The Atlantic's searchable database of AI music training data

For years the AI music fight ran on inference: labels argued models could only sound this good if they had ingested copyrighted recordings, and developers stayed quiet about their data. The Atlantic's tool collapses that ambiguity by making a previously opaque part of the pipeline publicly examinable ^[1]. The most consequential shift is in who gets named. This is not just two startups: Google and Stability AI are identified as confirmed users that drew tracks from the Free Music Archive ^[4], dragging well-resourced incumbents into the same consent and copyright scrutiny that has dogged Suno and Udio. And the datasets have been downloaded thousands of times, implying a network of developers far wider than the named companies ^[4].

The mechanism here is a gap between what is accessible and what is licensed. The Free Music Archive offers tracks free for personal streaming, but those rights do not extend to commercial AI training ^[2]. Similarly, LAION-DISCO-12M, the largest set at roughly 12 million tracks, was released in November 2024 for stated research purposes yet circulates among commercial developers ^[2]. That gap is what lets more than 21 million tracks become de facto training fuel without anyone formally clearing rights. The money side is already visible: the indie duo The American Dollar alleged that Suno cut their licensing revenue by nearly 80% ^[2], while Suno itself acknowledged training on copyrighted music and leaned on a fair-use defense ^[5].

The sharpest community read came from Reddit's reception, which reframed the story around supply chains rather than morality: if nobody can prove what went into a model, every launch inherits a legal supply-chain problem, and the datasets AI companies treated like exhaust are now being treated by courts like inventory. That framing matters because it ties directly to the litigation timeline. UMG settled with Udio in October 2025 and Warner settled with both Suno and Udio in November 2025, while Sony kept fighting ^[2]. A searchable, song-level record of what is inside these datasets gives the remaining plaintiffs and any new claimants concrete, examinable proof rather than circumstantial argument ^[2], and it explains why fresh evidence is being described as fuel for the copyright fight ^[6].

The reception across X, YouTube, and Reddit converged on one thing: musicians are discovering their own catalogs in the data only because a reporter built the tool. Ed Newton-Rex found recordings he sang on with the King's College Choir, and rapper Open Mike Eagle found 184 of his songs on stream. Named artists reacted the same way in the reporting: Tre Mission said he was certain he never consented, and Lunice did not know his work was included until told ^[3]. The consent problem is concrete and personal, with entries running into the hundreds for individual artists such as Drake (~800), The Weeknd (460), and Justin Bieber (458) ^[3]. YouTube coverage featuring an entertainment attorney also pointed at distributors like DistroKid and TuneCore as a plausible leak path, sharpening the question of how unconsented work ended up in these sets at all.