A federal judge in San Francisco has approved a $1.5 billion copyright settlement against Anthropic — the largest in history — after the AI company used pirated copies of more than 482,000 books to train its Claude chatbot. Authors will receive roughly $3,000 per book. The ruling closes the first major settlement in a wave of AI copyright lawsuits still working through US courts.
What is this case? Thriller novelist Andrea Bartz and two other authors sued Anthropic in 2024, alleging the company downloaded millions of pirated books from the internet and used them to train Claude. US District Judge William Alsup issued a mixed ruling last summer: training AI on copyrighted books qualified as fair use, but Anthropic’s acquisition of those books through pirate websites was unlawful. The settlement resolves the piracy claim while leaving the fair-use precedent standing.
🔍 THE BOTTOM LINE
Anthropic is paying $1.5 billion — and calling it a victory. That sum buys closure on the piracy claim while preserving the legal precedent that training AI on copyrighted books is fair use. Every other AI lab facing similar lawsuits is watching this template: accept liability for how you acquired the data, not for what you did with it.
The Numbers That Matter
The AP News reporting lays out the settlement in stark detail:
- $1.5 billion total payout — described by plaintiff attorney Justin Nelson as “the largest known copyright recovery in history”
- 482,000+ books covered by the class-action ruling
- 91% of books have been claimed by authors or publishers who are now due payment
- ~$3,000 per book — the average payout per title
- District Judge Araceli Martínez-Olguín approved the settlement on Monday, calling it “meaningful relief”
Judge Alsup, who issued the preliminary approval last September before retiring, had drawn a line that now defines the AI copyright landscape: using books to train models is legal, but downloading them from pirate sites is not. That distinction is the settlement’s foundation.
What Anthropic Is Actually Admitting
Anthropic’s deputy general counsel Aparna Sridhar framed the settlement as a win, highlighting Alsup’s fair-use ruling as “a landmark showing that training AI on books is fair use under copyright law.”
Read that carefully. Anthropic is not apologising for training Claude on copyrighted works. It is paying $1.5 billion for the acquisition method — downloading from pirate websites — while preserving the legal right to train on those same books if obtained lawfully. The settlement is a penalty for theft, not for use.
This distinction matters enormously for the dozens of other AI copyright cases still pending. If every lab can settle the piracy claim while keeping the fair-use precedent, the path is clear: pay for stolen data, keep training.
The NZ Angle
New Zealand has no equivalent copyright precedent. The Copyright Act 1994 predates the AI training era by three decades, and the previous government’s AI policy paper acknowledged the gap without proposing legislative reform. If a Kiwi author’s work was among the 482,000 pirated titles, they can claim under the US class action — but there is no domestic mechanism for pursuing an AI company for training on copyrighted works.
The broader question for NZ: if fair use on training data becomes the global standard through US court rulings, does NZ copyright law need to explicitly address AI training? The Ministry of Business, Innovation and Employment has not signalled a position. Australia’s AI copyright office is further along, but even Canberra has not legislated on training data.
The Industry-Wide Problem
Anthropic is the first to settle. It will not be the last. The pattern — download pirated books, train models, settle the piracy claim, keep the fair-use precedent — is available to every AI lab that did the same thing, which is most of them.
This is not an Anthropic-specific problem. It is a structural issue in how the AI industry built its training datasets between 2022 and 2024. OpenAI faces similar lawsuits. Meta faces similar lawsuits. Google faces similar lawsuits. Google’s Gemini publisher lawsuit is still in discovery. The question of who owns AI-generated code is separate but related.
The $1.5B figure sets a ceiling — or a floor, depending on how many books were involved and how the per-book rate compares across cases. Anthropic’s 482,000 books at $3,000 each is one data point. If another lab used two million pirated books, the math gets uncomfortable fast.
What Happens Next
The settlement approval triggers distribution to the 91% of authors who have already filed claims. Nelson said distributions will be made “as promptly as possible.” The remaining 9% of unclaimed books represent authors who have not yet participated in the class action — they still can, but the window is narrowing.
For the broader AI copyright landscape, the key question is whether other defendants adopt the same strategy: settle the piracy claim, fight for fair use on training. OpenAI has signalled it may fight rather than settle, arguing its data acquisition was lawful. Meta has taken a similar posture. If either loses the fair-use argument — which Alsup’s ruling does not guarantee — the calculus changes.
❓ FAQ
Will authors actually receive $3,000 per book? The $3,000 figure is an average. The settlement structure may weight some books differently based on length, commercial success, or other factors. The AP reports that 91% of the 482,000 books have been claimed, meaning most authors will receive payment — but the exact per-book amount may vary.
Does this mean AI training on copyrighted books is legal? Partially. Judge Alsup ruled that training AI on copyrighted books qualifies as fair use — meaning the use is legal. But acquiring those books through pirate websites is not. The settlement resolves the piracy claim without challenging the fair-use ruling.
Could this settlement affect NZ copyright law? Not directly — US settlements do not bind NZ courts. But the fair-use precedent may influence how NZ courts and policymakers think about AI training. NZ’s Copyright Act 1994 has no fair-use provision (it has fair dealing, which is narrower), so the gap remains.
Are other AI companies facing similar lawsuits? Yes. OpenAI, Meta, and Google all face active copyright lawsuits from authors and publishers. Anthropic’s settlement is the first major resolution. The others are in various stages of litigation.
What is fair use vs fair dealing? Fair use is a US legal doctrine that allows limited use of copyrighted material without permission. Fair dealing is the equivalent in NZ and other common-law jurisdictions, but it is narrower — it covers specific purposes like criticism, review, and research. Whether AI training qualifies under fair dealing has not been tested in NZ courts.
🔍 THE BOTTOM LINE
$1.5 billion sounds like a punishment. It is also a business decision. Anthropic is paying to make a piracy problem disappear while keeping the legal right to train on copyrighted books. If that template holds across the industry, the cost of building AI models just went up by the price of stolen data — but the right to train on that data remains intact. Authors get paid. AI labs keep training. The copyright system bends without breaking.
