What EU law asks of anyone who mines the web for AI, and how to prepare your answer.

EU law allows text and data mining of content you can reach without passing a barrier, unless the rightsholder has reserved its rights.

The AI Act then asks providers of general-purpose models to respect those reservations and to say in public how they did so. The rule sounds simple and is hard to apply, because member states and courts read the reservation differently and the highest courts have not yet spoken. This page explains the position in plain language and shows how RightSignal helps you follow reservations and keep the evidence. It describes RightSignal's understanding on 10 November 2026 and is not legal advice.

Request accessTalk to us

Article 4 and the right to opt out

Article 4 of Directive (EU) 2019/790 lets anyone mine text and data from "lawfully accessible" works, for any purpose, commercial purposes included. Article 4(3) gives rightsholders a way out. The exception does not apply where the rightsholder has expressly reserved the use "in an appropriate manner, such as machine-readable means" for content made available online. Article 4(2) adds that copies may be kept only for as long as the mining requires.

Article 4(3), Directive (EU) 2019/790

in an appropriate manner, such as machine-readable means

Two duties follow for a collector. You must not take content from behind a barrier. You must find out whether the rightsholder has reserved its rights, and stop if it has.

The opt-out is read differently from country to country

Each member state wrote Article 4 into its own law, and the wording differs.

CountryHow the reservation is described
GermanySection 44b(3) of the Copyright Act says that a reservation for works available online is effective only if it is made in machine-readable form.
FranceArticle L. 122-5-3 of the Intellectual Property Code lets a rightsholder object by appropriate means. The implementing rules name machine-readable means for online content, including metadata and a site's terms and conditions.
ItalyArticle 70-quater of the Copyright Law allows mining where the rightsholder has not expressly reserved the use. The provision does not prescribe a format.
NetherlandsArticle 15o of the Copyright Act asks for an express reservation made in an appropriate manner, such as machine-readable means for a work made available online.
DenmarkSection 11 b of the Copyright Act implements Article 4.

A model trained in one country and offered in all of them meets every one of these readings. A team that follows only the narrowest idea of a reservation is exposed wherever a wider one applies.

What the courts have said so far

  1. Kneschke v LAION, Germany

    A photographer sued the organisation behind a large image dataset. In September 2024 the Hamburg Regional Court dismissed the claim under the exception for scientific research, and it remarked that a reservation written in ordinary language might count as machine-readable. In December 2025 the Hamburg Higher Regional Court dismissed the appeal. It held that building the dataset was text and data mining, and that the reservation in the photo agency's terms had not been shown to be machine-readable. The Federal Court of Justice heard the case on 3 September 2026. It indicated that it may refer questions to the Court of Justice of the European Union, and it has said it will announce its decision on 17 December 2026.

  2. DPG Media v HowardsHome, the Netherlands

    In October 2024 the Amsterdam District Court dismissed a claim by news publishers against a media monitoring service. The publishers had not shown a machine-readable reservation that covered the service's crawler. Robots.txt rules that named particular AI crawlers did not amount to a general reservation. An appeal has been filed.

  3. BoligPortal v ReData, Denmark

    In May 2026 the Eastern High Court lifted an interim injunction against a company that mined a property portal. The portal's reservation was written in ordinary language in a policy linked from its footer. The court held that a reservation must be one that automated systems can interpret and act on, and that this one did not meet the requirement. The decision came in interim proceedings.

  4. Like Company v Google, Court of Justice

    This is the first case before the Court on AI chatbots and the rights of press publishers. The Advocate General delivered an opinion on 3 November 2026. The opinion does not bind the Court, and the judgment is still to come.

What a collector should take from these cases

The courts so far have looked for a reservation that a machine can act on, and they have not accepted reservations written in prose on the evidence before them. The highest courts have not ruled, and French law names a site's terms as one way to reserve. Two lessons follow. A robots.txt rule or a reservation file is the signal most certain to count, and the Dutch case shows that it must be read for the exact crawler you use. You also cannot yet treat a reservation written in words as if it were not there. RightSignal reads both kinds and follows both.

The AI Act duties for general-purpose models

The AI Act places two copyright duties on every provider of a general-purpose AI model.

