The Atlantic opens searchable database of music used to train AI

Musicians have long suspected that AI companies scraped their life's work without permission. Now, they have the receipts.

The Atlantic reporter Alex Reisner uncovered four massive music datasets currently powering AI training and turned them into a searchable public database under the outlet's AI Watchdog project.

Millions of tracks, hiding behind simple links

These collections are massive. Two of the datasets contain 12 million and 9 million tracks, while two smaller ones hold over 100,000 songs each.

According to Reisner, the files have been downloaded thousands of times. While tracking down every single user is impossible, both Google and Stability have officially confirmed in research papers that they trained on them.

The artist roster spans nearly every genre imaginable, including Lady Gaga, Bruce Springsteen, Fred Again.., Radiohead, Aphex Twin, Wu-Tang Clan, and experimental artist Hainbach.

Some of the sources, such as the Free Music Archive, explicitly state that songs are free for personal streaming but require licensing for commercial use. AI trainers clearly skipped that last part.

Getting the audio into the models took some rule-breaking. Three of the datasets don't host audio files directly; instead, they distribute lists of links to YouTube and Spotify.

Developers then use automated rippers to harvest the tracks. These scrapers bypass user logins, skip platform ads, and strip out the mechanisms creators rely on for ad revenue and subscriber growth, explicitly violating the platforms' terms of service.

Why it matters

For artists, this database is a massive turning point. Proving that an AI model ingested your copyrighted music used to be near impossible, but now creators can search their own names and find tangible evidence.

For the AI industry, the "we only shared links" defense looks increasingly flimsy. Relying on scraping scripts to dodge streaming platform paywalls and terms of service feels less like clever engineering and more like blatant evasion.

Google and Stability acknowledging these datasets in official research papers gives rights holders a direct paper trail. We think this exposure significantly ramps up legal exposure for major AI labs, handing record labels the exact smoking gun they need for their next wave of copyright lawsuits.