Two separate questions that keep getting mixed up
The input side: is anyone free to train a model on your content, and can you stop them?
The output side: is what you make with the machine protected, and is it yours?
The two questions have completely different answers, and they are settled in different places. Let us take them in order.
The input side: the text and data mining exception
The EU copyright directive gives two exceptions. One is for research organisations and cultural heritage institutions, for the purposes of scientific research. The other one — the general exception in Article 4 — is the one that concerns you: reproductions and extractions of lawfully accessible works for text and data mining are permitted — provided the rightholder has not reserved that use in an appropriate manner, and for content made publicly available online, "by machine-readable means".
That wording is the directive's. What actually binds you is your own member state's transposition of it, and those texts are not identical word for word. The three conditions are the same everywhere — lawful access, no valid reservation, copies kept only as long as necessary — but the phrasing differs, and with it the room for interpretation.
So wherever in the EU your site is hosted, the objection has to be in machine-readable form. That is not a technicality: it is the condition the whole exception turns on.
Hungary — Section 35/A asks for a machine-readable objection word for word
The Hungarian transposition is Section 35/A(1) of the copyright act: reproduction needed for text and data mining is free use where the work was accessed lawfully, where the rightholder has not objected in advance in an appropriate manner — in machine-readable form for content made publicly available by electronic means — and where copies are kept only as long as necessary.
That phrasing is the directive's, word for word. So for a Hungarian site, a visible sentence in plain language in the terms of use is not enough on the text of the act: for content published online, machine-readable form is the condition.
Spain — the act says "machine-readable or other appropriate means"
The Spanish transposition is not even in the copyright act but in a separate instrument, Article 67 of Real Decreto-ley 24/2021. It merged the directive's two exceptions into a single article, so mining for scientific research and the general exception both ended up in the same place.
The substantive difference is in the form of the reservation. Under Article 67(3) the exception does not apply where the rightholder has expressly reserved the use "medios de lectura mecánica u otros medios que resulten adecuados" — by machine-readable or other appropriate means. That is wider than the Hungarian or the German wording: under Spanish law a visible reservation in plain language is at least arguable.
It has a consequence that Spanish commentators treat as a drafting fault: because paragraph 3 applies to the whole article, a reservation also switches off mining for scientific research — while the directive makes precisely that exception one you cannot contract out of. The tension will surface in some future case.
And here is the problem: nobody knows exactly what counts as machine-readable.
The German case about precisely this — the one before the supreme court in September
A German photographer sued the non-profit organisation that assembled one of the largest public image-text datasets. Part of the question was whether the stock agency's reservation, written in natural language into its general terms and conditions, counts as machine-readable.
| Court | When | What happened |
|---|---|---|
| Hamburg Regional Court | September 2024 | dismissed the claim, but accepted the natural-language reservation as machine-readable |
| Hamburg Higher Regional Court | December 2025 | dismissed the appeal, but reversed the position on machine-readability |
| Federal Court of Justice | 3 September 2026, oral hearing | no decision; the panel signalled that a reference to the EU court may become necessary |
The first-instance reasoning was logical: it would be a contradiction in values if the exception allowed language-understanding models to be developed, while we did not expect those same models to be used to read the reservation. The appeal court rejected that, and named robots.txt and a dedicated protocol as the machine-readable forms.
The claimant's counter-argument, though, went unanswered, and it came up again at the supreme court: nobody could say what should have been written into robots.txt back in 2021 to keep that particular image out of the dataset — not least because the organisation did not collect the data itself, but took over a ready-made dataset.
What has to be written down here: today there is no safe answer to how you can validly forbid AI training on your own content. That is not vagueness in this article — that is the state of the law.
On the practical side of machine-readable signals — what you put into a robots.txt and what it is worth — the piece on Making Your Site Machine-Readable goes into detail.
The Munich judgments: where the training happened does not matter
Two German judgments came out of the same court, and both are significant.
The first, in November 2025, between a collecting society and a large chatbot provider. The court held that the provider should have acquired the rights to the song lyrics for training and for operation; the systems contain copies of the original works, which they reproduce and make available on a simple user instruction. The court also held that the provider is not a privileged research organisation, and that the text and data mining authorisation cannot in any case justify storing and outputting protected lyrics. The judgment is not final: the provider appealed.
