SAN FRANCISCO – A federal court approved Anthropic’s $1.5 billion settlement with a class of authors and book publishers on Monday, resolving what became the largest copyright payout in United States legal history. The settlement distributes $3,000 per work across approximately 500,000 copyrighted titles. It also arrives alongside a legal ruling that may matter more than the money: a determination that training an AI model on copyrighted text constitutes fair use under federal law.
The case began as a class action filed in the Northern District of California, in which authors and publishers accused Anthropic of using their books without authorization to train its Claude AI systems. What emerged during discovery was more specific: significant portions of Anthropic’s training data had come from Library Genesis and Pirate Library Mirror, two platforms that operate as repositories of pirated literary and academic texts. Anthropic did not dispute that it had downloaded books from those platforms. What it disputed was whether doing so constituted copyright infringement.
Library Genesis, also known as LibGen, hosts an estimated 4.5 million books and 85 million academic articles, the large majority in violation of copyright. Pirate Library Mirror maintains a similar collection. Both platforms are routinely blocked by internet service providers in the United Kingdom and several European countries. That Anthropic sourced training material from these platforms was not, by the time of the settlement, in serious dispute. The $1.5 billion represents the damages the parties agreed on for that sourcing, not an assessment of what Anthropic gained commercially from using that specific collection.
U.S. District Judge William Alsup, during the pretrial phase, ruled that training an AI model on copyrighted text is transformative and therefore protected as fair use. He also found that Anthropic’s method of obtaining training data through piracy platforms was “illegal on its own terms” – a separate violation the fair use finding did not protect. That ruling effectively settled the legal argument over training. Judge Araceli Martinez-Olguin signed final approval of the $1.5 billion settlement on Monday.
The structure of what emerged matters. Anthropic did not pay $1.5 billion because AI training infringes copyright. It paid because the specific method it used to acquire training material – downloading from piracy sites – was illegal on its own terms. The fair use ruling, which found in Anthropic’s favor on the training question itself, has no binding precedential value because the settlement was reached before appeal. But it will be cited.
Publishers who sued Google over unauthorized Gemini AI training on copyrighted books are watching the Anthropic outcome closely, as are plaintiffs in still-pending cases against Meta, Midjourney, and OpenAI. Each of those cases will be argued under different factual circumstances and before different judges. None is bound by the Alsup ruling. But AI company legal teams now have a federal district court determination – affirmed through settlement rather than overturned on appeal – that they can point to when the training question arises again.
For the rights holders in the class, the settlement ends this specific dispute at a known number. A novelist with one book in the class receives $3,000. A publisher with a large backlist receives proportionally more. The settlement contains no provision restricting Anthropic’s future training practices, no requirement to license authors’ works going forward, and no mechanism for ongoing compensation if Claude models trained on the settled works continue to generate commercial value for the company.
Many authors view the outcome as an incomplete win. The money is real but the precedent cuts against them. If AI training is fair use, the avenue available to any future plaintiff is not that their work was used without permission – it is that their work was obtained through piracy. That is a narrower argument, dependent on provable sourcing violations rather than on a theory of what AI training does to the value of copyrighted material.
Anthropic’s commercial position has not been visibly constrained by the litigation. Claude models are powering institutional deployments including the European Parliament’s internal AI platform, built alongside systems from OpenAI, Meta, and Mistral. The company has continued releasing new model versions during the lawsuit’s pendency. The $1.5 billion settlement – significant in absolute terms – represents a known and bounded legal cost for a company that has raised billions in venture capital and holds expanding enterprise contracts.
The settlement does not resolve the broader question of how copyright law applies to AI systems once training has occurred. Whether Claude’s outputs can infringe copyright when they closely reproduce passages from training data remains open. Whether authors whose works were used through legitimate channels – not through Library Genesis – have any copyright claim at all is equally unresolved. The Northern District of California has produced several defining AI copyright determinations in recent years. None of them has produced a coherent framework. This settlement adds a data point but not a doctrine.
The authors in the class now have their checks. What they do not have is a rule preventing the next AI company from doing the same thing – as long as it sources its training data more carefully than Anthropic did.
