Artificial intelligence and copyright in Europe: what has been decided in 2026

Verified as of 27 September 2026. It will be updated once the pending rulings are handed down. If anything has changed, let me know.

The question of whether it is lawful to train an artificial intelligence on protected works does not yet have a single answer in Europe. The rules govern specific acts; courts have examined distinct phases and facts; several rulings are not final, and central issues remain pending before higher courts.

This guide separates the regulatory framework, judicial decisions, and matters that remain unresolved. It is not legal advice.

The Copyright Directive

Directive (EU) 2019/790 introduced two mandatory exceptions for text and data mining, defined as the automated analysis of works or other digital materials to generate information, including patterns, trends, and correlations.

Article 3 permits research organisations and cultural heritage institutions to perform reproductions and extractions for scientific research on materials to which they have lawful access. This exception does not permit a unilateral reservation by the rights holder. Rights holders may apply proportionate measures to protect the security and integrity of their networks and databases.

Article 4 permits reproductions and extractions for text and data mining for other purposes, including commercial ones, provided that access is lawful and the rights holder has not expressly reserved such uses. For content made publicly available online, the reservation must be made in an appropriate manner and by machine-readable means. Recital 18 mentions metadata and the terms of a website or service as examples, but it does not render every phrase on a website an effective reservation in all cases.

The Directive does not use the term 'artificial intelligence training'. Applying these exceptions to reproductions made during collection, cleaning, or training is a legal interpretation that depends on the act examined and its conditions. It also does not automatically cover what the model retains or the results it generates.

The AI Act

Regulation (EU) 2024/1689 requires providers of general-purpose AI models to:

establish a policy to comply with Union copyright law;
identify and respect reservations made in accordance with Article 4(3) of Directive 2019/790;
publish a sufficiently detailed summary of the content used to train the model, according to the AI Office template.

These obligations do not grant a licence nor do they declare training to be lawful. The AI Act adds compliance and transparency duties, but the lawfulness of each reproduction continues to be examined in accordance with applicable copyright law.

The rules for general-purpose models apply from 2 August 2025. Enforcement by the Commission regarding new models commenced one year later. Providers of models placed on the market before 2 August 2025 generally have until 2 August 2027 to comply with the corresponding obligations.

The Code of Practice for general-purpose models, published on 10 July 2025, is voluntary. Its chapters on transparency and copyright offer signatories a pathway to demonstrate compliance with Article 53. Adherence may provide a practical presumption of regulatory compliance within that framework, but it does not, of itself, resolve a copyright dispute or legitimise every work used.

Judicial findings

Pelham: pastiche is not a rule on training

On 14 April 2026, the Court of Justice ruled on case C-590/23, Pelham. It held that the pastiche exception may cover a creation that evokes one or more existing works, is perceptibly different from them, and uses characteristic elements to engage in a recognisable artistic or creative dialogue. The litigation concerned a sample of approximately two seconds from a Kraftwerk recording. It did not address the training of AI models.

The Federal Court of Justice subsequently applied this interpretation to the national dispute. The result may be relevant for certain generated content that engages with previous works, but it does not address whether it was lawful to collect or copy the works during training.

Kneschke v LAION: creation of a dataset

The Higher Regional Court of Hamburg ruled on 10 December 2025 that the downloading of a photograph to verify the correspondence between an image and its description during the creation of a dataset could be covered by the German mining exception for research. It also examined the reservation included in an agency's terms and concluded that its machine-readability had not been proven at the relevant time.

The ruling did not decide that all training was lawful. The dataset published by LAION contained links and metadata, not the images downloaded during the verification process. The case is under appeal. Following the hearing on 3 September 2026, the Bundesgerichtshof has scheduled the judgment for 17 December 2026, case I ZR 281/25.

GEMA v OpenAI: memorised lyrics

The Landgericht München I upheld GEMA's claim in essence on 11 November 2025 regarding nine song lyrics. The chamber considered it proven that they were reproducible from models 4 and 4o, classified the memorisation within their parameters as reproduction, and identified further infringements in the chatbot's responses. It rejected the argument that these acts were covered by the mining exception.

