Judge rules Anthropic pirated books but calls legal AI training fair use
Anthropic must pay a $1.5 billion settlement for pirating books, but a judge just ruled that legally acquired data is fair use for AI.
Policy · Source: Hacker News
What happened
A federal judge ruled that Anthropic illegally scraped books from over ten Stanford professors to train its Claude models. The company downloaded texts from shadow libraries like LibGen and PiLiMi to build a permanent central training database. This blatant piracy resulted in a massive $1.5 billion settlement covering nearly 500,000 registered works. The ruling strictly prohibits AI companies from pirating books to build permanent training libraries.
Judge William Alsup delivered a massive win for AI companies alongside the financial penalty. He explicitly ruled that training large language models on legally acquired books constitutes fair use. He compared the AI training process to a human reader internalizing themes and styles to create something entirely new. This sets a major legal precedent confirming that the act of training models is not inherently copyright infringement.
Not all authors accepted the initial settlement terms. Stanford professors Tobias Wolff and Chang-rae Lee opted out to file a separate lawsuit called Cruz v. Anthropic. They are demanding up to $150,000 per infringed work and a permanent injunction against Anthropic using their texts. Meanwhile, the Authors Guild is pushing Congress to create national legislation that would mandate strict licensing for all AI training data.
Key facts
- $1.5 billion — Settlement amount paid to authors and publishers in the Bartz v. Anthropic case
- 500,000 — Number of registered works covered by the initial settlement
- $150,000 — Statutory damages sought per infringed work in the new Cruz v. Anthropic lawsuit
- $3,000 — Approximate damages per work offered in the original Bartz settlement
- 70 — Number of infringement lawsuits filed against AI companies outside of Bartz and Cruz
Why it matters
Foundation model builders now have a much clearer legal framework for acquiring their training data. If you acquire text legally, using it to train models is protected under fair use. You no longer have to fear copyright infringement lawsuits just for the technical act of training a neural network. However, scraping shadow libraries or pirating data is now explicitly illegal and carries massive financial penalties that will bankrupt early stage startups.
The primary cost of building foundation models will rapidly shift from legal defense budgets to data licensing fees. Since pirated data is completely banned, companies must buy access to massive text repositories to stay competitive. This creates a massive barrier to entry for open source researchers and bootstrapped AI startups. Only highly funded tech giants will afford the massive transaction costs required to license millions of books legally.
For builders
Clean data pipelines are mandatory
You can no longer scrape shadow libraries or pirate sites without taking on massive legal risk. Builders must audit their data supply chains and prove they acquired every text legally. Startups ignoring data provenance will face lethal lawsuits from copyright holders who are actively hunting for infringement.
Licensing costs will crush small teams
Legal acquisition is now the absolute only safe path for foundation model training. Negotiating individual licenses for millions of works is impossible for small teams due to massive transaction costs. Data brokers who aggregate and license clean text will extract massive premiums from AI builders who have no other options.
Fair use protects model outputs
The judge ruled that models transform data rather than just replicate the original text. This protects application layer builders from copyright claims on standard AI outputs. You can safely build commercial products on top of legally trained models without inheriting copyright liability from the underlying training data.
My take
I see this ruling as a massive double-edged sword for founders building in the AI space. The fair use protection for legally acquired data is the exact legal shield we needed to keep building without fear. But the death of shadow library scraping means only the richest companies can afford the clean data required to train frontier models.
Original reporting: Hacker News. This is my rewrite and opinion.