A legal battle that has dragged on for years is now being resolved with real money.
Anthropic and nearly 500,000 authors have reached a $1.5 billion settlement in the copyright class-action lawsuit Bartz et al. v. Anthropic. A final approval hearing is scheduled for May 14, and if the judge approves, the money will officially be paid out.
This is the largest copyright settlement in the AI industry to date. However, the significance of this case goes beyond "how much Anthropic paid." What is more noteworthy is the legal boundary the court has drawn in this process—a line that will affect how the entire industry handles training data going forward.
Training on Copyrighted Books Is Permissible, But Pirated Sources Are Not
The core reasoning of this case is worth reading carefully.
The court determined that using copyrighted works to train large language models constitutes fair use. The original ruling stated that "the use of training data is transformative—even highly transformative." In other words, using copyrighted books to train Claude is, from a legal standpoint, not a problem in itself.
However, Anthropic's problem lay elsewhere: the training corpus it used contained pirated works. Downloading millions of pirated books—even if the ultimate purpose was to train an AI—is an independent act of infringement not protected by fair use.
Thus, this $1.5 billion resolves the issue of "pirated sources," not the "act of AI training itself."
Training is permissible, but it must use legally sourced data. This is the clearest AI legal red line of 2026.
As part of the settlement, Anthropic is also required to destroy that batch of pirated datasets. It has to pay the money and delete the data.
Each author will receive approximately $3,000 before legal fees are deducted. With 500,000 people sharing $1.5 billion, that is the average. It is not a huge amount, but class-action lawsuits are not meant to make everyone rich—their purpose is to set rules.
The Music Industry Is Also Settling Accounts, With More Flexibility
It is not just the publishing world. Two major cases in the music industry have also concluded recently:
- Universal Music Group vs. Udio: The parties reached a settlement and simultaneously signed a formal licensing agreement. Money was paid, and future use will be based on licensed data.
- Warner Music Group vs. Suno: As part of the settlement, Suno committed to developing a completely new product using a "more advanced licensing model." This means it is not just paying compensation but also replacing its existing model with one trained on properly licensed data.
The music industry's approach is somewhat smarter than the publishing world's: it is not just about seeking compensation but turning the settlement into a starting point for establishing long-term licensing relationships. Copyright holders get paid and secure future revenue from data licensing.
AI-Generated Output Cannot Be Copyrighted
Meanwhile, on March 2, the Supreme Court declined to review the Thaler v. Perlmutter case.
This case sought to obtain copyright protection for AI-generated content. By refusing to hear it, the Supreme Court effectively upheld the lower court's ruling: AI-generated works, without human creative contribution, cannot be copyrighted.
This has direct implications for the business logic of AI tools:
- An article written by a user using Claude: the user holds the copyright.
- Content "created" by Claude without any human prompt or guidance: no copyright exists, and anyone can use it.
This rule is essentially settled and will not change.
Thomson Reuters Won, Legal Data Is a Different Line
There is one exception in the AI copyright cases that followed a different logic: Thomson Reuters sued the AI legal research tool Ross Intelligence, alleging it copied Westlaw's legal headnotes for training.
The court ruled: this does not constitute fair use.
Why is it different from the Bartz case? Two key distinctions:
- The legal headnotes are the product of Thomson Reuters' editors' work, not raw source material.
- Ross and Westlaw are direct competitors, and the court found that this use was not transformative.
This ruling is currently under appeal. It shows that the boundaries of "fair use" are not set in stone; competitive relationships and the nature of the data both influence the judgment.
Several Major Cases Still Ongoing
Several key lawsuits have yet to be resolved:
- In Re OpenAI Copyright Infringement Class Action: 12 cases have been consolidated into an MDL. The core issue is whether ChatGPT's output constitutes infringement.
- Disney et al. vs. Midjourney: The most significant pending case in the image generation field, directly impacting the future of text-to-image tools.
- Kadrey et al. vs. Meta: Whether the distribution and use of open-source models constitute infringement remains undecided.
Precedents for image generation and AI output infringement have not yet been established, representing the biggest current legal uncertainties.
The Rules Are Largely Set
For AI companies, several clear signals can be read from this batch of rulings and settlements:
- Training itself is largely fair use, but pirated sources are an independent risk. Do not cut corners on this cost.
- The music industry's "licensing + settlement" model may become the new normal. Copyright holders are increasingly inclined to seek licensing deals rather than just compensation.
- Image generation and output infringement remain minefields. It is best not to bet on "waiting for a precedent to emerge."
- AI-generated output has no copyright—this rule is settled. Business models should avoid relying on copyright claims for "AI-original content."
Anthropic's $1.5 billion is a large number, but given its current revenue scale, this is money spent to buy certainty. The cost of data compliance is always cheaper than the uncertainty of litigation.
Sources: AI in litigation series: An update on AI copyright cases in 2026 (Norton Rose Fulbright); CocoLoop; Bartz v. Anthropic Settlement: What Authors Need to Know (The Authors Guild); AI Copyright Lawsuit Developments in 2025: A Year in Review (Copyright Alliance); Generative AI Lawsuits Timeline (Sustainable Tech Partner)