It's Not the Training, It's the Torrenting: How Strike 3 v. Meta Reframes AI Copyright Liability
For two years, AI developers have won copyright fights on fair use. A new Meta ruling points the litigation somewhere they can't defend as easily: not how the data was used, but how it was pirated. Why Strike 3 v. Meta is a roadmap for every AI copyright case already on file.
On June 11, 2026, a federal judge in the Northern District of California refused to dismiss a copyright suit accusing Meta of using BitTorrent to download films and train its AI models.
On its face, Strike 3 Holdings v. Meta Platforms looks like a lurid sideshow: the plaintiffs are adult-film producers, and the opinion quotes torrented file names at length.
But strip away the subject matter and the ruling does something the AI industry has spent two years trying to prevent. It lets a copyright case against an AI developer proceed without the plaintiffs having to prove their works were used in training at all.
That single move is why this order matters far beyond Meta. For two years, AI defendants have fought copyright claims on the ground of fair use, and on the question that has dominated the litigation, whether training a model on copyrighted works is transformative, they have mostly been winning.
Judge Eumi Lee's order points the fight at a different target: not how the data was used, but how it was obtained.
What the court actually held
The complaint alleges that, between 2018 and 2025, IP addresses tied to Meta torrented at least 2,396 of the plaintiffs' films more than 6,000 times, and seeded many of them back out to other BitTorrent users for days at a time.
The plaintiffs link the activity to Meta through patterns rather than a single smoking gun: addresses downloading files that share a key term on the same day, synchronized switches between languages, and television episodes pulled out of sequence, all of which the court found pointed to an algorithm rather than employees watching movies at their desks.
The pivotal holding is on direct infringement. Meta argued that the plaintiffs had to allege their specific films were used to train a specific model.
Judge Lee rejected that, distinguishing the cases Meta relied on, In re Google Generative AI and In re Mosaic LLM, where the plaintiffs had pleaded infringement that occurred during the training process and therefore had to show their works were in the training set.
Here, by contrast, the infringing act is the torrenting itself.
Downloading and uploading copies are acts of reproduction and distribution under the Copyright Act, so the plaintiffs stated a claim, in the court's words, regardless of whether their films were used to train specific AI models. The case can proceed to discovery on acquisition alone.
The acquisition problem fair use has not solved
To see why that is so consequential, look at where the fair-use fight actually stands.
In Bartz v. Anthropic, Judge William Alsup held in June 2025 that training a model on lawfully acquired books was exceedingly transformative fair use, but that downloading and storing pirated copies from shadow libraries was not.
Weeks later, in Kadrey v. Meta, Judge Vince Chhabria granted Meta summary judgment on the training question, though he stressed he was doing so because the authors had failed to develop a market-harm argument, not because AI training is categorically lawful. In both cases, the courts blessed the use while flagging the piracy.
That distinction is now the most valuable real estate in AI copyright law. Training may be fair use; acquiring the training data by piracy is a separate act that fair use has not excused.
The price of getting that wrong is no longer theoretical. Anthropic settled the pirated-library claims in Bartz for $1.5 billion, the largest copyright recovery in history, roughly $3,000 per work, and agreed to destroy the pirated data. Strike 3 v. Meta takes that acquisition theory and proves it can stand entirely on its own at the pleading stage, untethered from any claim about model outputs or training sets.
Napster logic and a brand-new Supreme Court rule
The order is not a one-off on its secondary-liability claims either, which is part of what makes it portable to other cases.
On vicarious liability, the court found a plausible direct financial interest by analogy to A&M Records v. Napster: just as Napster's value grew with its ever-expanding library, an AI model's value grows with the quantity and quality of its training data, so the infringing acquisition foreseeably draws paying users to the resulting product.
The court pointed to the complaint's allegation that Meta projects $460 billion to $1.4 trillion in AI revenue by 2035 as the financial benefit.
On contributory liability, the court applied a Supreme Court decision handed down while the motion was pending.
