Is training AI on copyrighted books legal? Courts are still fighting it out

Last year, Judge William Alsup ordered Anthropic to pay $1.5 billion to settle copyright claims. The money goes to authors whose work was used to train Claude, which at first glance looks like a win for writers. But what Alsup actually found unlawful wasn't AI training itself. His objection was narrower: Anthropic had obtained the books by illegally downloading them from unauthorized online "shadow libraries." Using copyrighted works as training material, on the other hand, he found to be lawful. According to reports, Alsup's opinion compared the way Anthropic's LLM absorbed vast amounts of text — not to copy or replace earlier works, but to produce something entirely new — to how an aspiring writer studies literature.

Gellis sees the ruling as actually working in AI companies' favor. The fine is barely a dent for a company like Anthropic, whose annual revenue is projected to reach $200 billion by 2028. She points out that Alsup treating AI training as closer to reading and experiencing a work, rather than straight-up copying it, is a positive signal for AI training more broadly.

Where copyright law draws the line: copying versus not

U.S. copyright law was enacted in 1976 and hasn't been substantially updated since, which leaves courts trying to apply 50-year-old principles to the age of AI. "Everyone's anxious right now because the law is a mess," said Jason Henderson, an IP and media attorney. He noted that everyone knows AI models have trained on enormous amounts of material, but the law simply hasn't caught up to that question yet.

Most of these cases hinge on the doctrine of "fair use." Fair use is an exception that allows copyrighted works to be used without the owner's permission for purposes like criticism, parody, or education, and courts weigh it by looking at the purpose and nature of the use, how much material was taken, and the effect on the market for the original work. "Copyright has always been about protecting and growing a market," Henderson said. If AI training is aimed at creating something that directly competes with the original work, courts tend to view it unfavorably; if it doesn't compete, they're more inclined to allow it.

The line drawn by the Ross Intelligence ruling

Henderson pointed to Thomson Reuters v. Ross Intelligence as an example. Media and technology company Thomson Reuters sued research firm Ross Intelligence, alleging that Ross copied its content to build a competing AI-powered legal research platform. In his opinion, Judge Stephanos Bibas found that Ross's use didn't have "a different purpose or character" from Thomson Reuters' own, and rejected the fair use defense. The core problem was that Ross had trained on Thomson Reuters' content to build a platform that directly competed with it. Authors could try to make a similar argument — that chatbots trained on their work compete with them by generating new, synthetic books — but no court has accepted that logic yet.

Who owns the copyright on AI-written text?

Gellis argues that discussions of AI and copyright need to separate two very different questions: using copyrighted works to train AI, and whether AI-generated output itself deserves copyright protection. In Thaler v. Perlmutter, the court ruled that a work created 100% by AI isn't eligible for copyright protection. Applying that standard raises a whole new set of practical questions — namely, how do you prove whether a given work was AI-generated, and if so, how much of it was. Gellis compared it to writing a novel in Microsoft Word and running spell-check: nobody thinks Word owns the copyright to that novel. In the same way, AI is forcing courts to revisit questions people had long put off answering.

Case	Core issue	Court's ruling
Authors v. Anthropic	Use of copyrighted works in AI training	Training ruled legal; only illegal downloading required compensation ($1.5B)
Thomson Reuters v. Ross Intelligence	Training aimed at building a competing platform	Fair use defense rejected
Thaler v. Perlmutter	Copyright status of 100% AI-generated work	Not eligible for copyright protection

The litigation is far from over

Most AI companies are still tangled up in related lawsuits, so a clear resolution isn't likely anytime soon. Gellis notes that while the rulings coming out now carry weight as early precedent, that influence could be overturned if other courts reach different conclusions. Which precedents ultimately stick won't be clear until later stages of litigation play out. Still, AI companies can't afford to ignore the rulings that have already come down, because they're already shaping how the whole industry moves.

Editor's take

The message Alsup's ruling sends to the industry is clear: training is fine, only stolen material is a problem. That's essentially a get-out-of-jail-free card for AI companies — a $1.5 billion penalty is so small relative to their revenue that copyright lawsuits are more likely to become just another line item in the cost of doing business than an actual brake on AI companies' growth. Paying the fine and carrying on becomes the rational choice under this structure.

This pattern isn't new. Music copyright battles in the Napster era, and the lawsuits over Google Books scanning, both ended the same way: the technology spread first, and the law drew boundaries only after the fact. Courts are once again reaching for the old copyright framework of "copying versus learning" and applying it to AI, and within that framework, the outcome is almost bound to favor AI companies. Copyright law was designed to punish copying, not to prevent reading and learning.

For domestic companies and content creators, there are two things worth taking away from all this. First, when handing over your own data for AI training, make sure the contract clearly spells out sourcing and licensing. Second, before using AI-generated output commercially, check whether it actually qualifies as protectable creative work. Under the Thaler v. Perlmutter standard, content generated entirely by AI may not receive legal protection, so it's practically safer to keep a record of human editing and revision in the process.

In the coming months, expect the Thomson Reuters v. Ross Intelligence standard — whether the use was aimed at direct competition — to keep coming up in similar lawsuits against OpenAI, Google, and Meta. The trend seems likely to solidify: companies that release AI products directly competing with a given industry will face growing litigation risk.