Copyright and Creativity

AI models learned from millions of books, songs and pictures that people made. Was that fair use, a new kind of reading, or the largest copying exercise in history? Here are the arguments, the court cases and where they stand, and what creators can do now.

A Kuala Lumpur illustrator types her own name into an image generator and watches it produce pictures very like her style — the brushwork and palette she spent ten years developing. She never agreed and was never paid. Across the world, a novelist finds his books on a list of titles downloaded from a pirate library to train a chatbot. And an engineer at an AI company says, sincerely, that the model did nothing different from what every human artist does: it looked at a great deal of work and learned. All three are describing the same event. Which description the law accepts is one of the most important open questions about AI.

This page explains a fast-moving area of law for general understanding. It is not legal advice; laws differ by country and change often. For a decision that matters, ask a qualified lawyer.

Why training raises a copyright question

Copyright gives the creator of an original work — a novel, photo, song, painting, piece of code — the exclusive right to copy it and to make adaptations of it, for a limited time. To build a generative AI system, developers collect enormous amounts of training data, much of it by web scraping: billions of web pages, images with captions, digitised books, code and recordings. During training those files are copied, stored and processed many times. The finished model does not usually contain the works as files; it contains billions of numbers (parameters) adjusted to capture patterns across them.

So there are really three questions, and courts treat them separately:

  1. Input: Was it lawful to copy the works to train — and does it matter whether they were bought, licensed, scraped or pirated?
  2. The model: Is the trained model itself a copy, especially where it has memorised material?
  3. Output: When it produces something close to a protected work — a lyric, a famous character, an article — who is responsible?

Consensus Lawyers on all sides agree that copyright protects expression, not ideas or styles. Nobody owns "watercolour landscapes" or "detective novels". The fight is over whether wholesale copying, as a step towards learning patterns, needs permission.

Fair use, data mining and the opt-out

Different countries answer the input question with different tools.

US law has a flexible defence called fair use. Judges weigh four factors: the purpose of the use (is it "transformative" — does it add something new?), the nature of the work, how much was taken, and — often decisive — the effect on the market for the original. AI companies compare training to Google Books' searchable index, accepted as fair use in 2015; rights holders say the products compete with the works they were built from.

In May 2025 the US Copyright Office issued a pre-publication report on training. It concluded that some training uses will be fair and others will not, that market harm is central, and floated "market dilution" — a flood of AI work competing with the originals — as a possible harm. The report is advice, not law, and was released the day before the head of the Office was dismissed, which has left its status contested As of 2026.

The court cases so far

Dozens of lawsuits have been filed since 2023, mostly in the United States. The rulings so far point in different directions, partly because each turns on its own facts. Filter by place, and remember that statuses change month by month As of October 2026.

  1. Dec 2023New York Times v. Microsoft and OpenAIThe Times alleges millions of its articles were used without permission and that chatbots can reproduce them. Judge Sidney Stein let the core claims proceed in 2025, and other newspaper suits were joined to it. In June 2026 the Times amended its complaint, narrowing some claims. No trial or settlement had been reported as of October 2026. A 2025 order in the case to preserve user chat logs is discussed on Privacy and Surveillance.
  2. Jun 2025Bartz v. Anthropic — split rulingJudge William Alsup held that training Claude on books was "exceedingly transformative" and fair use, including scanning print books Anthropic had bought. But keeping millions of books downloaded from pirate libraries (Library Genesis and the Pirate Library Mirror) in a central library was not, and that part was headed for trial.
  3. Jun 2025Kadrey v. Meta — Meta wins, narrowlyTwo days later Judge Vince Chhabria granted Meta summary judgment on training its Llama models, because these authors had not shown market harm. He stressed the ruling did not mean Meta's use was lawful and suggested "market dilution" could defeat fair use in other cases. A separate claim about Meta sharing pirated books while torrenting continues, with a hearing scheduled for February 2027 As of October 2026.
  4. Sep 2025 – Jul 2026Bartz v. Anthropic — $1.5 billion settlementAnthropic agreed to pay US$1.5 billion, roughly US$3,000 for each of about 482,000 books on the class list, and to destroy the pirated files. After Judge Alsup retired, Judge Araceli Martínez-Olguín granted final approval on 20 July 2026. The release covers past conduct only, not future training or outputs. It is the largest copyright settlement reported in the US.
  5. Sep 2026Thomson Reuters v. Ross IntelligenceThe Third Circuit Court of Appeals held that Ross's use of Westlaw headnotes to train a legal research tool was not fair use — the first federal appeals ruling on AI training. Ross's tool competed directly with Westlaw and was not generative, so its reach is debated.
  6. 2024–2026Music labels v. Suno and UdioUniversal, Sony and Warner sued the two music generators in June 2024. In late 2025 Universal and Warner settled with Udio in deals combining licences with plans for new licensed services; other claims continued.
  7. Jun 2025Disney and Universal v. MidjourneyThe first major Hollywood studio suit, over characters such as Darth Vader and the Minions; Warner Bros. Discovery later filed a similar case. No outcome had been reported as of October 2026.
  8. Nov 2025Getty Images v. Stability AI (High Court, London)Getty dropped its main training claim because the training happened outside the UK. On what remained, Mrs Justice Joanna Smith ruled on 4 November 2025 that the Stable Diffusion model was not an "infringing copy", because it does not store Getty's images. Getty won only a narrow trade-mark point about some outputs showing its watermark. Permission to appeal was granted in December 2025; no appeal decision had been reported as of October 2026.
  9. Nov 2025GEMA v. OpenAI (Munich Regional Court)The German music-rights society won: the court held that song lyrics memorised by GPT models were reproduced in the model itself, and that the EU data-mining exception did not cover this. The judgment was not final and could be appealed.

