Summary
Authors say Meta used books shared through BitTorrent to train its Llama AI model. BitTorrent is a file-sharing network. In 2025, a judge ruled for Meta on the named authors’ training claim. Their evidence did not show harm to the market for their books. Other claims remained. The court still had to decide whether Meta shared copies. It also had to decide whether the authors could represent more writers.
Introduction
The case asks where the books used to train AI came from. It also raises questions about permission and file sharing. Authors say Meta used books shared online without their permission to train Llama. They also say Meta’s file-sharing software may have sent copies to other users. The court ruled on one claim. It did not resolve every claim or set a general rule for AI copyright disputes.
This article summarizes the allegations and court orders. Fair use is a U.S. legal rule about when copyrighted work may be used without permission. This case does not set a general rule for AI training.
What the authors alleged
The case is Kadrey et al. v. Meta Platforms. Plaintiffs include Sarah Silverman, Richard Kadrey, and Christopher Golden. They alleged Meta downloaded copyrighted books from LibGen and Anna’s Archive. They said it used BitTorrent. They said Meta used the books to train Llama. The authors also allege that Meta uploaded some files as it downloaded them. Those uploads could have shared copies with other users. These remain lawsuit claims; the court has not decided every allegation.
The case raised several ethical and business questions:
- Data provenance: Can a company identify where its training material came from?
- Permission: Did the company have a license or another legal basis to use the works?
- Distribution: Did a file-sharing method create copies for other users?
- Oversight: What did the team know about the data source and its risks?
The court’s June 2025 order recorded a dispute about BitTorrent. The parties did not dispute that Meta downloaded works from LibGen and Anna’s Archive. They did dispute whether Meta uploaded copies during downloads. See the court’s order on the training claim.
What the court decided about training
The U.S. District Court for the Northern District of California ruled on June 25, 2025. It ruled for Meta on the authors’ claim about training Llama on their books. They argued that this use infringed copyright. Summary judgment lets a court decide a claim without a trial. The judge must find the evidence cannot support a jury verdict. The judge found the authors had not shown harm to the market for their books. The ruling applied to the evidence and claims before the court. It did not decide that all AI training on copyrighted books is fair use.
The ruling applied to the named authors. They had hoped to sue for a larger group of writers. In a class action, people with similar claims ask the court to hear them together. The order left open the separate claim that Meta shared copies through BitTorrent.
What remained in the case in 2026
On March 25, 2026, the court let the authors file a Fourth Amended Complaint. It added a claim called contributory infringement. The authors say Meta helped others infringe copyright by uploading books through BitTorrent. They say those uploads let users share copies. In the same order, the court denied the plaintiffs’ request to gather evidence for a class action. Permission to amend the complaint was procedural. It did not decide whether the new claim is true. Read the March 2026 court order.
The training claim, distribution claim, and contributory-infringement claim are distinct. The 2025 decision addressed the named authors’ training claim on its record. It did not decide the later-added claim about uploads or establish a result for other authors.
Practical steps for AI teams
A company building or buying AI should keep a record of how it got its data and what each license allows. These steps can help teams spot questions early; they do not replace legal advice.
- Record each source. Note the source, owner, collection date, license, permitted uses, and any restrictions.
- Separate cleared data from unknown data. Do not treat “available online” as permission to copy or use a work for training.
- Keep evidence of permission. Save licenses, contracts, and any limits on storage, training, or redistribution.
- Review vendor terms. Check how the vendor stores your inputs and trains its models. Also check its rules for outputs and legal claims.
- Ask counsel about uncertain uses. Fair use depends on the facts and the applicable law.
- Plan a removal path. Know how you would stop using a source if its rights or license change.
Other copyright cases involving AI have different parties, evidence, and claims. Do not treat one case as deciding another. For broader context, see our complete guide to AI tools for entrepreneurs.
Sources
- Kadrey v. Meta, June 25, 2025 order on the training claim
- Kadrey v. Meta, March 25, 2026 order on amendment and class discovery
