On Friday (June 20), French streaming service Deezer launched its latest tool to combat what it calls the “spamming of AI-generated songs.” Now, the service will tag every song it identifies as fully AI-generated so that users have transparency when they find the content on the service.
The tagging tool is the latest in a series of announcements from Deezer about how AI-generated music is increasing rapidly on its service — and causing harm in the process. According to Deezer, up to 70% of the streams generated on fully AI-generated tracks are deemed fraudulent or artificial.
Deezer began sounding the alarm in January, when it reported that its proprietary AI detection tool identified that 10% of songs uploaded to Deezer daily were fully AI-generated. While the company says it has no issue with human artists using AI as part of their creative process, the purely AI-generated songs tend to be used to spam the platform or siphon royalties away from human artists. To combat this, Deezer began removing fully AI songs from their recommendations.
Then, in April, Deezer claimed the figure nearly doubled, reporting that 18% of daily uploads were fully AI-generated. In a new interview with Billboard, Deezer director of research Manuel Moussallam attributes that jump to two key reasons: the further adoption of AI music tools like Suno and Udio; and, simply, the fact that “our data got better.”
“Part of the stuff we were catching as AI-generated in January, we weren’t totally sure about, so we were very conservative in the numbers we reported. We didn’t want any false positives,” Moussallam explains. “I think the 18% [figure] is actually much more accurate and closer to what we actually saw from January, but still, that number is increasing.”
While most of its competitors have not yet taken a public stance on AI-generated music or created any rules specifically to tackle issues it may present, Deezer believes there’s no way to ignore what is happening. In a new interview with Billboard, chief innovation officer Aurelien Herault, along with Moussallam, explains that the clock is ticking: “At some point, the industry will need to make a decision on this because AI music will exist forever now. There’s no going back.”
You’ve both said before that your team had a lot of conversations with the music business about the rules you were developing at Deezer before you launched them. Have you had those same conversations with AI companies, like Suno or Udio?
Herault: Not officially. We are in touch with some of the researchers who work there because we go to the same conferences, but no, we didn’t make contact. Actually, I don’t think they care a lot about us being able to detect their outputs because the issue is on our side, because people are using these tools massively to spam our catalog.
Are you starting to see patterns of how people are using AI content for spam? For example, are most people creating multiple accounts and hopping between them to spread out their mass uploads? Are they artificially running up the streams?
Herault: What I can say is that they are creative.
Moussallam: Well, not creative on the music side, but on the spam side? Very creative.
Herault: The reason why we got worried about AI-generated music in the first place is that it’s really close to the fraud behavior we were already seeing. You create a lot of accounts. You deliver a massive amount of albums every day.
Moussallam: I think maybe the only new phenomenon we see from AI music is it can also be used for impersonation. You take one artist that streams a lot and then try to create stuff that resembles the style of that artist. That’s quite worrying on our side.
Herault: They even try to deliver these songs on official artist pages. That’s why it’s important to combine our AI efforts with our fraud detection system to avoid this kind of behavior.
Take me back to when you first discovered how much AI-generated content was being uploaded to Deezer and why you decided to devote so much time and so many resources within the company to work on this issue.
Moussallam: I think it dates back to maybe 2022 when the first large open-source generative models were released within the scientific community. It was before you had commercial models available, and we had a chance to play with it ourselves. Over time, we started to see momentum with these models. Soon they could generate full songs, which was not possible a few years before.
Obviously, when Suno launched, it was a huge moment. The first thing we did was to go on the platform and download a few songs that people were sharing over there, and then, using fingerprints, we were able to check whether people were also distributing these AI songs into our catalog. We soon found out that, yeah, they were. Back then, we had no idea of the amount, but we knew it was happening. It was really important for us to quantify this phenomenon, and this is what led us to actually researching it.
It took us a few months, and we finally found a system to identify the scale. Eventually, we felt confident enough to share the number in January. We had to double-check a lot, and to make sure we didn’t have too many false positives. We knew that would be a huge problem. We only communicated to the general public once we were confident.
Herault: Today, we are talking about AI detection, but we’ve seen the same spam and artificial streaming behavior with other kinds of content on Deezer, like public domain songs, noise and rain sounds. So we’ve been developing tools to understand what is on our platform before AI came along. It’s been 12 years now of research and development, so that’s why we can react pretty quickly when new developments occur.
I often hear mixed reviews about the products that are on the market right now that claim they can detect AI-generated music — and, even more controversially, can attribute AI-generated music back to the songs that inspired it from the training data. I imagine you’ve had folks doubt how good your tools are. How do you assure them of their accuracy?
Herault: We made a lot of presentations to the music industry, showing our results. And we also published a paper in the scientific community, so it’s accessible to everyone.
Moussallam: In the first paper we published, we stated that this problem cannot totally be solved by our tool. If anything, we are the most in doubt about our own technology because we did the work — we know exactly how limited it is. This is still an open problem, and it is still easy to bypass our detector.
Herault: That’s why we share data. We want to open the discussion. We explained to the industry that we cannot solve everything, but we can solve part of the problem with these tools.
How feasible is AI music attribution right now?
Moussallam: It’s impossible to know if we’re going to be able to do it. It’s a really fast-moving and exciting field, so we want to dive into that. But this is going to take years, probably. The only thing I’m sure of is that there is nothing on the shelf right now that’s able to do this in a satisfactory manner.
One debate I hear in the music tech world is whether or not this flood of AI-generated music is going to be a serious burden for streaming services, in terms of servers and storage space. Someone at a rival music streamer told me they don’t think this is a real concern. Others believe it is. What is your take?
Herault: This is a serious topic for us, and we already removed some content that was using too much space — because it’s not only an economic issue, it’s also an ecological one. Do we need to store all this content so we all have the same catalog and all have a copy of the same things? It’s really a question worth thinking about, and we already have had some discussion to remove some content, especially because a lot of content is not listened to at all.
That sounds like an issue that streaming services are going to have to face, then.
Herault: Not only music streaming, I mean, there’s all kinds of AI-generated content — images, audio, video. It’s everything.
Moussallam: Audio is actually not such a big file to store when you compare it to movies or some images.
I always wonder whether younger generations will care whether or not their music came from an AI model or a person. It’s hard for us to wrap our heads around, but if these kids are raised on AI content, maybe their listening habits are different from ours. Have you taken this into account when deciding how you want to treat AI music on Deezer?
Moussallam: It’s an amazing research question, but it’s one which doesn’t have an answer yet. I think a lot of people are wondering if we are seeing a shift in terms of the relationship that people have to music. It also raises the question of, why do people love music in the first place? Do you need to have a human relationship, either directly or indirectly, to music to attach meaning to it? I believe a lot of people need that. But I am not sure what people are going to do in the future.
