The Atlantic reported that "four giant datasets of songs” — the largest containing 12 million songs — are being shared in the AI music development community, including data from many Canadian acts, including Luna Li, Lunice, Valley and others.
A new data leak is showing artists if their music has been used to train AI models.
Earlier this week, The Atlantic published an AI model-training database that compiles "four giant datasets of songs that are being shared within the AI-development community” — the two largest containing 12 million tracks.
It’s not new information that AI-generated databases have been scraping music to train AI music-generating models without proper licensing, permission or compensation for artists. But now artists can see if their music was used for training.
Included in the database are songs from Canadian acts like Backxwash, Tre Mission, Lunice, Luna Li, Leith Ross, Valley and Nemahsis, among others. The publication’s tracker highlights exactly which songs are included in the two biggest databases.
advertisement
A handful of artists have spoken out after finding their music on the database.
Rapper and record producer Tre Mission, who has 20 songs on the database, said: “I’m 100% sure I never consented to this 🤔Anyone who knows me, knows I HATE the use of AI in music, so this is very disappointing.”
In response to artist Sophia Hjkleightninezeroone — who kicked off the conversation on X about the publication’s database and has 138 entries — Montreal DJ and record producer Lunice discovered that at least 93 of his own works were included. “That’s wild… Thanks for bringing it up,” he said.
Some artists put their discovery quite simply: “I dont like this,” Polaris Prize-winning rapper Backxwash wrote on X.
Following the virality of her post, Hjkleightninezeroone has requested that anyone with connections to lawyers in the music industry get in touch with her.
The database also includes songs by some of the country's biggest artists.
Drake has almost 800 entries, while The Weeknd notches 460 and Justin Bieber has 458. Hall of Famer Nelly Furtado has 307, whilst Alanis Morissette, Tate McRae and Sarah McLachlan have 259, 210 and 183 entries each, respectively.
advertisement
Reportedly, the four datasets have been downloaded thousands of times, but there isn’t much information on which companies have used them.
“Companies often claim to use only content that is freely available online, but the datasets reveal the quantity of downloadable music that developers can access even though it is not supposed to be free,” claimed writer Alex Reisner.
One of the biggest AI music platforms is Suno. Per The Atlantic, in a 2024 court filing, the music generator detailed that it trained its models on "essentially all music files of reasonable quality" available for download online. Music created on the platform has even made its way onto the charts.
Last December, HAVEN.’s “I Run.” debuted at No. 59 on the Billboard Canadian Hot 100, later climbing to No. 45. Before its charting debut, the song drummed up a notable amount of controversy after the electronic duo teased the track on TikTok with a female vocalist that resembled British singer Jorja Smith. After Smith said that it wasn’t her voice, the duo replaced the AI-generated vocals with real ones from up-and-coming singer Kaitlin Aragon, who had posted a cover of the song.
advertisement
Aside from Suno, The Atlantic reported that Google trained a model on 44 million songs in 2022, while two years prior, OpenAI scraped 1.2 million songs from online to train its Jukebox platform, which generates full-length songs with vocals and instrumentation from scratch.
