A stock photographer found one of his pictures in a dataset of 5.85 billion images and sued. Two German courts have ruled against him, and the reason why matters to every image library with a terms-of-use page.
Robert Kneschke is a German stock photographer. LAION is a Hamburg non-profit that compiled LAION-5B, a dataset of 5.85 billion pairs of image links and text descriptions that has been used to train image generators. At least one of Kneschke's photographs was among them. He sued, and so far two courts have found against him. The most consequential part of the second ruling is a short point about what counts as saying no.
This article is general information, not legal advice. Checked on 18 November 2026.
What is the case about?
LAION did not store Kneschke's photograph in its dataset. The dataset holds web addresses and captions. To build it, however, LAION downloaded each image so that software could compare the picture with its accompanying text. That download is a reproduction, and reproduction needs either permission or an exception.
Germany has two relevant exceptions, both implementing the EU's 2019 Copyright in the Digital Single Market Directive:
- Section 60d of the German Copyright Act, implementing Article 3 of the Directive: text and data mining for scientific research. Rights holders cannot opt out.
- Section 44b, implementing Article 4: text and data mining by anyone for any purpose, unless the rights holder has reserved the use. For works available online, the reservation is effective only if it is machine-readable.
Kneschke's photograph had been published on a stock agency's website, whose terms contained a restriction on automated use. The case therefore put two questions: did either exception cover what LAION did, and if the general one did, had the rights been validly reserved? Our wider comparison of text and data mining exceptions in the UK, EU and US explains how these provisions differ from British and American law.
What did the Hamburg courts decide?
| Date | Court and reference | Outcome |
|---|---|---|
| 27 September 2024 | Hamburg Regional Court, 310 O 227/23 | Claim dismissed. LAION's download was covered by the scientific research exception, section 60d |
| 10 December 2025 | Hanseatic Higher Regional Court, 5 U 104/24 | Appeal dismissed. Covered by section 60d and also by the general exception, section 44b. Reservation not machine-readable |
| 3 September 2026 | Federal Court of Justice, I ZR 281/25 | Oral hearing held. No decision announced |
First instance. The Regional Court decided the case on the research exception. LAION is a non-commercial association, and the court treated creation of the dataset as scientific research.
Appeal. The Higher Regional Court upheld that and went further. According to Morgan Lewis's summary of the judgment, it held that:
- downloading and processing the photographs was a reproduction;
- the reproduction was permitted under section 44b, because automated analysis to extract information is data mining, and the information extracted did not have to be "hidden";
- section 60d also applied, because the non-profit's processing served research. The Hogan Lovells case tracker adds that, in the court's view, access to the dataset by commercial users did not change that;
- the photographer had not proved an effective opt-out.
Why was the opt-out held invalid?
This is the part that reaches beyond the parties. The reservation relied on was written in natural language. The appeal court held that this was not machine-readable. As Morgan Lewis reports the reasoning, the text had to be interpretable by a machine in an automated process, not merely capturable by one. A crawler can copy the words of a terms-of-use page; that is different from being able to act on them.
The Hogan Lovells tracker records a second problem: the wording did not expressly mention text and data mining.
Two cautions are needed before treating this as settled.
First, it is a decision of one German appeal court, and it is under appeal.
Second, the first-instance court had approached the point differently. It decided the case on the research exception, so what it said about the general exception was not the basis of its ruling. The appeal court's holding is the one that currently stands.
What happened at the Federal Court of Justice?
The First Civil Senate of the Federal Court of Justice, Germany's highest civil court, heard the case in Karlsruhe on 3 September 2026. According to the legal news service beck-aktuell, no date for a decision was announced and it was initially unclear when one would come. The senate indicated that it saw a need to clarify points of EU law, which means it could refer questions to the Court of Justice of the European Union in Luxembourg.
We found no report of a decision or a reference in the weeks since. If the court does refer, the proceedings in Germany would pause until Luxembourg answers, which commonly takes more than a year.
A related case is already before the EU court. In Like Company v Google Ireland (C-250/25), Advocate General Szpunar advised on 3 November 2026 that training a language model is as a rule permitted under Article 4, while recommending that the questions on training be treated as inadmissible. His opinion is not binding and does not address what makes a reservation machine-readable.
What counts as machine-readable today?
The Directive says only that a reservation must be made "in an appropriate manner, such as machine-readable means". It names no protocol. In practice several exist, and they work at different levels.
| Signal | Where it lives | What it covers | Recognition |
|---|---|---|---|
| robots.txt | A file at the root of a website | Crawling of that site, by named crawler | Named in the EU Code of Practice for general-purpose AI |
| TDMRep | HTTP header, a well-known file, or HTML meta tags | Text and data mining of a site or resource | W3C Community Group report, built around Article 4(3); not a formal W3C standard |
| IPTC "Data Mining" property | Metadata inside the image file | That image, wherever the file goes | In the IPTC Photo Metadata Standard since 4 October 2023 |
| Really Simple Licensing (RSL) | Machine-readable licence terms referenced from robots.txt | Terms on which content may be used | Version 1.0 released 10 December 2025; supported by publishers and infrastructure companies |
| IETF AI Preferences | Proposed vocabulary for robots.txt and headers | Categories such as AI training and search | Still in draft; no published standard that we could confirm |
The European Commission consulted between December 2025 and 23 January 2026 on which protocols AI providers must respect under the AI Act's Code of Practice. Seven were assessed, and CEPIC, the picture-industry association, reported strong disagreement among rights holders. We could not verify a final list. The broader compliance framework is described in our guide to the EU AI Act training data summary.
