Three legal systems, three answers to one question: may a machine copy a photograph in order to learn from it? A side-by-side guide for counsel and policy readers.
The United Kingdom lets machines mine copyright works for non-commercial research and nothing more. The European Union lets anyone mine for any purpose unless the owner has said no in a way a machine can read. The United States has no exception at all and leaves the matter to fair use. Those three sentences are the whole comparison; the rest is detail, and the detail is where disputes are being fought.
This article is general information, not legal advice. Checked on 18 November 2026.
Training a model on photographs involves copying them: downloading, storing, resizing, feeding them through a pipeline. Copying is the act copyright restricts. A text and data mining (TDM) exception is a statutory permission to make those copies for computational analysis without asking the owner.
The exceptions differ along four lines: who may rely on them, for what purpose, whether lawful access to the work is a precondition, and whether the owner can switch the exception off.
The three regimes side by side
| United Kingdom | European Union | United States | |
|---|---|---|---|
| Legal basis | CDPA 1988, s.29A | Directive (EU) 2019/790, Arts 3 and 4, as implemented nationally | No exception; fair use, Copyright Act s.107 |
| In force | 1 June 2014 | Transposition deadline 7 June 2021 | Fair use codified in the 1976 Act |
| Who may rely on it | Anyone with lawful access | Art. 3: research organisations and cultural heritage institutions. Art. 4: anyone | Anyone, as a defence |
| Purpose | Non-commercial research only | Art. 3: scientific research. Art. 4: any purpose | Any; weighed case by case |
| Lawful access needed | Yes | Yes | Not a formal condition, but the source of copies has weighed heavily |
| Can owners opt out | No, and contract terms to the contrary are unenforceable | Art. 3: no. Art. 4: yes, by express reservation | Not applicable |
| Commercial AI training | Outside the exception | Permitted under Art. 4 unless validly reserved | Depends on the four factors |
| Transparency duty | None in statute | AI Act Art. 53 training summary | No federal duty; California's AB 2013 in force since 1 January 2026 |
How narrow is the UK exception?
Section 29A is short. A person who has lawful access to a work may copy it to carry out a computational analysis of anything recorded in it, for the sole purpose of research for a non-commercial purpose, with sufficient acknowledgement unless that is impractical. Subsection (5) makes any contract term that tries to prevent this unenforceable. The copy cannot be passed on or used for anything else.
Two attempts to widen it have failed. In June 2022 the Intellectual Property Office proposed an exception for any purpose with no opt-out; it was abandoned in February 2023. In December 2024 the government consulted on a commercial exception with a rights-reservation opt-out and stated that as its preference; the report of 18 March 2026 says that is no longer the preferred way forward. Neither was enacted.
The result, on the plain reading of the statute, is that commercial training on UK copyright works will usually need permission where the copying happens in the UK. No UK court has yet ruled on training directly. The territorial qualifier matters, as the Getty litigation showed. Both points are developed in our article on UK copyright and AI after the 2026 report.
How do the EU's two exceptions work?
The 2019 Directive on Copyright in the Digital Single Market created two exceptions with different beneficiaries.
Article 3 covers reproductions made by research organisations and cultural heritage institutions for scientific research, on works to which they have lawful access. Rights holders cannot opt out.
Article 4 covers reproductions by anyone, for any purpose including commercial ones, of lawfully accessible works. It applies on condition that the rights holder has not expressly reserved the use "in an appropriate manner, such as machine-readable means in the case of content made publicly available online".
Because these are provisions of a directive, each member state has its own implementing law. In Germany they are sections 60d and 44b of the Copyright Act, and it is German courts that have produced most of the case law so far.
What counts as a valid EU opt-out?
This is the live question for image libraries. In Kneschke v LAION, a photographer argued that a reservation written into the terms of the website hosting his picture was enough. On 10 December 2025 the Hanseatic Higher Regional Court in Hamburg disagreed: a reservation in natural language was not machine-readable, because it has to be interpretable by a machine in an automated process and not merely capturable by one. The court also held that building the LAION dataset was covered by both German exceptions. The Federal Court of Justice heard the further appeal on 3 September 2026 and had not announced a decision or a decision date in the reports we found. The case is covered in full in our article on Kneschke v LAION and machine-readable opt-outs.
Other German rulings mark the exception's outer edges. On 11 November 2025 the Munich Regional Court held in GEMA v OpenAI that song lyrics memorised in a model are reproductions which the TDM exception does not cover; that decision is under appeal. On 31 July 2026 the same court held in GEMA v Suno that training on works obtained by stream-ripping fell outside the exception, in part because the works were not lawfully accessed; that judgment is not final.
At EU level, Advocate General Szpunar advised the Court of Justice on 3 November 2026, in Like Company v Google Ireland, that training a language model is as a rule permitted under Article 4. He also advised that the questions on training were hypothetical and inadmissible, so the remark was made for completeness. An opinion does not bind the Court, and the judgment is still to come.
The AI Act then layers a compliance duty on top. Providers of general-purpose models must have a policy to identify and respect Article 4(3) reservations, and must publish a summary of training content, as set out in our guide to the EU AI Act training data summary.
Why does the US have no exception at all?
