TRW & Co · Intellectual property and AI
Copyright and AI: Training Data and Infringement Risk
The copyright AI UK position is unsettled but not empty. Training on protected works in the UK still needs permission unless the narrow non-commercial research exception applies. The government has dropped its opt-out proposal without replacing it, and the leading case, Getty Images v Stability AI, is under appeal. This guide explains the law, the policy and the practical steps to take now.
Key pointGetty Images v Stability AI turned on where training happened: the UK training claim fell away because training took place abroad, so the location of compute and data is now a legal risk decision, not just an engineering one.
At a glance
Narrow TDM exception
Section 29A of the Copyright, Designs and Patents Act 1988 only permits copying for computational analysis for non-commercial research by people with lawful access.
Opt-out plan dropped
The government's March 2026 report confirmed that a broad exception with a rightsholder opt-out is no longer its preferred approach, without proposing a replacement.
Getty v Stability
The High Court held in November 2025 that a model whose weights store no training images is not an infringing copy; Getty has permission to appeal.
No transparency duty
UK law does not currently require AI developers to disclose training data, so contracts are the main tool for rightsholders and deployers.
Computer-generated works
The government proposes removing copyright protection for computer-generated works with no human author under section 9(3) of the 1988 Act.
Swipe or scroll →
Copyright AI UK: the short answer
The copyright AI UK position in October 2026 is that the law has not changed. Copying protected works to train an AI model in the UK still needs the rightsholder's permission unless an exception applies, and the only text and data mining exception covers non-commercial research. The government has dropped its preferred plan for a broad commercial exception with an opt-out, but has not replaced it with anything.
The courts have not filled the gap either. In Getty Images v Stability AI, the High Court decided in November 2025 that an AI model trained abroad was not an "infringing copy" because it did not store the training images, but the core questions about training were never decided, and Getty has permission to appeal. For now, licensing, contracts and evidence of where training happens matter more than any statute.
The legal framework: the CDPA and the narrow TDM exception
UK copyright is governed by the Copyright, Designs and Patents Act 1988 (CDPA). Copying a work, or a substantial part of it, is a restricted act, and copying includes storing the work in any medium by electronic means. Building a training dataset usually involves making copies, so training in the UK on protected works without a licence is likely to be a primary infringement unless an exception applies. Databases can also attract a separate database right.
The relevant exception is section 29A of the CDPA, in force since 2014. It allows a person with lawful access to a work to copy it for computational analysis for the sole purpose of non-commercial research, with sufficient acknowledgement. Copies cannot be passed on or used for other purposes, and a contract term that tries to prevent such copying is unenforceable. Commercial model training falls outside it.
The UK did not implement the EU's 2019 Digital Single Market Directive, so it has no equivalent of the EU's broader commercial text and data mining exception with a rights reservation. That difference sits at the heart of the copyright AI UK debate.
Secondary infringement
The CDPA also makes it an infringement to import, possess or deal with an "infringing copy" in the course of business, knowing or having reason to believe it is one. Section 27 treats an imported article as an infringing copy if making it in the UK would have infringed. This was the route Getty used to try to reach a model trained outside the UK.
Outputs and authorship
Separate questions arise for AI outputs: whether an output reproduces a substantial part of a protected work, and who, if anyone, owns copyright in AI-generated material. Section 9(3) of the CDPA gives protection to "computer-generated works" with no human author, but the government now proposes to remove it. Our article on AI-generated works ownership looks at that question in detail.
Getty Images v Stability AI: what was decided
In Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch), judgment was handed down on 4 November 2025 by Mrs Justice Joanna Smith. Getty had alleged that Stability used its images to train the Stable Diffusion model, that outputs reproduced its works, and that outputs carried its watermarks.
The case narrowed sharply. Getty abandoned its primary copyright claims before closing submissions: the training claim fell away because it was accepted that training did not take place in the UK, and the output claim became academic after Stability blocked the relevant prompts. The database right claims were also dropped.
On secondary infringement, the judge accepted that an "article" can be intangible, but held that Stable Diffusion was not an "infringing copy" because the model's weights do not store or contain copies of the training images. Getty succeeded only on a limited trade mark claim about watermarks appearing on images generated by some older versions of the model. Our article on trade mark infringement explains those rights.
The judge later granted Getty permission to appeal the secondary infringement ruling, describing the meaning of "infringing copy" for an AI model as a novel and important point of law. Stability was refused permission by the judge to appeal the trade mark findings. As of October 2026, the appeal is pending in the Court of Appeal, so the High Court ruling is the current law but may not be the last word.
Getty shows that the hardest part of an AI copyright claim is often not the law but the evidence: proving where training happened and what the model actually holds.
Government policy: the consultation and the March 2026 report
The government's Copyright and Artificial Intelligence consultation ran from 17 December 2024 to 25 February 2025. It offered four options: do nothing; require licences in all cases; a broad text and data mining exception; or an exception allowing rightsholders to opt out, combined with transparency measures. The government initially said it preferred the opt-out model.