What can image libraries consider in the meantime?
The appeal ruling points one way while the higher courts deliberate. The points below are general observations drawn from it, not advice; what suits a particular library is a matter for its own advisers.
- Terms and conditions alone may not be enough. They still bind people who agree to them, but on the appeal ruling they work as a contract and not as an opt-out.
- Site-level signals. A robots.txt file that addresses AI crawlers by name is the one protocol the EU Code of Practice mentions. It covers the site it sits on and nothing else.
- A reservation in the file. Images are copied, re-hosted and scraped from places their owners do not control. That was the situation in Kneschke: the picture was on an agency's site, not the photographer's. Of the signals in the table above, the IPTC Data Mining property is the one designed to travel inside an image file. Our free AI training opt-out metadata tool writes it into a JPEG in the browser without uploading anything.
- What survives. Many platforms strip metadata on upload. The EXIF and metadata viewer shows whether a published copy still carries the property.
- The words "text and data mining". The appeal court noted that the wording before it did not name the use being reserved.
- Dated evidence. A record of when each signal was put in place helps to show what a crawler would have met on a given date.
No signal compels a crawler to behave. What a valid reservation does, in EU law, is remove the Article 4 exception, so that mining afterwards generally needs permission. It cannot be used against research organisations relying on Article 3.
What the case does not decide
- It is not about training. The claim concerned LAION's download while building a dataset. Whether a company that trains a commercial model on that dataset can rely on the exceptions is a different question.
- It is not about the UK. Britain has no equivalent of Article 4, as our article on UK copyright and AI after the 2026 report sets out.
- It is not final.
The case sits alongside the other pending disputes in our AI copyright cases tracker.
A different starting point
Kneschke's difficulty began with a photograph that was already online, on a third party's website, under terms a crawler could not interpret. Most of the Robert Harper Photograph Library is in the opposite position. Of 95,487 unique photographs on the archive's inventory count, 52,684 are analogue, 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. Much of it is still film, as the analogue collections describe.
That means decisions about metadata and reservations can be made as the material is scanned, not retrofitted afterwards. The archive is owned by 825 Ventures Ltd, the owner and rights holder; licences are negotiated directly and are subject to rights review. If you work on dataset procurement and would like to discuss licensed image data, see AI training data licensing.
Frequently asked questions
What was decided in Kneschke v LAION?
Two German courts held that LAION, a non-profit, did not infringe copyright when it downloaded a stock photograph to check it against its description while building the LAION-5B dataset. The Hamburg Regional Court relied on the scientific research exception in September 2024. The Higher Regional Court added the general text and data mining exception in December 2025. A further appeal is pending.
Is a notice in website terms and conditions a valid opt-out from AI training in the EU?
On the Hamburg appeal ruling of 10 December 2025, no. The court held that a reservation in ordinary language was not machine-readable, and noted that the wording did not expressly mention text and data mining. That ruling is under appeal to Germany's Federal Court of Justice, so the point is not finally settled.
What does machine-readable mean for an opt-out?
According to the Hamburg appeal court, the reservation has to be interpretable by a machine in an automated process, not merely capturable by one. A crawler should be able to detect and act on it without a human reading the page. Protocols such as robots.txt, TDMRep and the IPTC Data Mining metadata property are designed to work that way.
Has the German Federal Court of Justice ruled on Kneschke v LAION?
Not as far as we could establish on 18 November 2026. The First Civil Senate heard the case in Karlsruhe on 3 September 2026. Reports from the hearing say no date for a decision was announced and that the court saw a need to clarify EU law, which could lead to a reference to the Court of Justice in Luxembourg.
How can photographers opt out of AI training in a way machines can read?
Several signals exist and they can be combined. A robots.txt file addresses crawlers at website level. TDMRep can express a reservation through an HTTP header or a file on the site. The IPTC Data Mining property sits inside the image file itself, so it travels with copies, though it disappears if metadata is stripped. None is guaranteed to be honoured.
Sources
- Morgan Lewis: AI and copyright, judicial landscape in Germany (25 June 2026)
- Hogan Lovells AI litigation case-law tracker: Robert Kneschke v LAION e.V.
- beck-aktuell: Federal Court of Justice hearing on AI training and photography (3 September 2026)
- Recht im Internet: press statement on the LAION proceedings
- Directive (EU) 2019/790 (DSM Copyright Directive), EUR-Lex
- European Commission: the General-Purpose AI Code of Practice
- CEPIC: AI opt-out protocols, results of the EU consultation
- IPTC: Exclude images from generative AI with the Photo Metadata Standard 2023.1
- W3C: Text and Data Mining Reservation Protocol Community Group
- RSL Standard: RSL 1.0 specification announcement
- Court of Justice of the EU: Advocate General's opinion in C-250/25 Like Company v Google Ireland
- Lantad: The AI preferences standard you cannot deploy yet (IETF AIPREF status)
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.

