‘What we're really fighting is people scripting computers to generate thousands of songs and uploading them in batches to overwhelm, flood and replace’: Deezer's Head of Research on that industry-leading AI-filter — and why we need it now more than ever

Deezer has never been backwards about coming forwards regarding the scourge of AI slop infecting our music streaming platforms. Just days ago, the French streaming site announced that over half of all new daily uploads to its site are AI — up from 44% in April and just over 30% at the end of last year. So, things are only getting worse.

Luckily, the platform not only implements its own proprietary filter to find, label and in some cases purge AI tracks from its site; it also recently made the tech available to everyone, irrespective of your chosen streaming platform.

But how was Deezer's own in-house AI filter developed, why is Deezer so keen on tackling AI tracks, and what does the company think the future of music streaming will look like — you know, for actual humans still trying to make it their career?

I spoke with Manuel Moussallam, Deezer's Head of Research, and Jesper Wendel, the company's Head of Communications, to get their takes on all this and more.

Moussallam agreed to speak with me on the morning after the company's summer party, held at a music venue in Paris. As he was "Checking if the hangover is showing on my face", I couldn't help but think it's good to know that proper live gigs and parties still matter to a firm whose success hangs on our love of music.

Identifying an issue (back in 2021)

How long has Deezer's research arm been thinking "Okay, this AI thing is an issue, and we have to do something about it"? Moussallam explains that the site's comprehensive AI filter, now available to anyone who wants it, has been almost five years in the making.

"It all started in scientific conferences back in later 2021 or early 2022, because the generative models that were already working for images and videos were starting to make real progress — like real scientific progress — for music," he says. "That was probably a couple of years before the first commercial services were appearing.

"So, since we participate a lot in these scientific conferences, we saw people from Google and from Meta sharing these foundation models that were starting to work quite well for creating whole pieces of music. Not just snippets of 10 seconds like before, but really whole songs. So I guessed that eventually, if you're able to generate a lot of music, you can also distribute it quite easily — and down the road, it would end up in our catalogs."

"I would say we started working on the project in 2022, first with some complicated methods, because that's unfortunately a reflex for people like me — to try to do AI to capture AI — but then eventually we simplified a lot because we made some mathematical discoveries on how those models work.

"We realized that we didn't actually need a big AI to understand and to detect these pieces. Just by relying on some plain signal processing and mathematical tools, we were able to actually detect these artifacts for most of the models out there.

"Yeah, I think we deployed the first version of the detector in September 2023 — and we communicated the first numbers when we were able to scale it up to the catalog a few months later."

A big problem — and it's only getting worse

I wonder why, given that various rival music streaming sites have implemented 'Transparency Tags' (read: optional AI badges that place the onus on distributors and labels to tell listeners that what they're listening to was not written or performed by a human), Deezer chose to create a comprehensive filter — and then to release it to all? While admirable and commendable, that's a lot of R&D budget on AI, no?

Moussallam laughs. "There are many aspects, but I guess firstly, we did it just because we wanted to see if we were able to do it. On my team, we do a lot of audio analysis. So, knowing was the first motivation."

"But then, I think since Suno, Udio, and all these tools are trained on unlicensed catalogs, I was pretty convinced that eventually we would have to take down all this music, because it would be considered as plagiarism or copyright infringement in some sense — which may be the case if the legal battles end up that way.

"So at some point we thought that, okay, if people agree with us that it's kind of unlawful to train this model with all this music and they ask us to remove all the content from Suno in the catalog, we need to be prepared. We need to be able to do it. So, that was another big motivation.

"It didn't turn out like that — for now — but it may be that one day people tell us, 'OK, everything that was generated with that AI model prior to this date in the future, when people actually have agreements and compensate musicians for their work, all of this needs to be removed'. And we would have to be prepared for that."

Putting information into the hands of people who need to make decisions

I mention that across all of my audio streaming subscriptions (not a flex, I simply review audio kit for a living), I tried the Deezer AI filter, and none of my playlists contained AI, bar one: my Tidal catalog listed 1% AI content — and to know what the pesky tracks are I'd need to switch to Deezer. I ask what the reception has been since the release of the software for all.

"Mostly people are happy with the tool, I think everyone just wants to know," replied Moussallam.

"Ever since we started deploying this detector, we've been talking to major labels, obviously, but also indie labels, sometimes small publishers. All of them just wanted to know. Some of them made some pretty interesting discoveries on how their members were using — or actually not using — these technologies; so overall it's really about putting this information into the hands of people who need to have it to make some decisions.

"Implementing some transparency for our users was also a main motivation for us to flag the content on the website. And then we realized that people had to come to Deezer to check an album, say, to see if it had the label. And we said, 'Okay, we can also create a tool just for you to check — you will see for yourself how contaminated your whole library is on other platforms".

