A lawsuit over AI training data has taken an unexpected turn. Authors including Richard Kadrey and Sarah Silverman sued Meta, claiming its Llama language models were trained on pirated books obtained through BitTorrent — a peer-to-peer file-sharing protocol where users automatically upload to others as they download. A judge has already ruled that the AI training itself was fair use, but one claim survived: that Meta distributed pirated books to other torrent users in the process. That claim is still active. And now, to test Meta’s defense, publishers have subpoenaed Meta’s direct competitors.
What happened
Meta’s position is that any uploading that occurred was simply how BitTorrent works. The company framed it as a technical inevitability, not a deliberate act of distribution. Three additional publishers filed related lawsuits over the same alleged torrenting activity, and all the cases ended up with the same judge.
Rather than wait for Meta to produce its own internal records, the publishers went a different direction. They subpoenaed OpenAI and Anthropic — Meta’s competitors in the AI industry — asking for details about how those companies configured their torrent clients when acquiring similar training data.
The publishers offered a simple way out: if either company would confirm whether they took steps to prevent uploading during their own downloads, the document demands would be dropped. Both OpenAI and Anthropic declined and pushed back instead.
Evidence has also surfaced that a Meta engineer wrote a script to prevent one type of torrent activity, though apparently not the uploading side. That detail matters, because it suggests the behavior may not have been entirely unavoidable.
Who is affected
The most directly affected people are authors, journalists, and other creators whose published work appeared in shadow libraries — websites that host large collections of copyrighted books and articles without permission from the rightsholders.
Academic publishers and independent imprints are also involved. Several have filed coordinated AI training copyright lawsuits targeting the same alleged activity, all now consolidated before one judge.
But the reach goes further. Any creator who has ever had a book, article, or written work indexed by a shadow library without their knowledge has a stake in how these arguments are resolved. Freelancers who license written content, and founders who build products around original text, should pay attention too. The outcome will shape what AI companies can legally do with scraped material — not just what they do with it after training, but how they obtained it in the first place.
OpenAI and Anthropic are now pulled into a case where they are not defendants. They face pressure to disclose details about their own data acquisition practices, which may explain why both pushed back rather than offering even a limited confirmation.
What the real risk is
Meta’s core argument is that using a protocol that technically requires redistribution is enough to shield a company from liability for that redistribution. If that argument succeeds, it could set a precedent with wide implications for AI training copyright lawsuits going forward.
The engineer’s script complicates that argument. If Meta could suppress one type of torrent behavior, the question becomes why the upload function was apparently left active. That is not an accusation — it is the factual question the court is working through.
The competitor subpoenas go directly at this point. If OpenAI or Anthropic configured their torrent clients to suppress uploading, that would directly contradict the claim that seeding is an inherent and unavoidable feature of the protocol. It could undermine a key part of Meta’s defense.
For creators, the broader risk is this: a successful necessity defense could make it much harder to hold AI companies responsible for how they obtained training data, separate from what they did with it afterward. The acquisition process itself could become legally invisible.
The fact that two major AI companies refused even a narrow disclosure also raises questions. Companies generally do not push back against limited requests unless there is something they prefer not to confirm.
What to do today
These steps will not resolve a lawsuit on their own, but they put you in a stronger position if you ever need to act.
- Search your titles in shadow libraries. Look up your published work by name on the major shadow library sites. If you find your titles listed without your permission, take screenshots immediately and note the date. This documentation could matter later.
- Register your copyrights formally. In many countries, you own copyright the moment you create something, but formal registration strengthens your legal standing significantly. In the United States, for example, registration is generally required before you can file a lawsuit for infringement. Check the rules in your country and register anything you have not already registered. A lawyer can confirm what applies to you.
- Read your licensing agreements carefully. Look for language that permits sublicensing, data use, or transfer of rights to third parties. Even a single clause can affect your options. If you find language you do not understand, ask a rights professional before signing anything new.
- Follow the Kadrey and Silverman case and the related publisher lawsuits. The rulings on fair use and distribution liability will set the terms for future claims. You do not need to follow every filing — just watch for major decisions, which will be covered in the press.
- Treat AI licensing requests as a signal. If a company contacts you asking to license your work for AI training, that is confirmation your work has value in this context. Do not agree to anything immediately. Consult a rights professional first and find out what you are actually being asked to give up.
Why this keeps happening
AI companies built their models during a period when the legal rules around training data were genuinely unsettled. Many made data acquisition decisions before courts had weighed in. Shadow libraries made that easier by aggregating enormous amounts of copyrighted material in one place — a ready-made shortcut for assembling large datasets quickly.
BitTorrent was designed for decentralized sharing, which means the architecture itself creates legal ambiguity about who is responsible for redistribution. Copyright law was written long before generative AI existed, so there is no clean rule that covers how training data is acquired.
But there is a deeper reason this keeps happening. Online systems have no reliable way to tie ownership to a real, accountable person. When a book ends up in a shadow library, it is just a file. There is no persistent, verifiable link back to the author who wrote it, the publisher who holds the rights, or the terms under which it was created. Because ownership is invisible in the system, platforms and data aggregators can treat the material as effectively unclaimed — and act accordingly. The creator finds out what happened only after a lawsuit is already underway, often years later. Until digital systems can reliably establish who owns what and who is responsible for what, this pattern will repeat.
Creators rarely have the resources to monitor where their work ends up or pursue legal action individually. Class actions have become the primary tool for pushing back, which is why the cases keep consolidating and why individual rightsholders often have no voice in how they are resolved.
Frequently asked questions
Does the fair use ruling mean Meta can train AI on my books without consequences?
Not necessarily. The ruling covered the training itself, but the claim about distributing pirated books to other torrent users is still active in court. Fair use is also a case-by-case determination — it varies by jurisdiction and by the specific facts involved. A lawyer in your country can tell you what it means for your situation.
Why are OpenAI and Anthropic being dragged into a lawsuit against Meta?
The publishers subpoenaed them as third parties, not defendants. The goal is to find out whether those companies configured their torrent clients to prevent uploading when they acquired similar data. If they did, that would challenge Meta’s argument that uploading is technically unavoidable when using BitTorrent.
What can I do if my published work appeared in a shadow library used for AI training?
Start by documenting it — screenshots, dates, URLs. Then check whether your copyright is formally registered, because registration generally affects your legal options. Review any contracts you have signed for language about data use or sublicensing. Then consult a rights professional who can advise you based on your specific situation and jurisdiction. Do not assume you have no options, but also do not assume the path is simple.
Related reading
- 74 Sports Pirates Charged: What Creators Can Learn
- Unpatched Plex Servers: Protect Your Media and Your Work
Originally reported by torrentfreak.com. This article summarises that reporting and adds practical guidance.
Everyone deserves authenticity and accountability online. That is the mission we are working on. Subscribe to join us, and get a free copy of Wes Kussmaul’s Escape the Plantation.
This article is general information, not legal advice. Laws differ by country — consult a qualified lawyer about your situation.
