Music pieces, images, voices, styles, rhythms… what still remains artificial intelligence (ia) the genitive when they” trained » And they swallow billions of data collected on the web? , data scraping (in French, the act of sucking data) it’s like artist’s Now prefer not to put anything online for fear of having their art or their style seen and copied to perfection by one of the AI systems, these tools that make it possible to generate text, code, music, voice or images Are.
Is this radical solution justified? copyright Does that not adequately protect artists and rights holders from such extremists? Contrary to what one might think, the law, in the face of AI, is not on the authors’ side, at least for the time being.
AI deprived artist?
And it’s to prove exactly the opposite that a handful of artists and rights holders have initiated legal proceedings Against image building AI companies in US and UK. Getty Images, a forum for copyright-protected photographs, for example, realized that its images had been used to train Stable Diffusion without its authorisation.
The company, which filed a complaint last February, believes that this AI system may have massively plagiarized their photos. One month ago, attack on three artists Same stable spread, but also its competitors Midjourney and DeviantArt. Reason, same: billions of images have been collected for these AIs to “learn from”, including works protected by copyright.
Within the European Union, some go further by asking the legislator directly. change the law Which, according to him, would not adequately protect artists against artificial intelligence. The latest initiative, UVA (united voice artist), an organization that brings together unions and associations of voice-over actors for films, series and animation from several countries – including France. On May 24, this association sent a manifesto For European decision-makers asking to change the rules of the game, among other things, are concerned about AI’s ability to clone too many voices. The latter wants the collection of votes to be previously authorized and contracted for training purposes – against remuneration. That goes for any voice cloning that should be properly the subject of a contract with the person concerned, she adds.
Yes, AI can invalidate copyrighted works
Because, contrary to what one might think, AI can, in some countries, aggregate legally protected works for training. This is the case in Europe, where the 2019 Copyright Directive – transposed into France and therefore applicable in France – provides precisely for this particular case. Generally, the copyright in title forbids the use of any work except with the authorization of the creator or his successors. A company, for example, wouldn’t be able to use a piece by The Beatles for advertising without the agreement of their heirs and the payment of some salty royalties. But there are exceptions, and this Instructable provides one: the so-called exception to text and data mining (right of discovery, in French). It specifically targets AI systems that collect data for training. Security measure only: This archive is not possible when it comes to commercial use, if the author has explicitly opposed it.
To understand this exception, you have to put yourself in the context of 2019, when the directive was drawn up, says Arnaud Latil, lecturer in private law at the Sorbonne University and member Sorbonne Center for Artificial Intelligence (SCAI). At that point, ChatGPT, Midjourney, and all the others haven’t launched yet. , AI training issues weren’t front and center “, he explains. Legislators were therefore not necessarily aware of the consequences of this exception.
Additional problem: The way the author opposes this collection is also not specified in the text. Especially since practice, the AI will neither ask nor warn when collecting a work of art or music title to train on. Nor shall it differentiate between a work in the public domain or a work protected by copyright, since it scrap on a large scale, without any distinction. In addition, it is not necessary to mention copyright protection on the website. However, just because there is no inscription on an image does not mean that it has come into the public domain and is free of rights.
But today the context is quite different. Artists and rights holders are against generative AI and their training based on their works. , The question may seem to be settled legally, but politically it arises in a very important way today., What about the rights of authors whose works have been trained by generative AI? Arnaud asks Latil.
the question arises when scraped data The voice in question is. voice-over or dubbing actor who describes himself as ” ai loot », do not hesitate to talk about « flight About their work when generative AI collects their accessible voices through replays, podcasts, or ads served online. The association Les Voixes, which brings together dubbing voice artists, has for example launched several alerts and recommended that artists multiply mentions on sites and in their contracts, specifying that they are used for AI training purposes. Refuse to use your voice for.
Content Creation: For Now, It’s a Vacuum (Legal)
Another thorny topic, the question of AI-generated content that will perfectly emulate a style, a genre, or an artist’s voice, is the equivalent of creating new works.” Like Of an artist Last month, an AI-generated song that cloned the styles of Drake and The Weeknd found tremendous success before being pulled from Universal Music Group’s on-demand music streaming platform.
, In the United States, actors have also noticed listening to programs they’ve never participated in. It was about their voices, their timing. After being collected without authorization for training, which prompted the AI. , Ongoing trial, refund request. In France, the problem hasn’t arrived yet because we’ve raised the alarm », explains Patrick Cuban, voice-over actor and co-founder of the association Les Voix.
The latter asks politicians to take up the problem and agree on it ” A moratorium to prevent widespread looting “, in addition to a traceability of what has been generated by the AI - which will indicate which artists have given consent for the AI to use their voices, with license numbers, GDPR certification …
Because till now, this specific point of AI generated content has not been fixed by any text. , On the other hand, we have no answer. It is not provided anywhere. Arnaud confirms Latil. Because the law only protects a particular work and not its style – for example, copying a painting by Chagall would be considered infringing, i.e. infringing the painter’s copyright. But a painting done in the manner of Chagall would not. In law, it is often said that ideas, styles or genres are free class. Translation: We do not protect these elements by copyright. Only the result of creation will be protected.
The topic has become so sensitive that European lawmakers, during their review of AI regulation, added a provision specifying that generative AI must respect copyright, without indicating otherwise. For the teacher, it is necessary, in all cases, Adapt copyright to AI, either by providing a new text governing copyright for application to AI, or by reforming existing legislation. But, he adds, Debate should open on copyright and AI ,