During the passage of the Data (Use and Access) Act 2025, Parliament added sections 135 to 137, requiring the government to publish an economic impact assessment and a report on the use of copyright works in AI development within nine months of Royal Assent. Our guide to the Data (Use and Access) Act covers the wider Act.
The report and impact assessment were published on 18 March 2026. The report records that the originally preferred option, a broad exception with an opt-out, was rejected by most respondents, and says the government will consider other policy approaches and keep monitoring technology, litigation, international approaches and the licensing market. It proposes not to intervene in the licensing market at this stage, to consider the case for regulatory oversight of transparency only if legislation is introduced, and, absent evidence of its ongoing value, to remove protection for wholly computer-generated works while continuing to protect AI-assisted works.
For anyone advising on copyright AI UK questions, the significance is what the report does not do. It sets no timetable, proposes no Bill and leaves transparency, more targeted exceptions and digital replicas for further work. The legal risk therefore sits where it did before the consultation began, and businesses should plan on the basis that existing law will be applied by the courts case by case.
| Issue | Current position | Practical effect |
|---|---|---|
| Training on protected works in the UK | Needs permission unless an exception applies | Licence datasets or train outside the UK under the local law there |
| Commercial TDM exception | None; the opt-out proposal has been dropped | No statutory safe harbour for commercial training |
| Non-commercial research | Section 29A exception applies | Limited to computational analysis for non-commercial research with lawful access |
| Model trained abroad and used in the UK | Not an infringing copy on the High Court's reasoning in Getty | Under appeal; reassess when the Court of Appeal rules |
| Infringing outputs | Ordinary infringement tests apply | Output filters, prompt controls and indemnities matter |
| Computer-generated works (section 9(3)) | Still in force; government proposes removal | Do not rely on copyright in purely AI-generated material |
| Transparency of training data | No statutory duty in the UK | Contractual disclosure is the main tool |
Licensing and contract strategy
With no legislative fix, the copyright AI UK market is being shaped by contracts. The government's report notes a growing market for licensing content for AI training. For developers, a licence gives certainty about the source and scope of training data; for rightsholders, it turns a contested use into revenue.
Whatever side you are on, the contract terms do the work. Our guide to AI contracts explains these in more detail, but the core points are consistent.
- Map the data. List each source of training or fine-tuning data, how it was obtained, the applicable licence terms and where the copying took place.
- Define the permitted use. Specify training, fine-tuning, retrieval or output uses, the models covered, territory and duration; avoid vague 'AI purposes' language.
- Allocate output risk. Agree who is responsible if outputs reproduce protected material, including filters, takedown processes and indemnities with realistic caps.
- Agree transparency. Require disclosure of data sources at the level of detail each party needs, given that UK law does not yet require it.
- Protect confidential inputs. Restrict whether a deployer's prompts and documents can be used to train the provider's models.
- Plan for change. Include review clauses for the Getty appeal, any future UK exception and the EU AI Act, and specify what happens if the law moves.
Businesses that merely use AI tools are not immune. If an employee uses a generative tool to produce marketing copy or images that reproduce a protected work, the business publishing it can be the infringer. Put an internal AI use policy in place, check vendor terms for output indemnities, and keep a human review step for anything published.
An illustrative scenario
This scenario is illustrative only and does not describe any real client or matter.
A London publisher learns that a UK start-up is offering a writing assistant fine-tuned on a dataset that appears to include the publisher's back catalogue. The start-up says its base model was trained in the United States and that only light fine-tuning happened in London.
The publisher's copyright AI UK analysis turns on facts it does not yet have. Any copying of its books for fine-tuning in London would need a licence, because section 29A does not cover commercial work. Base-model training abroad is governed by foreign law, and on the High Court's reasoning in Getty the model itself is not an infringing copy in the UK. If outputs reproduce substantial passages, that is a separate infringement.
The publisher's practical options are a pre-action letter seeking disclosure of the fine-tuning data, a licensing proposal, or both. The start-up, for its part, should review its data records, its output filters and its own investors' expectations; a licence may cost less than a dispute and protect its fundraising.
Risks and common mistakes
Assuming publicly available means free to use
Content that is freely accessible online is still protected by copyright, and website terms often prohibit scraping. Section 29A only covers non-commercial research, and the government has not introduced a commercial exception. A business that scrapes web content to train or fine-tune a commercial model in the UK should assume it needs permission, and should keep records of what it collected, from where and on what terms.
Treating Getty as a green light
Getty decided a narrow point on secondary infringement for a model trained outside the UK whose weights did not store the training images. It did not decide that training in the UK is lawful, and the primary infringement claims were abandoned rather than rejected. The ruling is also under appeal. Developers should not restructure on the assumption that the High Court's reasoning will survive, and any fine-tuning or retrieval done in the UK needs its own analysis.
Ignoring the EU AI Act
UK businesses offering general-purpose AI models in the EU market can be caught by the EU AI Act, which requires providers of such models to have a copyright policy and to publish a summary of the content used for training. Those obligations sit alongside UK law and may be stricter in practice. See our guide to the EU AI Act for UK businesses and our overview of AI regulation in the UK.
