The EU's Text-and-Data-Mining Opt-Out Under Article 4 CDSM

Directive (EU) 2019/790 — the CDSM Directive — contains two text-and-data-mining exceptions. One cannot be opted out of and one can. Almost every dispute about AI training data in Europe turns on the second, and specifically on eleven words in Article 4(3) about machine-readable means.

Two exceptions, not one

Article 3 permits reproductions and extractions for text and data mining by research organisations and cultural heritage institutions, for the purposes of scientific research, of works to which they have lawful access. Article 7(1) makes this exception mandatory in the sense that contract terms trying to override it are unenforceable, and there is no opt-out for rightsholders. It is narrow because the beneficiaries are narrow.

Article 4 is the general one. It permits reproductions and extractions of lawfully accessible works for text and data mining by anybody, including commercially. Copies may be retained for as long as necessary for the mining. And then Article 4(3) attaches the condition that makes it different from Article 3: the exception applies only where the use has not been expressly reserved by the rightsholder in an appropriate manner, such as by machine-readable means in the case of content made publicly available online. The Directive’s text is on EUR-Lex, and Member States were required to transpose it by 7 June 2021 — so the operative text for any dispute is the national implementing law, not the Directive itself.

Two conditions therefore have to be satisfied before Article 4 helps you at all: lawful access, and no reservation. Lawful access is doing more work than it is usually given credit for — content behind a paywall you did not pay for, or obtained in breach of terms that governed the access itself, is a problem before any question about reservation arises.

What a reservation has to be

The reservation must be express, it must be by the rightsholder, and it must be made in an appropriate manner. Three practical consequences follow.

• Silence is not a reservation. The default under Article 4 is that the exception applies. A rightsholder who does nothing has opted in, which is the exact inversion of the position under Article 3 and the reason the provision is described as opt-out.
• It has to come from the rightsholder. A platform hosting user content can reserve rights in its own material; whether it can do so for its users’ works depends on what its terms actually assign or license, and that is frequently less than the platform behaves as though it has.
• “Appropriate manner” depends on the channel.Recital 18 distinguishes content made publicly available online, where machine-readable means are indicated, from other contexts where contractual terms or a unilateral declaration may be appropriate. A notice in a print contract can be an appropriate reservation for material that is not published online.

What counts as machine-readable

The Directive does not define it, and this is the central unresolved question in European TDM law. The candidate mechanisms in circulation are all partial.

# robots.txt — the oldest mechanism, designed for crawling, not mining
User-agent: SomeAICrawler
Disallow: /

# An HTTP response header, as proposed by the TDM Reservation Protocol
# community work: a machine-readable pointer to a rights policy
tdm-reservation: 1
tdm-policy: https://example.com/tdm-policy.json

robots.txt is universally understood and expresses a rule about crawler access rather than about downstream use, so a crawler that was permitted to fetch and a model that was trained on the result are not obviously the same permission. Response headers and rights-metadata schemes such as the TDM Reservation Protocol express the reservation directly but are only effective if the party mining looks for them. Files like ai.txt and similar conventions are proposals, not standards. Content-provenance metadata handles a different problem — see what C2PA does and does not carry.

The most-cited judicial treatment is the Hamburg Regional Court’s judgment in the LAION case of 27 September 2024, which decided the matter on the Article 3 research exception. Its remarks on Article 4 were therefore not necessary to the decision: the court indicated that a reservation stated in natural language in a website’s terms could be capable of being machine-readable, given that machines can now read natural language. That reasoning is contested, it comes from a first-instance German court, and it binds nobody else. It is not settled whether a reservation in prose terms of use is effective across the Union, and the question will not be settled until either the Court of Justice rules on it or the point is legislated. Anyone telling you the answer with confidence is describing a preference.

What actually happens if you opt out

An effective reservation removes the Article 4 exception for that material. The consequence is not a penalty and not a new right: it is that the ordinary rules apply again, so reproducing the work for mining requires a licence or another exception. The rightsholder’s remedy is the ordinary copyright remedy in the relevant Member State.

Which is why the practical difficulty is not legal but evidential. A rightsholder who reserved rights and believes a model was trained on their work has to show that reproduction happened, that it happened within a jurisdiction where their rights bite, and that their reservation was in place and effective at the time. Training is not observable from the outside, and this evidential gap has shaped litigation elsewhere — it is precisely what caused the training claims to fall away in the UK Getty proceedings, on different law but the same underlying problem.

There is also a timing question with no clean answer. A reservation added today says nothing about a copy made two years ago, and models trained before a reservation appeared are not retrospectively affected by it. Whether continued distribution of such a model raises separate questions is not resolved.

How the AI Act picked this up

Article 53(1)(c) of the AI Act requires providers of general-purpose AI models to put in place a policy to comply with Union copyright law, and in particular to identify and comply with reservations of rights expressed under Article 4(3) of the CDSM Directive, including through state-of-the-art technologies. This is the provision that converts a rightsholder’s opt-out from something enforced only by private litigation into something a regulator can ask a model provider about.

Recital 106 goes further in reach: a provider placing a general-purpose model on the Union market should comply with that obligation regardless of the jurisdiction in which the training took place. The legal basis and limits of that extraterritorial framing are debated, and it has not been tested. The related transparency duty — the sufficiently detailed summary of training content under Article 53(1)(d) — is covered in the training content summary, and the policy obligation itself in the GPAI copyright policy requirement. The Code of Practice for general-purpose AI published in 2025 contains a copyright chapter with commitments on crawler behaviour and respecting robots.txt; adherence to it is voluntary and it is not itself law.