The second, in July 2026, the same collecting society against a music generation service. The court found that by training its system on world-famous songs in the United States, and by storing and reproducing them in Europe, the service infringed both American and German copyright. According to the collecting society, this is the first time a European court has ruled on training carried out in the United States.
The business conclusion: on the Munich approach, what counts is not where the training happened but that the system is operated in Europe and the output appears here. A European company using the output of an American generative service faces European legal risk even if the model was trained in the United States.
The Hungarian case before the Court of Justice of the European Union
This is the most important pending European case on the subject — and it was referred by a Hungarian court.
| Item | Value |
|---|---|
| Case number | C-250/25 |
| Referring court | Budapest Környéki Törvényszék (Hungary) |
| Lodged | 3 April 2025 |
| Language of the case | Hungarian |
| Hearing | 10 March 2026 |
| Advocate General's opinion | none |
| Judgment | none |
The Hungarian court asked four questions, and together they cover the whole problem:
- Does it amount to communication to the public if a chatbot displays, in its answer, a stretch of text from a press publisher's website long enough to enjoy related-right protection — and does it matter that the chatbot produces the text by predicting words?
- Does the training itself amount to reproduction?
- If it does, is it covered by the text and data mining exception?
- Does it amount to reproduction by the provider if a chatbot, asked a question referring to a press publication, displays its content?
Why this is the most important case: it is the only forum whose decision will bind all 27 member states. And it asks about the input side and the output side at once.
Its legal basis is the press publishers' related right, created by Article 15 of the same copyright directive and transposed into every national act — into Sections 82/A to 82/C of the Hungarian copyright act, and into Article 129 bis of the Spanish one. The term of protection is short: two years, counted from the first day of January of the year following the publication of the press publication. And the right does not reach hyperlinking, the use of "individual words or very short extracts", or private, non-commercial use by individual users.
How much the national choice matters on this right is best shown by Spain: the stricter Spanish version of 2014 imposed an unwaivable payment on news aggregators, whereupon the largest aggregator shut down in Spain in December 2014 and came back only after the 2021 reform, in June 2022. Same right, different wording, ten years apart.
The contradiction sitting inside the legislation
Article 53 of the AI Act obliges providers of general-purpose models to put in place a copyright policy, and to identify and comply with reservations under Article 4 using "state-of-the-art technologies". They also have to publish a sufficiently detailed summary of the content used for training.
Meanwhile the German courts are arguing right now about whether a natural-language reservation has to be treated as machine-readable at all. So the legislation imposes an obligation whose content has not yet been decided. That is the most honest description of this area in September 2026.
The output side: is what you made with the machine yours?
EU law has no separate rule for AI-generated work. The condition for protection is that the work be the author's own intellectual creation, reflecting their free and creative choices.
The national acts say the same thing in their own words, and the Court of Justice fixed the standard in the line of cases starting with Infopaq: subject matter is protected if it is the author's own intellectual creation, and the copyright belongs to the person who created the work. On top of that, an "idea, principle, procedure or method of operation" cannot be the object of protection — and a prompt, read as an instruction, can easily fall through that sieve.
The first member state judgment on this is a refusal. In February 2026 a German local court refused authorship for three logos generated with a text-to-image model. The reasoning: the model has to work as a "mere tool", not as an "independent creative instrument"; open instructions — "make it more artistic", say — leave billions of possibilities to the machine; and picking one of four suggestions afterwards is not enough. The court did leave the door open for human intervention that leads to the prompter's personality being reflected in the output.
Beyond that one ruling I found no member state court decision and no official expert opinion on the authorship of AI-generated work. The national starting point is not in the same place everywhere, though — open the one that applies to you:
Hungary — a strict statutory text, and no applied practice at all
Under the Hungarian copyright act, protection attaches to a work "on the basis of the individual, original character arising from the author's intellectual activity", and the copyright belongs to the person who created the work. An "idea, principle, concept, procedure or method of operation" cannot be the object of protection.
I found no Hungarian court decision and no expert body opinion on the authorship of AI-generated work. The text exists, the application does not — anyone citing Hungarian precedent cannot name one.
