OLG Hamburg upheld LAION's AI-training scrape because the photographer's natural-language opt-out was never shown to be machine-readable

Q: Did OLG Hamburg rule that natural-language AI opt-outs are invalid?

No. The 5th Civil Senate held the machine-readability requirement in Section 44b(3) UrhG is technology-open. It did not bar natural-language reservations as a category. On the facts, the photographer failed to prove his natural-language reservation was machine-readable at the relevant time in the second half of 2021, so it did not block the scrape.

Q: Is the decision final?

No. The Senate admitted a further appeal (Revision), so the Federal Court of Justice (Bundesgerichtshof) could still review it. The judgment expressly states it is not yet final (nicht rechtskraeftig).

By Anthony Guerriero, Founder and AI Regulation Analyst, The Leveraged Years. Hamburg, 10 December 2025. Last verified: 2026-07-25.

Germany just produced the first appellate judgment on whether scraping copyrighted images to build an AI training set is lawful, and how a rights holder can stop it. The Hanseatic Higher Regional Court (Hanseatisches Oberlandesgericht, OLG Hamburg) dismissed a professional photographer's appeal and let stand the reproduction of his photo by LAION e.V., the non-profit that assembles the image-text datasets behind many generative models. The wedge that makes this case worth citing is narrow and specific: the opt-out was written in plain German, sat in a stock agency's terms of use, and still failed, because the plaintiff could not prove a machine could read and act on it in 2021.

The ruling

On 10 December 2025 the 5th Civil Senate of OLG Hamburg decided case 5 U 104/24, dismissing photographer Robert Kneschke's appeal against the Regional Court of Hamburg (Landgericht Hamburg), which had rejected his claim on 27 September 2024 in case 310 O 227/23. Kneschke sued for an injunction against LAION's reproduction of one of his photographs during the creation of a dataset that pairs image URLs with text descriptions. LAION downloaded the image from a stock agency's site to match the picture against its caption, one step in building the training corpus.

The Senate held the Regional Court was right to dismiss the claim, but reached the result on grounds the lower court had left open. LAION could rely on the text and data mining exception in Section 44b of the German Copyright Act (Urheberrechtsgesetz, UrhG), the domestic transposition of Article 4 of the EU DSM Directive. As a second and independent basis, the reproduction was also covered by the scientific research exception in Section 60d UrhG, the transposition of Article 3. Building the dataset, the court said, was a methodical, verifiable process aimed at later insight, and the fact that commercial actors could also use the dataset did not defeat the research privilege because no private company exercised determining influence over the research body (Section 60d(2) sentence 3 UrhG).

What the court actually held on the natural-language opt-out

Article 4 of the DSM Directive lets rights holders switch off the commercial TDM exception by reserving their rights "in an appropriate manner, such as machine-readable means in the case of content made publicly available online." Section 44b(3) sentence 2 UrhG copies that: a reservation is effective only if made in machine-readable form. Kneschke's reservation existed. It was written in natural language, both in the agency's website terms and human-readably in the page source. The whole case turned on whether that counted.

The Senate refused to treat natural language as automatically disqualifying. It read the statute as technology-open (technikoffen): the legislature prescribed no fixed format, only that whatever form is chosen must be machine-readable, and the reservation may even sit in an imprint or in general terms and conditions so long as it is machine-readable there. But machine-readable means more than machine-capturable. The text must be capable of being machine-interpreted, so that an automated process actually excludes the reserved content. That raised the bar, because Kneschke's reservation did not name text and data mining outright; a machine would have to read the clause and infer that this use was covered.

Then came the fact that sank the claim. The relevant moment was the download in the second half of 2021, and Kneschke offered no proof the reservation was machine-readable then. His ChatGPT demonstration dated from 2023; the chatbot did not launch until late 2022. His reference to the "weboptout" tool came with no showing it existed in 2021. His expert-evidence offer spoke to today's capabilities, not to 2021. His late argument that a skilled developer could have built such a tool in three hours came in a pleading filed after the oral hearing and was procedurally barred. So the reservation failed, not because plain language can never work, but because its machine-readability in 2021 was never established. The court expressly left open whether a user must build custom software to read an opt-out.

Primary-source excerpt

From the court's official press release on the effect of the reservation:

Der auf der Webseite der Bildagentur zum Zeitpunkt des Downloads der Fotografie vorhandene Nutzungsvorbehalt habe vorliegend aber nicht die gesetzlich vorgesehene Form (Maschinenlesbarkeit) aufgewiesen (Paragraf 44b Abs. 3 S. 2 UrhG), so dass die streitgegenstaendliche Vervielfaeltigung zulaessig gewesen sei.

