The question looks simple: if a company trades in a writer's books to build a machine that then writes in their place, is that theft? The answer American courts are giving pleases neither side - and that is exactly why it matters to anyone who writes, films or creates anything in digital form.

Last year judge William Alsup ordered Anthropic to pay 1.5 billion dollars to a group of writers whose work ended up in the training data of its models. At first glance - a win for authors. Read the ruling and the picture changes: Alsup ruled that the training itself is legal. What Anthropic paid for was how it obtained the books - pirated copies from illegal online libraries.

Alsup justified this with a comparison everyone now quotes: the model, he wrote, read the works "like any reader aspiring to be a writer" - not to copy them out, but to create something else. With one sentence, training on other people's labour got judicial legitimacy, and the penalty stayed attached to the method of acquisition.

A billion and a half for a company projecting two hundred billion

"I think that is generally good news for AI training," says Cathy Gellis, a lawyer specialising in intellectual property and technology. "Copyright rests on copying, but it does not rest on using the work, on experiencing it, on reading it."

When 1.5 billion is measured against projections of around 200 billion dollars in annual revenue by 2028, the penalty stops being a penalty and becomes a budget line. That is not a conspiracy theory - it is arithmetic every chief financial officer does before deciding whether something is worth risking.

A law from 1976 judging technology from 2026

American copyright law has not been changed since 1976. Judges today have to interpret rules written half a century before the first large language model, and it shows in the outcome - every court decides differently.

"Everyone is very worried at the moment because the law is scattered all over the place," says Jason Henderson, senior counsel and founder of the intellectual property and media practice at JWL International. "They know the model was trained on an enormous amount of material, and the law really has not caught up with that question."

What is emerging as a line, though, is competition. When Thomson Reuters sued the research firm Ross Intelligence for using its content to build a rival legal platform, judge Stephanos Bibas ruled that this was not permitted use - the use had "no different purpose or different character" from the original. In other words: train on other people's work all you like, but not in order to kill the market of the person you are taking from.

Writers could argue exactly that - that chatbots generating synthetic books are eating their market. So far that argument has not succeeded in court.

And when the machine is the author?

The second half of the story is the reverse: what if the work was created by a machine? In Thaler v Perlmutter the court ruled that a work created one hundred percent by artificial intelligence is not subject to copyright at all. The question that follows is more interesting than the ruling itself - how do you even prove what percentage of a text a machine wrote?

"If you write a novel in Word and run a spellcheck, we are somehow comfortable saying that Word does not own your novel," says Gellis. "This forces us to look at a whole series of decisions we have long ignored."

Almost every major artificial intelligence company is currently in ongoing litigation on these questions, which means there will be no definitive answer soon. Until then, what Gellis formulates most precisely holds: the first rulings shape the whole game, even if they are later overturned. The law has not caught up with the question, but the decisions are already being made - and they will not wait for someone to refresh a law from 1976.