The ruling relies on the substantial retrieval of a specific group of lyrics. It does not establish that all training, all parameters, or all model outputs always constitute a reproduction. It should not be presented as a general and definitive rule for any system. The judgment was delivered at first instance.

GEMA v Suno: music, model, and outputs

On 31 July 2026, the same chamber largely upheld GEMA's claims against Suno regarding six musical compositions. It held it proven that the works formed part of the training data, had been extracted from YouTube via stream-ripping, and could be recognised in outputs obtained through numerous attempts and prompts that included title, style, and lyrics, but not melody, harmony, rhythm, or arrangement.

The official note indicates that the court examined alleged infringements in the United States and Germany, memorisation, and outputs. According to the ruling and published analyses, it applied US law to the training conducted there and rejected the fair use defence, in addition to identifying acts subject to German law related to the model and the outputs. It is a first-instance judgment and not a general declaration against any musical model.

Getty Images v Stability AI: training was not resolved

The High Court of England and Wales delivered judgment on 4 November 2025. During the trial, Getty withdrew its primary claim for direct infringement linked to training as it could not prove that this had occurred in the United Kingdom. The court also did not consider the model to be an 'infringing copy' for the purposes of importation, because it did not store or reproduce the protected works in that sense. Getty achieved limited success on certain trade mark claims related to outputs that displayed signs similar to their watermarks.

The judgment did not declare training to be lawful. It did not reach a resolution on that substantive issue due to territorial and evidentiary reasons.

Pending matters

Case C-250/25, Like Company v Google Ireland, refers questions to the Court of Justice regarding the reproduction of works during the training of a language model, the potential application of the Article 4 exception, and the value of rights reservations. As of 27 September 2026, no published judgment has been located.

Also pending are:

the judgment of the Bundesgerichtshof in I ZR 281/25, expected on 17 December 2026;
appeals or the finality of German first-instance judgments;
the practical development of machine-readable reservation protocols;
the judicial application of content summaries and policies required by the AI Act.

Spain

I have not located any published Spanish case law that directly resolves the lawfulness of training a generative model on protected works.

The Ministry of Culture has defended the principles of authorisation, remuneration, recognition, and transparency before the World Intellectual Property Organization. In May 2026, it promoted the debate on commercial mining, legal certainty for training, and transparency at the Council of the European Union's Ministers of Culture.

The Council of Ministers approved the draft Organic Law for the proper use and governance of artificial intelligence on 26 May 2026 and submitted it to the Cortes. The draft develops national supervision of the AI Regulation. This initiative does not equate to an approved reform of the Intellectual Property Law, nor does it, of itself, establish a Spanish system of licensing or remuneration for training.

A phase-based reading

What follows is my own interpretation. The cases do not yet form a coherent European rule on 'training' considered as a single activity. Courts are isolating different acts:

accessing and downloading;
selecting and cleaning a dataset;
copying during training;
retaining information reproducible in the parameters;
offering the model in a territory;
generating and communicating outputs.

An exception may cover one act without covering the next. A reservation may be effective against one mode of mining and not resolve the reproduction in the output. The coincidence between an output and a work may demonstrate memorisation in a specific case without proving how the entire model functions.

For a creator, the useful question is not only whether a work 'trained' a system. It also matters where it was obtained, what reservation existed, where the copies occurred, what remains retrievable, what the model produces, and in which territory the service is offered.

What to monitor

17 December 2026: German judgment on Kneschke v LAION.
Case C-250/25: future response from the Court of Justice regarding training, reproduction, and mining.
Munich judgments: appeals, finality, and potential revision of their concept of memorisation.
Getty v Stability AI: evolution of the appeal and territorial scope of the claims.
Rights reservations: recognised technical standards and their judicial interpretation.
AI Act: initial enforcement measures regarding rights policies and training summaries.

On the open conversation

This guide summarises regulations, judgments, and official communications published up to 27 September 2026. It distinguishes between final rulings and first-instance decisions, and between what has been decided and what remains pending. The phased reading is an interpretation. If you work in copyright, machine learning, or creative practice and notice data that requires correction, the guide remains open.