English gloss: the reservation of use present on the stock agency's website at the time the photograph was downloaded did not have the legally required form (machine-readability) under Section 44b(3) sentence 2 UrhG, so the reproduction at issue was permissible.

And from the judgment itself, the standard the Senate set for machine-readability:

Bei der Frage der Maschinenlesbarkeit kommt es nicht nur darauf an, dass der Text maschinell erfasst werden kann, sondern dass er auch in dem Sinn maschinell interpretiert werden kann, dass er bei einem automatisierten Vorgehen dazu fuehren kann, dass die vom Vorbehalt erfassten Inhalte nicht ausgewertet werden.

English gloss: on the question of machine-readability, what matters is not only that the text can be captured by a machine, but that it can be machine-interpreted so that, in an automated process, the content covered by the reservation is not analyzed.

How the EU opt-out regime compares with the US posture

EU DSM Article 4 opt-out (as read in OLG Hamburg 5 U 104/24) vs the US fair-use posture
Feature	EU: DSM Art. 4 / Section 44b UrhG	US: fair use (17 U.S.C. 107)
Default for commercial AI training	Permitted, unless the rights holder reserves rights	No statutory permission; case-by-case fair-use defense
How a rights holder blocks it	Opt-out, effective only in machine-readable form	No opt-out mechanism; owner sues and litigates fair use
Who carries the burden	Rights holder must show a valid, machine-readable reservation	Defendant must prove the use is fair
Test applied	Statutory exception plus the three-step test (Art. 7(2) DSM, Art. 5(5) InfoSoc)	Four-factor fair-use analysis, including market harm
Research carve-out	Separate, opt-out-proof research exception (Art. 3 / Section 60d UrhG)	Research weighs inside the fair-use factors, not a stand-alone shield

Cross-border read for US AI training

US courts frame AI-training disputes as fair use, where the fight is over transformative purpose and market harm and there is no formal way for an owner to pre-empt the analysis. The EU regime is structurally different: commercial mining is allowed by default and the owner's lever is the opt-out. OLG Hamburg's contribution is to define what that lever has to do. A reservation is not effective just because a human can read it; a machine has to be able to interpret it and act on it at the moment of the use.

For US and global model builders scraping EU-hosted content, that has two practical consequences. First, timing is a defense. Liability is judged against the technical state of the art when the copying happened, so reservations that only became machine-parseable later may not reach back. Second, the ambiguity cuts both ways: because the Senate insisted the standard is technology-open and left open whether scrapers must build their own readers, a well-drafted, explicitly TDM-tagged reservation using tools like robots.txt or the TDM Reservation Protocol is far more likely to bind than vague plain-language terms. The Federal Court of Justice may yet reset the line, since the Senate admitted a Revision.

Key Facts

Instrument: Judgment, OLG Hamburg (Hanseatisches Oberlandesgericht), 5th Civil Senate, case 5 U 104/24 (source: court press release)
Date decided: 10 December 2025, upholding LG Hamburg 310 O 227/23 (27 September 2024)
Parties: Photographer Robert Kneschke (appellant) v. LAION e.V. (respondent)
Holding: Scrape lawful under Section 44b UrhG (Art. 4 DSM TDM exception) and independently under Section 60d UrhG (Art. 3 research exception); natural-language opt-out not shown machine-readable in 2021
Status: Not final; further appeal (Revision) to the Bundesgerichtshof admitted (source: full judgment, Landesrecht Hamburg)

What this ruling does NOT do

It does not hold that natural-language opt-outs are invalid. The Senate said the machine-readability requirement is technology-open and declined to bar plain-language reservations as a class.
It does not settle the law. The decision is not final; the Bundesgerichtshof could still overturn or refine it on the admitted Revision.
It does not decide whether a scraper must build custom software to read an opt-out. The court expressly left that question open.
It does not rule on the legality of the earlier crawling that produced the source dataset, only on LAION's own download of the image.
It does not create a general AI-training license. Both exceptions carry conditions, and the three-step test still applies.

Frequently asked questions

Did OLG Hamburg rule that natural-language AI opt-outs are invalid?

No. The court held the standard in Section 44b(3) UrhG is technology-open and did not reject plain-language reservations as a category. Kneschke lost because he did not prove his reservation was machine-readable in the second half of 2021, the time of the download.

Which law did the court apply?

Section 44b UrhG, the German transposition of Article 4 of the DSM Directive, and, as a separate ground, Section 60d UrhG, which transposes the Article 3 scientific research exception.

Is the decision final?

No. The Senate admitted a Revision, so the Bundesgerichtshof may still review it. The judgment states it is not yet final.

Does this ruling apply outside Germany?

It binds only the parties, but it interprets Article 4 of the DSM Directive, which is EU-wide, so its reading of machine-readability is a reference point across the bloc.