American copyright law handles new technology through fair use, a defence that weighs four factors: the purpose and character of the use, the nature of the work, the amount taken, and the effect on the market for the original. Nothing in the statute mentions data mining.
That makes US law both more permissive and less predictable than a written exception. A use for any purpose can qualify, but nobody knows whether it does until a court says so. The decisions to date:
- Bartz v Anthropic (23 June 2025): training a language model on lawfully purchased books was fair use; retaining a library of pirated books was not.
- Kadrey v Meta (25 June 2025): summary judgment for Meta on training, on a limited record about market harm.
- Thomson Reuters v Ross (Third Circuit, 29 September 2026): copying Westlaw headnotes to train a non-generative legal search tool was not fair use. The court distinguished generative models.
No US court has ruled on training an image model. There is also no opt-out mechanism with legal force, because there is no exception to opt out of; a rights holder's stated refusal is at most one fact among many.
Which signals do rights holders actually have?
Across all three systems the tools are the same, though their legal weight differs.
| Signal | Works at the level of | Legal standing |
|---|---|---|
| robots.txt | Website | Named in the EU Code of Practice for general-purpose AI. No statutory status in the UK or US |
| TDMRep | Website or file, via HTTP header, well-known file or HTML | Designed around Article 4(3). A W3C Community Group report, not a formal standard |
| IPTC "Data Mining" property | Individual image file | Part of the IPTC Photo Metadata Standard since 4 October 2023. Disappears if metadata is stripped |
| Licence terms | Contract | Binding on the parties to the contract only |
Of the signals in the table, the IPTC property is the one built to travel inside an image file. Our free AI training opt-out metadata tool adds it to a JPEG in the browser, and the EXIF and metadata viewer shows whether a file already carries one.
What is decided and what is pending?
Settled: the text of section 29A; the text of Articles 3 and 4; the absence of any US exception; the application of AI Act Article 53 from 2 August 2025.
Pending: the Court of Appeal judgment in Getty v Stability AI, heard on 11 and 12 November 2026; the Federal Court of Justice's decision in Kneschke; the Court of Justice's judgment in Like Company; any request for rehearing in Ross; the first US ruling on an image model; and the list of opt-out protocols that EU providers must honour, which we could not verify as final.
Where a licence replaces the question
Every exception discussed here is an answer to the problem of using works without permission. A licence removes the problem for the works it covers. That is the practical case for dealing directly with an owner, and it is the footing on which Archive.London works: the Robert Harper Photograph Library, 95,487 unique photographs by its inventory count, is owned by 825 Ventures Ltd, the owner and rights holder, and more than half of the archive is still on film and has yet to be scanned, and only a small selection has been published online by Archive.London. Its range runs from studio still life to the street and travel pictures in the London collections.
Licences are negotiated directly and are subject to rights review; the general framework is on our image licensing terms page. Counsel and data teams comparing licensed sources can find the starting points at AI training data licensing.
Frequently asked questions
What is a text and data mining exception?
It is a rule in copyright law that lets someone copy protected works without permission in order to analyse them computationally, for example to find patterns or to train a model. The copies are a means to an end. Exceptions differ by country in who may rely on them, for what purpose, and whether rights holders can opt out.
Does the UK text and data mining exception cover commercial AI training?
No. Section 29A applies only where the copy is made for the sole purpose of research for a non-commercial purpose, by someone who already has lawful access to the work. Commercial model training falls outside it, so on the plain reading of the statute it will usually need permission from the copyright owner where the copying takes place in the UK. No UK court has yet ruled on the point.
How do rights holders opt out of text and data mining in the EU?
Under Article 4(3) of Directive 2019/790, by expressly reserving their rights in an appropriate manner, which for content published online means machine-readable means. robots.txt is the one protocol named in the EU's Code of Practice for general-purpose AI. The opt-out is not available against scientific research covered by Article 3.
Is there a text and data mining exception in the United States?
No. US copyright law has no specific exception. Developers rely on fair use, a flexible defence decided on four statutory factors. Courts have reached different results on different facts, and as of 18 November 2026 none has ruled on fair use for training an image model.
Sources
- Copyright, Designs and Patents Act 1988, section 29A
- Directive (EU) 2019/790 (DSM Copyright Directive), EUR-Lex
- GOV.UK: Report on Copyright and Artificial Intelligence (18 March 2026)
- Reed Smith: UK copyright and AI training, where the law stands after the opt-out rejection
- Morgan Lewis: AI and copyright, judicial landscape in Germany (25 June 2026)
- Court of Justice of the EU: Advocate General's opinion in C-250/25 Like Company v Google Ireland
- Regulation (EU) 2024/1689 (AI Act), EUR-Lex
- Haynes Boone: Third Circuit rejects fair use for Ross's copying of Westlaw headnotes
- Simmons & Simmons: GEMA v Suno AI, key EU decision on AI training, outputs and copyright
- IPTC: Exclude images from generative AI with the Photo Metadata Standard 2023.1
- Polsinelli: AI vs authors update (Bartz v Anthropic and Kadrey v Meta)
- Cooley: State AI laws, where are they now (24 April 2026)
This article is general information, not legal advice. It reflects the position as we understood it on the date shown and may be out of date. Published by Archive.London. Corrections are welcome under our editorial policy.

