Does AI steal from authors?

Copying to train. The 2019 European directive allows lawfully accessible works to be reproduced "for the purposes of text and data mining", provided that the author has not "expressly reserved" their use "in an appropriate manner, such as machine-readable means in the case of content made publicly available online". In December 2025, the Hanseatic Higher Regional Court (Hamburg) held lawful the copying of a photograph by the association LAION, which builds datasets used to train image models: the reservation displayed on the website of the agency distributing the photograph did not have the machine-readable form required by German law. The court also allowed the copying under the scientific research exception; the case concerns the building of a dataset, not the training of a model by a company. The Regional Court of Munich I accepts, for its part, that the text and data mining exception covers the copies made to build a training corpus. The Advocate General of the Court of Justice of the European Union suggested on 3 November 2026 that the training of a language model should, as a general rule, be covered by this exception, while judging the question hypothetical in the case before him.

Keeping in memory. A model can learn a text well enough to reproduce it. Researchers at Stanford and Cornell showed that Llama 3.1 70B, an open-weight model from Meta, allows long passages to be extracted from some books, including nearly all of the first Harry Potter and of 1984, while most of the books tested are not memorised. The Regional Court of Munich I drew a legal consequence: if a model reproduces song lyrics identically, they are reproduced in its parameters, and this reproduction is not covered by the text and data mining exception. The High Court in London ruled the opposite for an image model: Stable Diffusion "does not store" Getty's images and is therefore not an infringing copy of them. The 2 cases concern neither the same works nor the same right.

Reproducing. This is the action on which the position is clearest. In Munich, OpenAI is held liable for the lyrics of 9 songs that ChatGPT reproduced, and the court holds the provider responsible, not the user who asked the question. The European code of practice for general-purpose models, of July 2025, asks signatories for technical safeguards against such reproduction.

To take action, an author must first show that their work was used. Reproduction can be observed: you ask the question and compare the text. Use for training does not show in the answers when the model has not memorised the work, which is the most frequent case. Only the provider knows what it copied.

The EU AI Act, applicable to general-purpose models since 2 August 2025, asks their providers for a copyright compliance policy and "a sufficiently detailed summary" of the training content. The summary template published by the Commission in July 2025 asks for the main data collections to be listed and for other sources to be explained: it does not ask for the list of works. The code of practice commits signatories to receiving complaints from rightsholders that are "sufficiently precise and adequately substantiated", which assumes that the author already has the means to substantiate them.

On 8 April 2026, the Senate voted unanimously for a bill that creates a presumption of use of works by an AI system when an indication makes it plausible: a technical expert report, reproduction of extracts, an output "in the style of" the author. This is a rebuttable presumption: the provider can prove the contrary. In the National Assembly, the cultural affairs committee adopted it on 2 June 2026; its examination in plenary did not finish on 11 June, and it is scheduled for 19 November 2026.