigation asks whether the provider’s training pipeline may create legal exposure or disruption. Teams choosing models for long-lived products should avoid treating those three questions as interchangeable. They should maintain fallback model paths and record why a particular provider was selected.

## What happens next
- **The court docket will show Meta’s first substantive response or a procedural move.** (3 months, 75%). Watch for: an answer, motion to dismiss, or consolidation-related filing.
- **The allegations about Project Vellum will become a focal point of factual dispute if the case survives early motions.** (6 months, 60%). Watch for: filings that identify documents, witnesses, or a competing account of the scanning program.

## Limits and uncertainty
This account reports allegations in a complaint, not findings by a court. The source does not provide Meta’s response, evidence establishing that any plaintiff’s work was actually downloaded, scanned, or used in training, or a judicial ruling on whether the alleged conduct infringes copyright. The Levine exhibit described as listing registered copyrights is not publicly available, according to the source. Nor does the supplied material quantify the number of books allegedly scanned, identify the operational scope of Project Vellum, or establish which released Muse versions were trained on any particular material.

## Skill to build now
Add provenance risk to model-selection records this week. For each production dependency, capture the provider, model version, access route, fallback option, contract owner, and the reason it was chosen. This is not a legal conclusion about any provider. It is a way to make a model swap, a procurement review, or an internal risk discussion faster if litigation or policy changes alter the team’s assumptions.

## Reader takeaway
The filing does not prove Meta trained Llama or Muse on the materials alleged. It does make one operational point concrete: a model’s deployment terms and its training-data history can create different kinds of risk. Builders should design for both, rather than assuming API access settles the question.

## Open question
Will Meta contest the existence or use of Project Vellum, the plaintiffs’ inference that their works were included, or both? The answer will determine whether this remains a pleading-stage provenance dispute or produces evidence that changes how teams assess the Muse family.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)