Training data

Training data means the body of texts, images, sounds, code or measurements from which an artificial intelligence model adjusts its parameters. Their collection and use engage several intellectual property rights at once: copyright in the content reproduced, the sui generis right of the database producer, trade secrets and, where applicable, related rights. The main framework derives from the text and data mining exception laid down in Article L. 122-5-3 of the French Intellectual Property Code, which transposes Directive (EU) 2019/790, and from the transparency obligations imposed on model providers by Regulation (EU) 2024/1689.

Reproduction of protected content

Building a training corpus requires copying the content, which falls within the reproduction right of the author or of the holder of related rights. That copying is lawful only if it falls within the scope of text and data mining, which permits the reproduction of lawfully accessible content for the purposes of automated analysis. The exception for scientific research is mandatory, whereas the general exception yields to an opt-out expressed by the rightholder, in particular by machine-readable means for content made available online.

Databases and secrecy

Where the data are extracted from a structured database, the operation may amount to a substantial extraction within the meaning of the database producer’s right, which exists alongside copyright and is subject to its own exceptions. The datasets assembled by a company, their selection, annotation and cleaning may themselves be protected as databases or as trade secrets, provided that reasonable protective measures have been taken.

Transparency imposed on model providers

The Artificial Intelligence Act requires providers of general-purpose models to put in place a policy to comply with copyright law, including the identification of and compliance with the reservations expressed by rightholders, and to publish a sufficiently detailed summary of the content used for training. Those obligations create no new right for authors, but they make it easier to prove that a work has been used and open the way to copyright infringement actions or to licence negotiations.

Key points

Training a model relies on large-scale reproductions that are lawful only within the text and data mining exception or under a licence. Rightholders who wish to oppose the use of their content must express their reservation in machine-readable form and document it. Developers must trace the origin of each dataset, check the reservations, keep evidence of lawful access and anticipate the transparency obligations that will allow them to justify their policy in the event of a claim.

---

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