Copyright and artificial intelligence

Copyright and artificial intelligence meet at two levels. Upstream, the training of generative models involves the large-scale reproduction of protected works, the lawfulness of which depends on the text and data mining exceptions derived from Directive (EU) 2019/790. Downstream, generated content raises the question of its protection, as French copyright law and EU law require an original human creation. Regulation (EU) 2024/1689 adds transparency obligations for providers of general-purpose models.

Model training and the exceptions

The reproduction of works to build training data falls within the reproduction right. It is covered by the mining exception for scientific research purposes for the benefit of research organisations and cultural heritage institutions, and by the general mining exception, open to all, provided that access to the works is lawful and that the rightholder has not reserved its rights in an appropriate manner, for instance by machine-readable means for online content. The conditions for the validity of this reservation, known as an opt-out, and the scope of the exception are the subject of litigation in several Member States, the outcomes of which are not yet settled.

Transparency imposed by the Artificial Intelligence Regulation

The Artificial Intelligence Regulation requires providers of general-purpose models to put in place a policy to comply with EU copyright law, including the identification of and compliance with reservations of rights against mining, and to publish a sufficiently detailed summary of the content used for training, according to a template drawn up by the Commission. These obligations apply wherever the training takes place, as long as the model is placed on the EU market. They create no right to remuneration but make it easier to prove that works have been used.

The protection of generated content

Copyright protects original works of the mind, which presupposes a natural person as author and free and creative choices. Content produced autonomously by a system, from a mere prompt, is not protected, for want of the imprint of a personality, and falls into the public domain. By contrast, AI-assisted content, in which the user has made sufficient creative choices in the conception, selection or reworking, may be protected on account of the human contribution. The degree of control required remains debated. The reproduction, in generated content, of the original elements of a pre-existing work constitutes infringement attributable to the person exploiting it, whatever the machine’s share.

Key points

Model training is lawful under the text and data mining exception, unless the rightholder has reserved its rights in an appropriate manner, and providers must publish a summary of the content used. Content generated without human creative input is not protected, whereas assisted creation may be. A rightholder should express its reservation in machine-readable terms. A user should keep a record of his or her creative choices and check that the output does not reproduce an identifiable work.

---

This glossary entry is general in scope and does not constitute legal advice.

Dhenne Avocats is a patent litigation firm. Matters outside that field, such as trade marks, designs, copyright and IT contracts, are handled by practitioners the firm works with regularly, and by the firm itself when they arise in a patent dispute.

Intellectual property attorney · Tell us about your situation