General-purpose AI models and opt-outs: legal key for IPR holders

1. Application of the obligations of the Artificial Intelligence Regulation (EU) 2024/1689 for general-purpose AI models

Artificial intelligence ("AI"), as it has become popular and is widely used in society, arises from the emergence of general-purpose AI models ("AI models"). Specifically, those capable of generating content (text, graphics, music, etc.), and which fall within the so-called "generative AI", such as ChatGPT, Copilot, Midjourney or DALL-E.

These AI models have structurally altered the relationship between technology and copyright by resorting to the massive use of copyrighted content, without authorisation, for training and output generation.

In response to these challenges, the Artificial Intelligence Regulation (EU) 2024/1689 ("AI Regulation"), already in force, establishes a demanding regulatory framework whose implementation will be phased in over time.

The next step in its implementation process will take place from 2 August 2025. By that date, AI model providers will have to comply with, among other obligations, those under Article 53 of the AI Regulation, which includes a duty to establish guidelines to comply with EU copyright law and, in particular, to identify and respect any reservation of rights ("opt-out") expressed in accordance with Article 4(3) of Directive (EU) 2019/790 ("DSM").

This obligation requires AI model providers to implement appropriate technical mechanisms, such as tracking technologies, automated opt-out or machine-readable protocols, that allow for the exclusion from the training process of works in respect of which rightholders have expressly exercised their right to object in a clear, specific and technically recognisable manner.

2. What is an opt-out?

An opt-out is an express reservation of rights made by the rightholders of protected works, informing the providers of AI models that certain content may not be used in data mining processes to train their AI models.

This exclusion, recognised in Article 4.3 of the DSM, does not prevent public access and indexing of works when they are legally available online, but it does prohibit their automated reuse in mass data mining activities for AI model development purposes.

Recommended technical tools for expressing an effective opt-out:

3. Legislative developments and proposals

On 10 July 2025, the final version of the Code of Best Practices was published, which sets out the voluntary commitments made by the main providers of generative artificial intelligence models, in order to ensure an adequate technical implementation of the European legal framework. In particular, this Code establishes a series of minimum technical obligations that developers must comply with when carrying out data mining activities on content accessible online:

• Prohibition of circumvention of effective technological measures designed to restrict unauthorised access, including payment or subscription mechanisms (paywalls), in accordance with Article 6(3) of Directive 2001/29/EC.
• Actively exclude websites that have been identified by EU authorities or courts as persistent sources of copyright infringement on a commercial scale. The European Commission will maintain an updated list of such sites to facilitate their exclusion.
• Respect reservations of rights expressed by machine-readable means, in accordance with Article 4(3) of Directive (EU) 2019/790, by using:
  • Crawlers compatible with the robots.txt protocol according to the IETF RFC 9309 standard or its future versions recognised as technically feasible.
  • Other appropriate exclusion protocols (e.g. metadata by asset or location), which have been adopted by standardisation bodies or widely implemented by the cultural sectors concerned.
• They should provide accessible information on the crawlers they use and the opt-out respecting protocols, and enable automatic notification mechanisms (e.g. via a web feed ) so that rightholders can be aware of any changes in these practices.

In addition, the European Parliament's draft report on copyright and Generative AI was published on 27 June 2025, in which it proposes additional measures of particular relevance:

• Request to the Commission to propose a presumption of use of works and other content protected by copyright or related rights for AI model training, where providers have failed to comply with transparency obligations (provision 11).
• Establishment of a remuneration obligation for the use of protected works, including with retroactive effect (provision 4).
• Recommendation to use digital watermarks or identifiers as a complementary control mechanism (Recital Q).

 4. Relevant case law precedent

In September 2024, the Hamburg Regional Court ruled on a relevant case concerning the unauthorised use of copyrighted images embedded in a dataset used to train an AI model. The defendant invoked the exception of use for scientific research purposes, provided for in European law, as justification for the automated processing of the works.

Although the Court accepted the applicability of the exception of data mining for scientific research purposes, it did not at first rule conclusively on whether the generic formula of reservation of rights in natural language was sufficient as an effective opt-out under article 4.3 of the DSM. However, in an obiter dictum, the Court suggested that a clear expression of opposition formulated in natural language could be legally valid and technically machine-readable, given that technologies capable of identifying such exclusions already existed in 2021.

The ruling expressly recognised that there were no harmonised technical criteria that would allow rightholders to validly express their opposition to the use of their works in data mining contexts, which creates an area of legal uncertainty to be resolved by clear technical standards and interoperable mechanisms, such as those foreseen by the AI Regulation and the future repository managed by the European Union Intellectual Property Office ("EUIPO").

5. What can you do as an IPR holder?

If you own intellectual property rights in content accessible online, whether text, images, video, audio, software, databases or other protected creations, and you wish to prevent your works from being used in the training of AI models, we recommend that you consider adopting an opt-out mechanism in accordance with Article 4.3 of the DSM.

For this opposition to be legally valid and technically effective vis-à-vis AI providers, it must meet at least the following requirements:

• Apply machine-readable protocols: it is essential to use standardised systems that allow automated agents to identify and respect your will.
• Formulate a clear and specific expression of the uses to be prohibited, differentiating between:
  • Indexing or mere public viewing (which are not excluded by default), and.
  • Automated re-use for data mining and AI model training purposes (which can be excluded).
• Registering the reservation of rights in the future centralised opt-out management system to be managed by EUIPO, once operational. This registry will facilitate traceability, publicity and consultation by AI model providers and authorities.
• Consider the use of additional traceability technologies, such as visible or invisible digital watermarking, hash fingerprinting or fingerprinting techniques, to detect unauthorised use after the fact and to support potential legal action.

Early adoption of an effective opt-out mechanism not only preserves rights against unauthorised use in AI training, but also strengthens the bargaining position for future licence applications, compensation schemes or legal disputes.

In an evolving regulatory environment, a proactive and technically robust strategy will be key to protecting the value of content.

Briefing note written by the TMT area of ECIJA Madrid.