The simple version
- AI models learn by reading huge amounts of text and looking at huge numbers of pictures.
- A lot of that material is owned by someone: authors, newspapers, photographers.
- Copyright law decides when you may copy someone’s work without asking.
- Different countries answer differently, and courts are still deciding.
- So far, learning from work is often allowed, but copying it back out is a problem.
Why it matters now
On 3 November 2026 an adviser to the EU’s top court, Advocate General Maciej Szpunar, gave his opinion in the first generative AI case to reach the Court of Justice of the European Union. A Hungarian publisher had sued Google over its Gemini chatbot. He said a chatbot that fetches news articles and reuses their wording so closely that you no longer need the original may infringe publishers’ rights, while training a model is, as a rule, allowed under the EU’s text and data mining exception. His opinion is not binding; the judges will rule later. Our brief has the details.
That one opinion touches every question in this debate: what may go into an AI model, what may come out, and who pays.
How it works, step by step
Think of a student preparing for an art exam by visiting a gallery. Looking at the paintings and learning how light falls is fine. Taking photographs of every painting to study at home is a grey area. Selling exact copies of the paintings is not fine. AI copyright fights are about where on that line each step sits.
Step 1: Collect the data
Before a large language model can do anything, it goes through pretraining: it is shown an enormous collection of text, images or sound. Much of this is gathered from the public web by automated programs, and some comes from books, news archives and licensed datasets. Gathering it means making copies, and copying is what copyright controls.
Step 2: Turn the data into patterns
During training the model does not file away each article like a library. It adjusts billions of internal numbers, called weights, to get better at predicting what comes next. In Getty Images v Stability AI, the High Court of England and Wales described an image model’s weights as “purely the product of the patterns and features which they have learnt” and found that they “do not store an infringing copy” of the photos used to train it.
There is a catch. A model that sees the same passage many times, such as a famous speech or a popular song lyric copied across thousands of websites, can learn it so well that it can repeat it almost exactly. Researchers call this memorisation. It is usually rare for any single work, but it is the reason output, not just training, matters in the cases below.
Step 3: Check the law where the copying happened
Copyright is national, so the answer depends on the country.
- European Union: the 2019 copyright directive created an exception for text and data mining, the automated analysis of large amounts of material. Anyone with lawful access may mine works unless the rightsholder has opted out in a machine-readable way, for example in a website’s settings. Research bodies get a wider version with no opt-out.
- United States: there is no specific exception. Instead courts apply fair use, a four-part test that weighs the purpose of the use, the nature of the work, how much was taken and the effect on the market for the original.
- United Kingdom: the Getty case turned partly on geography. Getty accepted there was no evidence the model had been trained in the UK, so it dropped its main claim about training, and the court only ruled on whether the finished model was itself an infringing copy. It found it was not.
Step 4: Look at what comes out
Even where training is allowed, the output can still be a problem. If a chatbot reproduces a news article almost word for word, or an image generator spits out a near-copy of a famous photo, that copy can infringe. This is the Advocate General’s distinction: publishers’ rights protect how a story is written, not the facts in it.
Step 5: Settle, license or sue
Many disputes end in deals. The US Copyright Office’s May 2025 report on AI training noted that licensing agreements for AI training “are fast emerging in certain sectors”, though their availability “so far is inconsistent”. Where no deal is struck, rightsholders sue.
A worked example: a novel, a newspaper and a chatbot
Imagine a novelist’s books and a local newspaper’s archive both end up in a model’s training data.
- Training: in the EU, if the newspaper had opted out of text and data mining and the developer ignored it, the newspaper has a stronger case. In the US, a judge would weigh fair use. In June 2025 a US federal judge ruled in Kadrey v. Meta that training on authors’ books was fair use in that case, but called the ruling limited: it stood “only for the proposition that these plaintiffs made the wrong arguments”, and suggested a case built on harm to the market for the authors’ books might succeed.
- Output: if a user asks the chatbot for today’s local news and it copies the newspaper’s article almost word for word, that answer may infringe regardless of how training was judged.
- A photographer’s picture: if an image generator, asked for “a dog on a beach in the style of a stock photo”, produces something new, that is learning. If it produces a near-exact copy of one photographer’s picture, complete with a faded watermark, that looks much more like copying, and watermarked outputs were part of what Getty complained about in its UK case.
- Business model: the Advocate General suggested a provider is liable where such copying is built into its business model, or where it has not taken credible steps to stop it.
What people get wrong
“If it’s on the internet, it’s free to use.” Publicly visible is not the same as copyright-free. Most web pages are protected.
“The model contains a copy of everything it read.” Courts so far have accepted that model weights are learned patterns, not stored copies. But models can sometimes reproduce passages they saw many times, which is why output matters.
“Courts have settled this.” They have not. The cases so far are early, narrow and in different countries. The CJEU ruling is still to come.
“Fair use exists everywhere.” It is a US doctrine. The EU and UK use specific listed exceptions instead.
Where it is heading
Three threads to watch. First, the CJEU judgment in Like Company v Google (C-250/25), which will be the EU’s first word from its top court on generative AI and copyright. Second, US courts testing the market harm argument the Kadrey judge flagged. Third, licensing: the US Copyright Office said in May 2025 that government intervention would be “premature at this time” and that licensing markets should continue to develop. Meanwhile, labelling and watermarking rules are making it easier to trace what AI produced, if not yet what it learned from. For how the wider rules differ by country, see How Is AI Regulated Around the World?
Key terms
Sources
- Court of Justice of the European Union, press release No 155/26, Advocate General’s opinion in C-250/25 Like Company v Google, 3 November 2026. curia.europa.eu
- Directive (EU) 2019/790 on copyright in the Digital Single Market, Articles 3 and 4 (text and data mining). eur-lex.europa.eu
- Getty Images v Stability AI [2025] EWHC 2863 (Ch), judgment of 4 November 2025, paragraph 600. judiciary.uk
- US Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version), May 2025. copyright.gov
- Kadrey v. Meta Platforms, N.D. Cal., order on summary judgment, 25 June 2025, case 3:23-cv-03417; quotes as summarised by Goodwin, 25 June 2025. goodwinlaw.com
- 17 U.S.C. § 107, the four fair use factors. copyright.gov
- Latham & Watkins, on why Getty dropped its training claim. lw.com