This guide’s framing Read together, three patterns stand out so far. First, how the data was obtained matters: buying or licensing helps, piracy hurts. Second, market harm is where the fights are won and lost. Third, memorisation — a model reproducing protected text almost word for word — is the weak point for developers, especially in Europe.

Who owns what AI makes?

The output question is covered in practical terms on Making Images and Media; here is the short version. In January 2025 the US Copyright Office concluded that material generated purely from a prompt is not protected, while human selection, arrangement and real creative editing can be. In March 2025 the US Court of Appeals for the D.C. Circuit held in Thaler v. Perlmutter that copyright requires a human author, and on 2 March 2026 the Supreme Court declined to hear the appeal.

Malaysia's position is less settled. A government minister has said that work produced autonomously by AI, without significant human creative contribution, is not currently eligible for protection under the 1987 Act, and the MyIPO consultation asks whether that should change As of 2026. Other countries differ: in 2023 the Beijing Internet Court protected an AI-assisted image because of the user's creative choices in prompting and refining it.

Creators and developers: the real disagreement

Debated Behind the legal technicalities is a disagreement about fairness that no court ruling will fully settle. Try each side.

The question

Should AI developers need permission to train on copyrighted work?

Choose one to hear its case.

Other views
  • Creators' organisations (for example the Authors Guild in the US, and many illustrators and musicians who have spoken out in Malaysia and elsewhere): training without permission is infringement, and settlements like Bartz show the copying had real value.
  • AI developers and many technologists: training is transformative fair use or lawful data mining; the problem, if any, is specific outputs, which can be filtered.
  • Some economists and policy writers favour collective licensing — a fund or collecting society that pays creators, as radio does for songs — rather than either free use or one-by-one permission.

Licences, opt-outs and practical tools

While courts deliberate, the market has moved. Many AI companies now sign licensing deals: with news publishers (the Associated Press, Axel Springer, News Corp and others have announced deals with OpenAI since 2023), with platforms (Reddit licensed its posts to Google in 2024), with stock-image libraries, and — after litigation — with record labels. Critics note that individual creators rarely see much of this money.

If you publish work online, these are the main levers. Tap each one.

Quick check

In Bartz v. Anthropic (June 2025), what did the judge treat differently from training itself?

Training on lawfully bought books was held to be fair use; building a library from pirated copies was not, which led to the US$1.5 billion settlement.

Under the EU's 2019 copyright directive, commercial text and data mining of online works is allowed unless…

That reservation is the "opt-out". The AI Act adds a duty for general-purpose model providers to respect it.

// Pause and reflect

  1. Is an AI model learning from a painting more like an art student studying it, or like a photocopier? Where does the analogy break?
  2. If you make things — writing, photos, music, code — would you opt out of AI training, license it, or let it be used freely? What would change your mind?
  3. Should Malaysia follow the EU's opt-out, Japan's permissive rule, or something else? Who gains and who loses under each?

Your notes are saved privately in this browser.

Key terms on this page

Sources & further reading