Spain — there is already a concrete refusal on the record
Article 5(1) of the Spanish copyright act is even more categorical: "Se considera autor a la persona natural que crea alguna obra" — the author is the natural person who creates the work. Article 10(1) additionally requires an original creation.
And there is a concrete decision of the kind Hungary lacks: on 14 March 2023 the Madrid territorial copyright registry refused to register the rights in a work presented as "the first novel written by an artificial intelligence", applying Articles 5 and 10 together — and reasoning that it could not be established what exactly within the result was the protectable human creation.
In practice this lands in the same place as the Hungarian text, only on the record: the question is not whether the machine took part, it is whether you can separate out what in it is your own creation. If you cannot, you already fail at registration, not just in litigation.
But three routes do exist, and they follow from the law
1. The collection. EU law protects a collection — a database, in the directive's language — where the selection or arrangement of its contents is the author's own intellectual creation, and the next part is the key one: that copyright covers the collection itself and does not extend to its contents, which means the collection can be protected even where the individual items are not protectable at all. The rights in the collection as a whole belong to the person who compiled it.
In plain terms: the twenty generated images may not be protected on their own, but the way you selected and ordered them in a publication can be.
2. Substantive human reworking. An adaptation, arrangement or translation of someone else's material is itself protected if it has an original character of its own — the national acts all say this, and they all measure it against the same originality standard. If real human work is done on the generated material, that can create protection in its own right.
3. The trade mark. Trade mark protection does not depend on authorship. A logo generated with AI can be registered as a trade mark even if it enjoys no copyright protection. In practice that is often exactly enough — because the real business question is not "am I the author" but "can someone else use it".
On the image side of this question there is a separate piece: Who Owns an AI-Generated Image? takes the same problem from the picture end.
And one exception worth knowing about: pastiche
EU copyright law contains an exception for use for the purpose of caricature, parody or pastiche — Article 5(3)(k) of the 2001 Information Society Directive — within the limits justified by that purpose. Pastiche means imitation of a style; the exception is optional for member states, so whether you can rely on it, and in what wording, depends on your own national act.
This is the only express exception in EU copyright law that could serve as a defence for content "made by machine, in the style of X". And the subject is live: in September 2026 Germany's federal supreme court — independently of the training-data case above, in separate proceedings — widened the reading of the pastiche exception to cover sound sampling.
Hungary — pastiche is named in the act
Under Section 34/A(1)(b) of the Hungarian act, anyone may use a work "by way of evoking it and expressing humour or mockery, for the purpose of caricature, parody or style imitation (pastiche)", to the extent justified by that purpose. All three cases appear in the text, expressly.
Spain — only parody made it in, pastiche did not
Article 39 of the Spanish copyright act names parody only, and even that as a limit on the transformation right: the parody of a disclosed work is not treated as an adaptation requiring the author's consent, as long as it carries no risk of confusion. Caricature and pastiche do not appear in the article.
This does not mean style imitation is forbidden in Spain — it means there is no express exception to invoke. EU law does state separately that a user uploading to an online content-sharing platform can rely on caricature, parody and pastiche, but that rule is about platform use, not about use in general.
The United States — briefly, because it works differently
There is both administrative practice and case law there: a federal appellate court held in 2025 that under the Copyright Act every protectable work must be primarily of human origin, and the supreme court declined review in March 2026. The copyright office is producing a three-part report on the subject, the third part of which, on training, is still only available in a pre-publication version in September 2026.
The substantive difference in one sentence: the United States has a registration system, and there the machine-made part has to be excluded from the application. The European Union has no registration, no official guidance, and even the first member state judgment comes only from a local court.
What to take away
On the input side the exception exists and so does the right to opt out — we just do not know how to exercise it validly. What you can do today: put out a machine-readable signal, and know that its legal weight is not yet settled.
On the output side the purely machine-made work is probably not protected, but arrangement, substantive reworking and the trade mark are three working detours.
And the most important date: the Court of Justice of the European Union will rule on both sides, on a Hungarian reference. The case is pending, the hearing took place in March 2026, and there is no opinion and no judgment. This article is updated quarterly — that is the line to look at again next time.