Rightsholders may reserve their content against certain uses for training artificial intelligence. But how should a reservation be expressed so that it can be understood? How can it remain recognisable as the work circulates? And how can compliance be verified? Europe has just closed the third of its consultations seeking answers to these questions.
Europe is examining various solutions to give effect to reservations of rights against the training of artificial intelligence systems. The work focuses on three closely related questions, which concern how to express the reservation, how to link it to the content and how to verify subsequently that it has been respected.
How can AI systems be made to recognise a reservation of rights?
As is well known, Article 4 of Directive (EU) 2019/790, transposed into Spanish law by Article 67 of Royal Decree-Law 24/2021, permits text and data mining (TDM) of content to which there is lawful access, unless the rightsholder has appropriately reserved its rights, which, for content made publicly available online, requires machine-readable means. The difficulty, which has been debated for some time, lies in the fact that there does not yet appear to be a generally accepted technical format that meets this requirement.
Is it sufficient, for instance, for a photo agency to prohibit AI training on its images in the terms of use of its website? In the LAION case, discussed here, the Hamburg Higher Regional Court held, in its judgment of 10 December 2025, that a natural-language clause did not, in that instance, meet the machine-readability requirement. As also noted here, a reservation should arguably be not merely “readable” by machines but “actionable”, that is, capable of being interpreted automatically when content is extracted. In the absence of a uniform judicial approach, uncertainty persists.
The Commission therefore launched, in December 2025, a consultation on technical protocols for reserving rights against TDM, which closed on 23 January 2026, with a view to identifying accepted mechanisms that allow reservations to be expressed and recognised automatically.
Would a European registry help reservations travel with content?
Even if the previous issue were resolved by establishing a common protocol, reservations might still be difficult to detect once works circulate beyond their original location. A robots.txt file, for example, does not necessarily accompany a photograph that has been scraped and subsequently republished by a third party.
To examine this difficulty, on 13 July 2026 the Commission published a feasibility study on a possible European registry of rights reservations, which considers whether a common infrastructure could associate reservations with works identifiable through metadata, identifiers or digital fingerprints (fingerprinting), even where the content is hosted on a platform other than the original one.
The creation of the registry has not yet been decided and leaves several questions open, such as who may make the reservation or how it should be updated if a licence is granted. It also has an inherent limitation, in that it would help to establish which works are reserved, but would not by itself demonstrate that a provider used them to train a model.
How can compliance with reservations be verified, and what about remuneration?
This last limitation helps to explain the targeted consultation launched by the Commission on 29 September 2026, the deadline for which was extended until 10 November. Among other matters, the consultation addresses AI training and puts forward measures such as the following.
- Recognition of reservations. Possible common standards, a European registry and an obligation for AI crawlers and agents to disclose their identity and purpose.
- Verification of compliance. Record-keeping obligations, independent audits and the provision of information, subject to confidentiality agreements, to rightsholders and collective management organisations that request it.
- Burden of proof. A possible rebuttable presumption that protected content has been used, subject to specific conditions, for instance where the provider fails to comply with its transparency obligations under the AI Act.
The consultation ties in with Article 53 of the AI Act which, as discussed in this post, requires providers of general-purpose AI models to identify and respect reservations of rights and to publish a sufficiently detailed summary of the content used for training.
Lastly, the Commission seeks views on the difficulties of negotiating licences and on remuneration. An effective reservation may require authorisation for certain acts of TDM, but does not in itself facilitate agreement between rightsholders and providers. The options under consideration therefore include a licensing marketplace, model contractual clauses, an obligation to negotiate in good faith, mediation or arbitration, and collective licensing, including extended collective licensing, distinguishing between training, inference and outputs. The questionnaire also invites views on possible equitable compensation or a new remuneration right.
The three initiatives therefore address a sequence of questions, from how to express a reservation and how to keep it recognisable as the work circulates to how to verify that it has been respected. The September consultation adds the question of how to facilitate authorisations and, where appropriate, remuneration.
While European solutions take shape, rightsholders would be well advised to review their reservation mechanisms, their consistency with the licences granted and the evidence of when and how they communicated them, while AI providers should document the provenance of their content and how they detect reservations. Both will also need to keep a close eye on a landscape that, like the technology itself, changes by the day.
Cristina Delgado

