If you are a musician or creator relying on AI tools, the latest intelligence from The Atlantic suggests your work is likely already being harvested without consent. The publication’s AI Watchdog project, which has recently gained significant traction among creatives, indicates that the scale of unauthorised data scraping is far more extensive than previously suspected.
While independent artists previously filed class-action lawsuits against platforms like Suno and Udio in late 2025 for infringing on their rights, the scope of the issue appears to have been underestimated.
Launched in September 2025, the AI Watchdog tool identified a vast repository of unauthorised training material. This included over 7.5 million books, 81 million research articles, 15 million YouTube videos, and written transcripts from tens of thousands of films and television programmes.
Massive music datasets exposed
Following earlier investigations into how AI models infiltrate platforms like YouTube Shorts, writer Alex Reisner recently uncovered specific repositories dedicated to audio. He discovered that giant datasets containing millions of tracks are being actively shared within AI development communities.
Reisner highlighted two particularly large collections: one containing 12 million songs and another with 9 million. Two smaller datasets each held more than 100,000 tracks. The largest of these collections, he noted, would take 91 years to listen to in its entirety.
These datasets have reportedly been downloaded thousands of times. Google is among the entities confirmed to have downloaded a smaller collection, sourcing material from the Free Music Archive. Similarly, Stability AI has admitted to training its Stable Audio Open system on this same archive.
Artists confront the reality
The revelation has sparked immediate backlash from the creator community. Producer DJ Sabrina The Teen DJ addressed the irony of her work being dismissed as “AI slop” once those very models began generating it.
to everyone who thought my music sounded like ai slop, did you ever think it was because Suno was using a dataset that contained 22 of my songs?
it’s funny how there were no accusations of my music sounding like ai slop until these datasets started getting used to generate slop
– DJSabrinaTheTeenDJ (@DJSTTDJ) June 18, 2026
Other creators have found their specific discographies catalogued within these illicit databases. A user searching for the catalogue of artist Quedeca found 295 instances of their work across eight different datasets, ranging from full releases to lyric snippets sourced from Genius. Quedeca responded with a single, sarcastic: “Yayyyyy!”
Smaller artists are not immune. Breakcore producer sophia_hjkl, who has a combined following of around 10,000, revealed that Suno and Udio had utilised 138 of their songs across two datasets. This represented nearly their entire catalogue of work released between 2017 and 2024.
the atlantic just published a searchable database of the music used by suno and udio. they used *one hundred and thirty eight* of my songs across two of their datasets. this is almost my entire catalogue of music. it’s just about everything i’ve released from 2017 to 2024.
– Sophiaaaahjkl;8901 (@sophia_hjkl) June 18, 2026
While major stars have remained silent, Vince Valholla, head of Valholla Records, shared a damning update. He stated that over 100 songs from his label’s catalogue had been used to train AI models.
Late last night I found out over 100+ songs from our catalog were used to train AI models. Thanks to The Atlantic, they leaked a database of millions of songs that have been used by the biggest AI music companies like Udio and Suno.
To be honest, until the major labels go through their lawsuits, there’s no way for artists or labels to fight back. They literally scraped the best songs from our catalogue. I’m sick.
– Vince Valholla (@VinceValholla) June 19, 2026
Official investigations launched
APRA AMCOS, Australia’s official music copyright agency, has announced an investigation into the findings. The organisation, which represents artists ranging from Nick Cave to Kylie Minogue, released a statement titled “PROOF OF THEFT” condemning the practice.
Following Australia’s official rejection of copyright exceptions for AI platforms in October, APRA AMCOS is particularly aggrieved. Chief Executive Dean Ormston emphasised that these works are not bargaining chips but the life’s work of songwriters.
No permission, no licence, no payment. These are not bargaining chips, they are the life’s work of Australian and New Zealand songwriters.
The discovery has also mobilised fans on social media. Users are frantically searching for ways to protect their favourite artists, with some noting that even niche or underground creators are being targeted.
Scott The Pisces, a small British producer, confirmed that ten of his own tracks had been found in the datasets, despite his lack of fame.
Key takeaways
- Massive scale of theft: The AI Watchdog project has identified datasets containing over 12 million songs, alongside millions of other copyrighted works, used to train AI models without permission.
- Targeting all levels: While major labels are the primary targets of legal action, independent and underground artists report their entire discographies being scraped and uploaded to public repositories.
- Corporate admission: Major tech firms including Google and Stability AI have confirmed training on unauthorised archives like the Free Music Archive.
- Limited recourse: As noted by label head Vince Valholla, individual artists currently lack the leverage to fight back until major copyright holders initiate large-scale litigation.