Claiming ownership of purely AI-generated output
Many businesses assume they own whatever their AI tools produce. Where there is real human creative input, copyright may arise in the ordinary way. For outputs with no human author, the business may be relying on section 9(3), which the government now proposes to abolish. Brands, publishers and agencies should record the human contribution to important works and secure assignments from staff and freelancers.
Cross-border angle: where training happens matters
Copyright is territorial. Training that takes place in a country with a broad text and data mining exception, or a fair use doctrine, may be lawful there even if the same activity would infringe in the UK. That is why Getty's UK training claim fell away. For multinational developers, the location of compute and data storage is now a legal decision as well as a technical one.
For rightsholders, the territorial point cuts the other way: a UK claim may need evidence of copying in the UK, or of infringing outputs made available here. Groups with creative assets in South Asia, including Bangladesh, should register and document ownership in each key market and align licences with where the AI developer actually operates. Our media and entertainment team advises on content licensing across jurisdictions.
What to do next: checklist
- Developers: document the source, licence and location of every dataset; licence content used for training in the UK.
- Deployers: adopt an AI use policy, check vendor indemnities and review outputs before publication.
- Rightsholders: record ownership, set clear website and API terms, consider technical controls, and decide your licensing position.
- Everyone: build review clauses into contracts and monitor the Getty appeal and any further government proposals.
We advise developers, deployers and rightsholders on copyright AI UK issues, from dataset audits and licences to disputes; our IP litigation team acts where agreement fails. Contact our London office at Holborn Gate, 330 High Holborn, London WC1V 7QH, by email at info@trw.co or on +880 1779 127165.
Frequently asked questions
01Is it legal to train AI on copyrighted material in the UK?
Copying protected works to train an AI model in the UK generally needs the rightsholder's permission unless an exception applies. The only text and data mining exception, in section 29A of the Copyright, Designs and Patents Act 1988, is limited to computational analysis for non-commercial research by someone with lawful access. Commercial training is outside it. The government considered a broader exception with an opt-out but, in its March 2026 report, said that is no longer its preferred approach. Training outside the UK is governed by local law.
02What did the court decide in Getty Images v Stability AI?
In November 2025 the High Court dismissed Getty's secondary copyright infringement claim, holding that Stable Diffusion was not an infringing copy because its weights do not store or contain the training images. Getty had abandoned its primary infringement claims, partly because training did not take place in the UK. Getty succeeded only on a limited trade mark claim about watermarks on some generated images. Getty has permission to appeal the secondary infringement ruling, so the position may change.
03Has the UK introduced a text and data mining exception for AI?
No. The government consulted from December 2024 to February 2025 on four options, including a commercial exception with a rightsholder opt-out, which it initially preferred. Its report and impact assessment, published in March 2026 under the Data (Use and Access) Act 2025, said that approach is no longer preferred and that it will gather further evidence and consider alternatives, such as more targeted exceptions. No legislation has been introduced, so the existing non-commercial research exception remains the only one.
04Who owns content created by AI in the UK?
Where a person makes a real creative contribution using an AI tool, copyright may belong to that person or their employer in the ordinary way. For works with no human author, section 9(3) of the 1988 Act currently treats the author as the person who made the arrangements necessary for the work's creation, but the government now proposes to remove that protection. Businesses should record human input into important works and use contracts to allocate ownership with staff, freelancers and AI vendors.
05Can my business be liable for using AI-generated content?
Yes. If an output reproduces a substantial part of a protected work, publishing or commercially using it can infringe copyright, whoever generated it. Trade mark issues can also arise, as Getty's watermark claim showed. Reduce the risk with an internal AI use policy, human review of anything published, vendor terms that include output indemnities, and records of prompts and sources for important content. Check whether your insurance covers intellectual property claims arising from AI use.
06Do UK AI developers need to disclose their training data?
Not under current UK law. The government's March 2026 report proposed working with industry on best practice for transparency rather than legislating, and did not create a regulator for it. Developers placing general-purpose AI models on the EU market are, however, subject to EU AI Act obligations that include publishing a summary of training content. In the UK, rightsholders and business customers rely mainly on contractual disclosure obligations and, in a dispute, the court's disclosure process.
Sources and official materials
- Copyright, Designs and Patents Act 1988 — UK copyright statute
- CDPA 1988, section 29A — Text and data analysis for non-commercial research
- Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch) — High Court judgment on AI training and infringing copies
- Copyright and Artificial Intelligence consultation — Government consultation, December 2024 to February 2025
- Report and impact assessment on copyright and AI — Government report published in March 2026
- Data (Use and Access) Act 2025, section 135 — Duty to publish the economic impact assessment
General information only, current as at 21 October 2026. It is not legal advice and does not create a solicitor–client relationship. It reflects the law of England and Wales unless stated otherwise; Scotland and Northern Ireland have separate legal systems, and the law changes frequently. Check current official sources and take advice on the facts of your matter.