  • A copyright policy. The provider must have a policy to respect EU copyright law, and in particular to identify and respect reservations of rights made under Article 4(3), "including through state-of-the-art technologies".
  • A public summary of training content. The provider must publish a sufficiently detailed summary of the content used to train the model, on the template the Commission has issued.

These duties have applied since 2 August 2025. The Commission has been able to enforce them since 2 August 2026, with fines of up to 3% of worldwide annual turnover or 15 million euros, whichever is higher. Models placed on the market before 2 August 2025 have until 2 August 2027. The recitals to the Act say that the copyright policy duty applies to any provider placing a model on the EU market, wherever the training took place.

The General-Purpose AI Code of Practice, published in July 2025, sets out how a signatory meets the copyright duty. A signatory commits to use crawlers that read and follow robots.txt. It commits to identify and follow other appropriate machine-readable protocols for reserving rights. It commits to reproduce and extract only "lawfully accessible" content, without circumventing measures such as paywalls. It commits to exclude sites recognised as persistently infringing copyright and to give rightsholders a point of contact for complaints. The Commission consulted on which machine-readable protocols should be generally recognised and has not yet published its list.

The box you will have to fill in

The Commission's template for the public summary has a section on data processing. Section 3.1 asks whether the provider has signed the Code of Practice. It then asks the provider to describe the measures taken before and during data collection to respect reservations of rights, including the opt-out protocols and solutions honoured by the provider or by third parties from which datasets were obtained.

Published summaries show how providers answer today. One provider's summary, last updated on 23 July 2026, states that the model's training data includes text from Common Crawl. It answers yes to the question about the Code. It says that the provider applies rights-reservation and opt-out signals to public web datasets, including robots.txt signals for its own crawler, where those signals are available for the domains in the datasets. It adds that the crawler is not designed to circumvent captchas or paywalls.

The protocol that answer names is robots.txt, read for one crawler. The template also asks for the protocols honoured, and it asks about datasets that came from third parties. A provider that can answer in more detail, and can support each sentence with a record, is in a stronger position when a rightsholder or the AI Office reads the summary.

How RightSignal helps you prepare

What the law or the Code asksWhat RightSignal gives you

Identify and respect reservations under Article 4(3).

The EU profile reads reservations in the site's reservation file, in headers and page tags, in other signal files and in the site's own terms. Where a site has reserved, the outcome is Do not collect under the EU profile.

Follow robots.txt.

RightSignal reads robots.txt for the exact crawler you name and records the rule that applied and the time it was read.

Take only content you can reach without circumvention.

RightSignal records every login, paywall and bot challenge it meets, never passes one and returns a stop.

Keep copies only for as long as the mining requires.

Where the EU profile allows collection, the decision lists this as a condition.

Describe your measures in the public summary.

Your receipts are a dated record, source by source, of which signals were read, for which crawler and under which signed version of the Algorithm.

Account for datasets from third parties.

A Trace assessment shows what each site in a Common Crawl corpus said at crawl time and what it says today.

Respond to complaints from rightsholders.

A receipt quotes what the rightsholder's site said on the day. Matters, which is coming soon, will compile the record for you and your counsel.

A provider that works this way can describe its measures in the summary in specific terms. It can name the signals it reads, which go beyond robots.txt. It can say that each source has a sealed record. It can say what it did about third-party datasets. Each statement rests on evidence that the provider can produce on request.

What RightSignal does not do for you

Using RightSignal does not by itself meet any duty under the AI Act or under any other law. RightSignal does not write your copyright policy or your summary, and it does not tell you what the law requires of you. It gives you decisions made under a cautious reading of the law and the evidence behind them. Your policy, your summary and your legal position remain yours, and your lawyer should advise you on them.

If you are building in India for the EU market

The AI Act reaches a model trained in India once that model is placed on the EU market. An Indian team that plans to sell into Europe should collect under the India and EU Bridge profile from the start. The Bridge applies the stricter of the Indian and EU readings at every step, so each source in your dataset already has a record of the reservations that EU law asks about. Building that record later, for data already collected, is far harder, because a site's past signals are rarely preserved.

Follow every reservation you can read, and keep the record that shows you did.