On January 28, Universal Music Publishing Group (UMGP), Concord Music Group, and ABKCO Music jointly filed a lawsuit against Anthropic, seeking $3.1 billion in damages.
The core allegation: Anthropic built Claude on a foundation of piracy.
Two months later, on March 18, BMG Rights Management filed its own suit, this time naming specific artists — lyrics by Bruno Mars and the Rolling Stones were used to train the model without authorization.
Together, the two cases may represent the largest legal action from content copyright holders against the AI industry to date.
First, look at the case that already reached a conclusion last year
In June 2025, the Bartz v. Anthropic case was settled out of court for $1.5 billion.
The reasoning in that ruling is worth examining closely:
- Training AI on copyrighted books = fair use. The court found the training to be sufficiently transformative, describing it in the ruling as "spectacularly so."
- Storing pirated files themselves ≠ fair use. Anthropic downloaded and stored large volumes of pirated material during training — this was deemed a violation.
The conclusion: training is permissible, but how you obtain the training data is something the court will scrutinize.
That line is drawn clearly, and it is a subtle one.
Why the music industry's logic is more complex
With books, the situation is: you download an entire book, use it for training, and that book is a complete, independent copyrighted work.
Lyrics are more complicated. They are typically short, widely reproduced on many platforms, and it is harder to argue they were deliberately collected as training data for Claude. This argument is much harder to make than with books.
But the $3.1 billion claim suggests the music industry's lawyers are confident in their case — or at least in their leverage for a settlement.
The $1.5 billion precedent has already shown Anthropic's tendency to settle rather than fight to the end.
Two other tracks are also moving forward
On the Meta front, in Kadrey et al. v. Meta, Meta won a partial dismissal at the initial stage — the court found that using pirated material for training could, under certain conditions, be considered fair use. However, claims related to distribution during the torrenting process are still being litigated.
This ruling is broadly favorable for the industry — it suggests that the training act itself already has a relatively clear defense path.
On the OpenAI front, the Southern District of New York consolidated 12 copyright lawsuits and ordered OpenAI to hand over millions of ChatGPT conversation records to the plaintiffs for discovery. This move indicates the plaintiffs are trying to prove that ChatGPT's output itself infringes, not just the training data.
Disney v. Midjourney is also still ongoing. This case focuses more on whether AI-generated images constitute derivative infringement of IP like Mickey Mouse, a different story from training data.
The focus of litigation is shifting from training to output
In a report earlier this year, legal consultancy Morrison Foerster concluded that the center of gravity in 2026 AI copyright litigation is shifting from whether training data infringes to whether model output infringes.
In other words, future lawsuits may be more about: your AI's output infringes, rather than what data you used to train the model.
If this shift materializes, the impact on AI companies will be deeper — because you can control what data you train on, but it is very difficult to control what the model outputs across millions of user interactions.
The real-world impact of these lawsuits on the industry
In the short term, AI companies are already changing how they acquire data. Google, OpenAI, and Anthropic have been signing large numbers of copyright licensing agreements in recent years — with news outlets, academic publishers, and book companies. This is not because they believe they would definitely lose without them, but because the uncertainty is too high and buying peace of mind upfront is more cost-effective.
In the medium term, the rulings from these lawsuits will gradually clarify the legal boundaries for AI training. The clearer the boundaries, the better for the entire industry — at least that is what investors and compliance teams will think.
The $3.1 billion figure is alarming, but what is more worth watching is: where will the courts ultimately draw the line between content that can be used to train AI and content that cannot. Once that line is drawn, that is when this war truly ends.
Sources: AI in litigation series: An update on AI copyright cases in 2026 (Norton Rose Fulbright); CocoLoop, Generative AI Lawsuits Timeline (Sustainable Tech Partner); AI Trends for 2026 - Copyright Litigation Shifts from Training Data to AI Outputs (Morrison Foerster); US copyright cases against AI training surge past 100 (Noah News)