Every generative AI product raises two copyright questions. First, can you train a model on books, articles, images and code that someone else owns? Second, who owns what the model produces, and can that output itself infringe someone’s rights?
Courts and governments are answering these questions at different speeds and in different ways. The United States is deciding case by case under fair use. The EU has a statutory exception with an opt-out. The UK has stepped back from reform. Japan and Singapore are permissive, and India is debating a licensing model of its own.
This guide explains where each major market stands in October 2026, what the leading rulings actually decided, and what a business building or buying AI should do now. It is general information, not legal advice for your situation.
Why training raises a copyright question at all
Copyright gives creators control over copying their work. Training a model usually involves making copies: scraping or downloading works, storing them in a dataset and processing them many times. Unless an exception applies, each of those copies needs permission.
Every legal system has exceptions, but they differ. The US uses a flexible “fair use” test that courts apply to the facts of each case. The EU, UK, Japan and Singapore use specific “text and data mining” (TDM) exceptions, which permit automated analysis of works under defined conditions. India relies on a narrower “fair dealing” list.
Outputs raise a separate question. A model may produce something new, or it may reproduce part of a work it was trained on. The second risk is often called “memorisation”, and courts are starting to treat it seriously.
United States: fair use, decided case by case
Fair use weighs four factors, including whether the use is “transformative” (serves a different purpose from the original) and whether it harms the market for the work. In June 2025 two federal judges in California reached the first major rulings on generative AI training, and both found training itself could be fair use.
In Bartz v. Anthropic, the court found training on lawfully acquired books to be fair use but treated pirated library copies differently. The case then settled for about $1.5 billion, and the settlement received final court approval in July 2026. In Kadrey v. Meta, the judge ruled for Meta on the evidence before him but warned that a better record of market harm could change the result in future cases.
The first appellate ruling went the other way on different facts. On 29 September 2026 the Third Circuit affirmed in Thomson Reuters v. Ross that copying Westlaw headnotes to train a competing, non-generative legal research tool was not fair use. Ross has said it will seek Supreme Court review.
The US Copyright Office’s report on generative AI training, released in pre-publication form in May 2025, takes a middle path: some training uses will be fair, others will not, especially where outputs compete with the works used. Many other cases, including claims by news publishers, are still pending.
European Union: an opt-out regime with transparency duties
The 2019 Copyright in the Digital Single Market Directive contains two TDM exceptions. Article 3 covers research organisations. Article 4 covers everyone else, including commercial developers, but only where the rightsholder has not reserved its rights “in an appropriate manner”, such as machine-readable means for content published online.
The AI Act adds obligations for providers of general-purpose AI models. Article 53 requires a policy to comply with EU copyright law, including respecting those opt-outs, and a public summary of training content using a Commission template published in July 2025. The voluntary General-Purpose AI Code of Practice expects signatories to use crawlers that honour signals such as robots.txt.
Courts are testing the limits. In November 2025 the Munich Regional Court held in GEMA v. OpenAI that song lyrics memorised by a model and reproducible through simple prompts were copies in the model, and that the TDM exception did not cover this. The ruling is first-instance. The Court of Justice is also considering its first generative AI reference, Like Company v. Google (C-250/25), about chatbot outputs that reproduce press content.
United Kingdom: no commercial TDM exception, and no reform yet
UK law permits text and data mining only for non-commercial research. Commercial training on UK copyright works therefore needs a licence unless it happens outside the UK, which is exactly what tripped up Getty Images.
In Getty Images v. Stability AI, judgment given on 4 November 2025, Getty dropped its main training claim because the model was trained abroad. The High Court then rejected the secondary infringement claim, finding that Stable Diffusion did not store copies of Getty’s images. Getty succeeded only on narrow trade mark points about historic watermarks, and it has permission to appeal on the copyright question.
On policy, the government consulted on a TDM exception with an opt-out. In its report to Parliament of 18 March 2026, it said that option was no longer its preferred approach and announced no immediate legislation. The status quo, and the uncertainty, continues.
