Is it legal to train AI on copyrighted books? The answer is still complicated

The models behind ChatGPT, Gemini, Claude, and others were trained on gigantic databases of published works: hundreds of millions of books, articles, academic papers, and basically everything that exists on the internet. Most authors contributed without knowing and without consenting. It sounds illegal, but according to a TechCrunch report, reality is far more complex, and the first court decisions in the US are starting to put the puzzle together.

The 1976 law trying to make sense of LLMs

The underlying problem is that US copyright law hasn't been updated since 1976. Judges must interpret 50-year-old guidelines to resolve disputes that could shape the entire AI industry.

"Everyone is very worried right now because the law is all over the place," Jason Henderson, senior attorney and founder of JWL International's IP and Media practice, told TechCrunch. "They know the AI model was trained on a lot of stuff, and the law still hasn't caught up with that issue."

Cathy Gellis, an attorney specializing in intellectual property and technology, sums up the mood: "It's very complex, and there are a lot of raw feelings about what's happening, both for and against."

The Anthropic case: billion-dollar fine, but training deemed legal

The most emblematic case so far is that of Judge William Alsup, who ordered Anthropic to pay $1.5 billion to a group of writers whose works were used in training. At first glance, a moral victory for the authors. But Alsup actually ruled that AI training was legal. What Anthropic paid for was piracy: downloading the books from illegal shadow libraries.

In his reasoning, Alsup compared the way an LLM ingests trillions of words to a writer's study of literature: "Like any reader aspiring to be a writer, Anthropic's LLMs trained upon works not to race ahead and replicate or supplant them, but to turn a hard corner and create something different."

For Gellis, the verdict is more advantageous for AI companies. After all, what is a $1.5 billion fine for a company projecting around $200 billion in annual revenue by 2028? The central point, according to her, is the distinction the judge made: "Copyright law depends on copying, but it doesn't depend on using the work, experiencing the work, consuming the work, reading the work."

The criterion deciding these cases: compete or not compete

Most of these disputes revolve around fair use, the exception in American copyright law that allows copyrighted works to be used without explicit permission for purposes such as criticism, parody, and education. Judges weigh factors like the purpose and nature of the use, the amount used, and the impact on the market, especially whether the use is "transformative" enough.

According to Henderson, a pattern is emerging: "What tends to win is: if you're training on someone's property because your purpose is to compete directly, courts disapprove. If what you're doing won't compete, courts tend to find a way to say it's fine."

The example is the case Thomson Reuters v. Ross Intelligence. Reuters sued Ross for copying its content to build an AI-based legal platform that would compete directly with it. Judge Stephanos Bibas ruled that it was not fair use: "Ross's use is not transformative because it does not have a 'further purpose or different character' from Thomson Reuters's."

Authors could, in theory, argue that chatbots compete with them by generating synthetic books. But that argument hasn't won in court yet.

Training AI and registering AI-generated work are different things

Gellis makes an important distinction: thinking about copyright in the context of training is quite different from thinking about the registration of AI-generated content.

In Thaler v. Perlmutter, the court ruled that a work 100% generated by AI is not eligible for copyright. That opens another can of worms: how do you prove whether something was generated by AI and, if so, what percentage? Gellis uses an analogy: "If you write your novel in Word and run the spell checker, we're comfortable saying Word doesn't own your novel. AI is forcing us to look at a bunch of decisions we've ignored for a while."

What this changes for developers in Brazil

These are all US decisions and have no direct legal effect here. But they matter for two practical reasons. First: the models most Brazilian devs use via API (OpenAI, Anthropic, Google) were trained under these rules still under construction, and the legal stability of these companies affects product, pricing, and availability. Second: Brazil is debating its own regulation, and the debate over AI copyright is still open, without a ruling that settles what's allowed to be trained on.

For those who build or fine-tune models with their own data, the practical lesson from the American courts is direct: the origin of the data matters. Anthropic wasn't punished for training, it was punished for pirating. And the competition test from the Ross case suggests that training on a competitor's data to build a product that replaces it is the riskiest scenario. Using data obtained lawfully and for a distinct purpose is the safer ground.

As Gellis sums up, nothing is settled: "What you see is that the early moves are being influential, and that influence can be undone if other courts decide things differently. But in the meantime, all these decisions are shaping everything. It would be foolish for AI companies to ignore them." The same goes for those building software on top of them.