Is anything fair game when it comes to training artificial intelligence?
Javier López warns that the idea is gaining ground that AI can freely draw on any content available online to train its models – a belief reminiscent of the false premise that ‘everything on the internet is free’. In response to this, he points out that behind such content there are rights holders and significant creative and financial investments, and that the rights holder holds the exclusive right to authorise or prohibit the reproduction, distribution, public communication and adaptation of their work. This rule, set out in Article 17.1 of the Consolidated Text of the Intellectual Property Law, in Directive 2001/29/EC (InfoSoc) and in Section 106 of the US Copyright Act, does not cease to apply simply because the use is carried out via AI systems.
The debate has now moved to the US courts, with numerous lawsuits filed against companies such as OpenAI and Microsoft. Of particular significance is the lawsuit brought by The New York Times over the unauthorised use of its archive, which illustrates the substitution effect that occurs when AI provides users with content that they previously obtained by visiting a website or purchasing a publication, thereby eroding the revenue streams of rights holders.
In the European Union, the text and data mining exception under Directive 2019/790 (transposed in Spain by Royal Decree-Law 24/2021) does not presuppose that content must be free of charge for the purposes of training AI, but rather seeks to strike a balance between legitimate access and the possibility of expressly reserving rights. In this regard, López cites the ruling of the Munich Regional Court in the case of GEMA v. OpenAI, which establishes that text and data mining cannot justify the storage of protected works and their reproduction in the results provided by AI. The discussion thus becomes a matter of evidence and technical detail: metadata, standards such as C2PA, opt-out registers or protocols such as ai.txt do not in themselves prevent misuse, but they do make it possible to demonstrate that the system could reasonably identify the reservation of rights.
The author also addresses the other side of the debate: the protection of works generated entirely by AI. Given that AI cannot be the holder of rights, one might consider that there is neither an author nor any rights, or attribute authorship to a natural person, without it being clear to whom. López concludes that the principle to be upheld is a model based on consent, traceability and the value of intellectual creations, so that technology does not override the law, but rather requires an appropriate legal solution.
Read the full article on LegalToday.