Asia: Japan, Singapore, China and India
Japan’s Copyright Act, Article 30-4, allows use of works for data analysis where the purpose is not to “enjoy” the expression in the work. The Agency for Cultural Affairs’ 2024 guidance narrows this in practice: it does not cover training aimed at reproducing the expression of particular works, and it may not apply where use would unreasonably prejudice rightsholders. Singapore’s Copyright Act 2021 has a computational data analysis exception that requires lawful access to the works.
In China, the Beijing Internet Court held in Li v. Liu (November 2023) that an image made with Stable Diffusion was protected because the user’s choice of prompts and parameters reflected real intellectual effort. China’s generative AI rules also require providers to use training data from lawful sources and respect intellectual property.
India’s Copyright Act has no specific TDM exception. In July 2026 the Delhi High Court refused ANI an interim injunction against OpenAI, finding on a preliminary view that training fell within fair dealing; the full case continues. Separately, a December 2025 government working paper proposed a mandatory licence for AI training, with royalties paid through a central collective once a model is commercialised. It is a proposal, not law.
Who owns AI outputs?
Most systems protect only human creativity. The US Copyright Office’s January 2025 report found that prompts alone generally do not make the user an author, while human selection, arrangement and modification of outputs can be protected. The courts agree: Thaler v. Perlmutter confirmed that a work made autonomously by AI cannot be registered, and the Supreme Court declined to hear the case in March 2026.
The EU test of originality, the author’s “own intellectual creation”, points the same way. The UK is an outlier: its 1988 Act protects “computer-generated works” with no human author for 50 years, though the scope of that rule for modern AI is untested. China’s Li v. Liu decision shows a more generous view of prompting.
The practical point for businesses: purely machine-made logos, code or copy may be hard to protect against copying by competitors. Keep records of human creative input where ownership matters.
Contracts do most of the practical work
For most companies, the live questions sit in contracts rather than courtrooms. Your model provider’s terms decide who owns outputs, whether your inputs can be used for training, and whether the provider will defend you if an output is said to infringe.
Several major providers now offer intellectual property indemnities for enterprise customers. These promises usually come with conditions, such as using the provider’s content filters and not deliberately prompting for protected material. Consumer and free tiers often offer much less.
- Output ownership and any licence back to the provider
- Use of your prompts, files and outputs for model training
- Scope, conditions and caps of any IP indemnity
- Warranties on training-data provenance, if any are offered
- Your obligations to your own customers, which should mirror what you receive
Checklist
- List every model you train, fine-tune or call, and the market where each is trained and deployed.
- Record the source and licence of each training and fine-tuning dataset, including scraped data.
- Configure crawlers to honour robots.txt and other machine-readable opt-outs.
- Remove pirated or unlawfully obtained copies from datasets and keep evidence of the clean-up.
- Test models for memorisation and add output filters for long verbatim reproduction.
- Prepare the EU training-content summary if you place a general-purpose model on the EU market.
- Review provider terms for output ownership, training use and indemnity conditions.
- Document human creative contribution where you need to own or register AI-assisted work.
FAQs
Is it legal to train AI on content that is publicly available online?
Public does not mean free to copy. In the EU it depends on whether the rightsholder opted out, in the UK commercial training generally needs a licence, and in the US it depends on a fair use analysis. Japan and Singapore are more permissive, within limits.
Do I own the images or text my team creates with AI tools?
As against your provider, the contract usually assigns outputs to you. As against the world, copyright may protect only the human creative contribution, so purely generated material can be hard to enforce against copiers.
Does the Bartz settlement mean AI companies must pay for training data?
No. The court found lawful training on purchased books to be fair use. The settlement related mainly to pirated copies. It shows how expensive poor data provenance can be.
If I only use a third-party model, am I exposed?
Less than the model developer, but not immune. If an output reproduces a protected work and you publish it, you may be liable. Check your indemnity and use output filters.
Which jurisdiction’s rules apply to my model?
Copyright is territorial. The rules of the place where copying happens apply to training, and the rules of each market where outputs are used apply to those outputs. Plan for every market where you train, deploy or could be sued.
Primary sources
US Copyright Office · Copyright and AIEUR-Lex · Copyright in the Digital Single Market Directive (EU) 2019/790European Commission · EU AI ActGeneral information, not legal advice.