To purge or not to purge, based on AI markers?

Speaking to Moussallam even briefly, it become clear that he's a scientist as well as a musician (among his many accolades is a Master's Degree in Acoustics, Signal processing and Computer science applied to Music, from the Université Pierre et Marie Curie).

So when I ask how this proprietary new filter works to pick out 100% AI-generated content — which I assume is fully text-prompted music — and whether Moussallam and his team are looking to take it further, to stuff that's only 50% AI generated, say, things get more serious.

"It's a very complicated question," Moussallam concedes. "Right now, we can't really say that we only detect fully AI-generated songs, because what we detect is the presence of markers of AI generation in a song, and we detect the strength of these markers in the song.

"When we flag an album, it's because these markers are strong enough for us to be confident that it's been AI generated at some point.

"But we don't know if you wrote the lyrics for yourself, for instance. We also don't know if the song contains very few elements of human creation; that's not something that our detector is able to do right now — but that's the focus of our current research.

"So we really want to be able to boil down to the stem level and be able to make a decision based on a stem by stem case, because obviously we're seeing a lot of hybrid cases, and a lot of artists are also integrating AI tools into their creative process.

"So, these very strong markers of AI, we're starting to see them spread in human creations, and we need to be able to disentangle that from what you mentioned, like purely text-prompted music curation".

Can the markers the filter picks up recognize whether it has flagged an AI vocal, say, or artificially generated instruments — and how about stolen melodies?

"No, no, we can't go that far," admits Moussallam. "I mean, we are working on the stem by stem detection, but right now we detect markers — and we are not so sure where they come from."

He smiles. "I mean, we have ongoing research, but not something that I can publicly share. When it comes to detecting chord progressions or melodic impressions or plagiarism metrics, we are not so interested in that, because it would lead us too far. And I mean, there are so many songs that share only four chords! I think that debate is way, way out of our hands in any case."

What's next at Deezer — and can AI and humans coexist happily in music?

Perhaps surprizingly for someone who's spent nearly five years building the tech to flag and potentially quash AI music tracks, when I ask Moussallam where he thinks Deezer and online music will be in 10 years time — and whether he'd like to see all flagged AI tracks gone from the site — he instead builds a quite compelling case for computation in music.

"Well, the very first piece of music that was written by a computer dates back to 1953 I think, back in the days where computers were actually whole buildings built around them.

"Ever since people built computers, they tried to make music with computers. I think that's the story of the 20th century. I'm confident people are going to make great things with these new creative tools. It enables some stuff, and I guess as long as it's being used by humans to create cool music for other humans to listen to, we are very fine with that.

"What we're really fighting at Deezer is automated music generation, people who script computers to generate songs — I mean thousands of songs — and uploading them in batches to streaming platform to try to overwhelm and flood and basically really replace human creation with automated creation.

"So, I guess I hope music will stay artisan, as we say in French, so something that humans do to have the attention of other humans. That's really what I guess drives our motivation. If the music includes computers, that won't be a such a revolution — computers are fine, numerical instruments are fine — it's just who is using them to create that matters, in the end."

Jesper Wendell is keen to hone in on the fraudulent aspect of AI slop and the payment dilution for for artists, because that's also one of the motivations for stifling the recent development of the AI-in-music game.

"Fraud has always been there, it's just easier now with AI music," he says. "And if you upload thousands of tracks, then you can sort of spread that fraud across more and more tracks — and that's something that we want to get rid of as much as possible".

I mention that from the perspective of the consumer (i.e., me), the key annoyance is tracks making it into our recommended playlists and ultimately usurping streaming royalties — small amounts maybe, but still — which we'd prefer would go to a band that we actually like.

Moussallam emphatically agrees. "That’s really something we wanted to avoid. Anyone listening to AI music without knowing that it's AI music, like just feeding it to people in a recommendation? That's something that was very clear for us from the beginning we wanted to avoid. Because that's an awful feeling, I think for everyone".

Given that Tidal only recently drew its anti-AI line in the sand, Bandcamp's strong and concise stance was set out in January of this year, and Spotify's Verified by Spotify badge certifies that an artist is human but doesn't help filter out the slop from any generated playlists, I think it's safe to say that Deezer is leading the charge here.

Based on Reddit threads I see every day, I think I speak on behalf of thousands of music lovers when I say we're grateful for everything the Deezer team is doing to combat the AI infestation trying to take hold on our streaming sites. Where do we go from here? For now, trust in Moussallam's work.

Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.

The best turntables for all budgets

➡️ Read our full guide to the best turntables
1. Best overall:
Pro-Ject Debut Carbon Evo
2. Best cheap beginner deck:
House of Marley Revolution
3. Best budget Bluetooth deck:
Audio-Technica AT-LP70XBT
4. Best premium Bluetooth deck:
Cambridge Audio Alva TT V2