« Some songs from this album may have been created using artificial intelligence Now indicate certain Deezer sheets. The service has indeed announced, Friday, June 20, the commissioning of a labeling tool for songs generated by AI on its online music listening platform.
To do this, he operates an automated detection technology, developed by him, which would have given rise to the filing of two patents, and above all allowed Deezer to measure the explosion of the publication of music generated by AI.
Last April, the company said that ” 18% of new music published every day on its platform are entirely generated by artificial intelligence ». She then mentioned a volume doubled in just three months.
0.5% of listening, but 70% fraud
Many songs, but few listenings: the music generated by AI would only represent 0.5% of the platform streams. However, it would serve another phenomenon: that of fraudulent listening, that is to say behaviors of simulated internet users on the platform, used to artificially inflate the audience and therefore the income of certain artists. “” Today, up to 70 % of listening to titles entirely generated by AI are considered fraudulent “Says Deezer, without detailing the terms of this calculation.
In response, the platform has therefore embarked on the development of a detection algorithm ” Capable of identifying 100 % music generated by the most advanced generative models – such as Suno or UDIO – with the possibility of adding new tools, provided you have the necessary data ».
Deezer claims to be able to detect, directly in the audio signal, markers associated with the tool used for generation, at least if it is Suno or Udio, considered as market leaders. “” When IA algorithms generate new song, they have species of small noises that identify them, specific to them (…) that we will be able to find. It is not audible in the ear but it is visible in the audio signal “, Described to AFP Alexis Lanternier, Managing Director of Deezer. The tool is, according to him, ” Reliable at 98% ».
« Concretely, you will finally know if what you listen to was created by a human … or not », Presents Deezer again on one of the pages of his forum. The platform specifies that the pieces generated by AI will no longer be highlighted in its editorialized reading lists or in algorithmic recommendations.
How to manage these productions generated by AI?
Why not purely and simply delete them? “” Or even better, why not block them at the entrance by subjecting them to this detection test before publication, in order to prevent them from unnecessarily invading the platform? Asks a user. Apart from cases of plagiarism or copyright violation, the question is not trivial: what about false positives, or pieces that would only be partially generated by AI?
In the absence of an a priori blockage, Deezer has already shown that it was possible to clean up: in April 2024, the company had indicated that it had deleted over six months some 26 million titles (the equivalent of 13% of its catalog), ” including non -artistic content, noises and duplicates », Thanks to its automated detection tools.
How far will Deezer automated tools be effective, especially if new music generation models appear? In the field of text or image, detection models are often pointed out for their poor efficiency, and our own work on the proliferation of news sites generated by IA conclude that the need for manual analysis.
The French platform is not the only one to study this automated approach. YouTube announced last September that he was working on the detection of the visual or sound elements generated by AI within his technology to combat copyright offenses, Content ID.
Beyond technical efficiency, Deezer mainly adopts a posture of control and supervision of practices related to AI, where his large Spotify rival is suspected of having, on the contrary, voluntarily disseminated pieces created by false artists within certain themed reading lists.