In Cox Communications v. Sony Music, decided in March 2026, the Court narrowed secondary liability, holding that merely providing a service used by some to infringe is not enough; the provider must have intended infringement, shown either by inducement or by tailoring a service incapable of substantial lawful use.
Judge Lee acknowledged that Meta's mere provision of servers and IP addresses would fail that test, then held that the alleged active steps, building an algorithm and configuring private cloud servers to torrent, were tailored to infringement and cleared the new bar.
A ruling that survives the most defendant-friendly secondary-liability standard in a generation is one other plaintiffs will study closely.
What it means for the cases already on file
The AI copyright docket is large and growing: a consolidated multidistrict proceeding in New York folds in roughly sixteen suits against OpenAI, led by The New York Times; In re Google Generative AI and In re Mosaic LLM proceed in California; and music publishers, visual artists, and news organizations have their own actions.
Many of these defendants are alleged to have built their training corpora from the same shadow libraries at issue in Bartz, including Library Genesis, Anna's Archive, and the Pirate Library Mirror.
For every one of those cases, the acquisition theory offers plaintiffs a route that does not run through the fair-use thicket.
The BitTorrent wrinkle sharpens the point. A defendant who pirates a file at least makes a reproduction; a defendant who pirates it over BitTorrent also distributes it, because the protocol uploads pieces of the file to other users as it downloads.
Distribution of complete copies to the public is among the hardest conduct to defend as fair use, and it exposes the defendant to the separate distribution right.
Meta's own admissions in the parallel Kadrey case, that it configured private cloud servers to torrent training archives, supplied much of the circumstantial foundation the court relied on here, a reminder that discovery in one AI case now routinely arms plaintiffs in the next.
Two coasts, two questions
The timing was almost too neat. The same day Judge Lee issued her order, the Third Circuit heard argument in Thomson Reuters v. Ross Intelligence, the first federal appeal to consider whether AI training can be fair use.
The split screen captures where the law is heading. On one coast, appellate judges probe whether the use is transformative; on the other, a trial judge signals that the answer may not matter if the data was stolen to begin with.
Plaintiffs who lose the fair-use argument on training can still win on acquisition, and acquisition is the one fact pattern that discovery, not legal theory, tends to settle.
A note of caution belongs here. Strike 3 Holdings is, by most counts, the most prolific copyright filer in the United States, with more than twenty thousand BitTorrent suits to its name, and its tactics have drawn sharp judicial criticism.
In a 2018 ruling, Judge Royce Lamberth wrote that the company treated his court "as an ATM" and sought to have it oversee a "high-tech shakedown", though the D.C. Circuit later reversed that particular decision.
None of that changes the legal significance of the ruling here.
A favorable pleading standard does not depend on the sympathy of the plaintiff who wins it, and the doctrine Judge Lee articulated is now available to the news publishers, authors, and record labels litigating against the largest AI developers.
Why the order matters for AI copyright cases
This is a motion-to-dismiss ruling, not a finding of liability. Meta has admitted nothing, the allegations remain unproven, and the court was required to read the complaint in the plaintiffs' favor.
But the practical effect is real.
By holding that the acquisition of training data can be infringement in its own right, the order lowers the bar for a wave of cases to reach discovery, where the decisive question becomes simple and concrete: how did you get the data?
For AI developers, that question is harder to answer well than the fair-use debate they have been winning, and it is the one that produced a $1.5 billion settlement the last time a court forced it.
For plaintiff firms, Strike 3 v. Meta is a template for asking it.
Know which legal theory shifts the exposure next.
AI copyright moved from a fair-use debate to a data-acquisition problem in a single order, and the exposure followed. The only question is which firms saw it first.
By the time a ruling like this is news, the best matters are already spoken for.
Rain Intelligence puts your firm ahead of the filing, with daily, personalized intelligence matched to each attorney's expertise, clients, and jurisdictions. Book a demo.