Generative artificial intelligence has reopened a classic copyright question with renewed intensity. Can a company train a model using books, images, songs, articles or databases created by third parties without obtaining individual authorisation from each rights holder? In Europe and Spain, the answer is not absolute. It depends on the content concerned, how it was obtained, whether the rights holder has reserved their rights, and the subsequent use made of the model.<\/p>\n
European legislation starts from a practical premise. In order to develop certain technologies, it is necessary to analyse large volumes of information. For this reason, Directive 2019\/790 introduced text and data mining exceptions, which were implemented in Spain through Royal Decree-Law 24\/2021. This legislation defines text and data mining as an automated analytical technique aimed at analysing digital texts and data in order to generate information, patterns, trends or correlations.<\/p>\n
Article 67 of Royal Decree-Law 24\/2021 permits reproductions and extractions of lawfully accessible works for text and data mining purposes without the authorisation of the rights holder. However, this legal authorisation is subject to important limitations. The content must be lawfully accessible, the copies may only be retained for as long as necessary, and data protection rules must be complied with. In addition, for uses that are not strictly scientific, the rights holder may expressly reserve their rights by means of machine-readable tools or other appropriate means. In such cases, the developer must exclude those contents or obtain the relevant licence.<\/p>\n
Therefore, training an AI system on copyright-protected works is not automatically unlawful, but it is not exempt from legal scrutiny either. The key issue is not merely whether the work was available online, but whether access to it was lawful and whether a valid reservation of rights was in place.<\/p>\n
Courts are still developing their approach. In the United States, several proceedings against OpenAI, Meta, Anthropic, Midjourney and other developers have considered whether model training amounts to transformative use or to unauthorised reproduction<\/a>. Some rulings have favoured developers where the training did not replace the original work and the model’s outputs did not reproduce substantial parts of protected content. However, courts have also drawn a clear line in relation to the use of pirated copies.<\/p>\n In Bartz v. Anthropic, the court held that training could be covered by fair use, but that the retention of pirated copies could not. In Kadrey v. Meta, the court adopted a more favourable position towards Meta in relation to training, while leaving open the debate on potential market harm. In Thomson Reuters v. Ross Intelligence, concerning an AI-based legal tool, the court rejected fair use where protected content was used to create a competing product. In the United Kingdom, Getty Images v. Stability AI has not definitively resolved the lawfulness of training either, but it has highlighted the difficulties of applying traditional infringement categories to models trained outside the territory.<\/p>\n For rights holders, the first step is to document the reservation of rights in a clear, express and technically recognisable manner. For online content, this may be done through legal notices, terms of use, metadata, exclusion protocols or machine-readable instructions addressed to crawlers and automated systems. Such reservation should be consistent across the website, and evidence of its implementation should be retained.<\/p>\n For AI developers, the recommended approach is to maintain robust dataset traceability. This means identifying sources, verifying licences, excluding content subject to opt-out, avoiding materials from unlawful repositories, documenting the filters applied and reviewing whether the model’s outputs reproduce protected works in a substantial manner.<\/p>\n The EU Artificial Intelligence Act reinforces this approach. Providers of general-purpose AI models<\/a> will be required to have policies in place to comply with EU copyright law, respect reservations of rights and publish a sufficiently detailed summary of the content used to train their models. Ultimately, the future does not lie in opposing copyright and technology, but in combining innovation, transparency and licensing where necessary.<\/p>\n <\/p> How to Exercise Opt-Out Rights Over Your Content and Train Models Without Infringing Copyright<\/h2>\n
Contact Us<\/h2>\n
<\/ul><\/div>\n