Artists discovered their songs in the training data through a tool built by one journalist. The revelation hit hard. Beth Crowley posted on Instagram that she saw her own recordings listed alongside her book. Other musicians followed with similar reactions. Some expressed shock. Others voiced resignation.
The Atlantic published four searchable databases last week detailing more than 21 million tracks circulating among AI developers. One dataset alone holds 12 million songs. Listening to every one would consume 91 years. Another contains 9 million. Two smaller collections each exceed 100,000 tracks. The Atlantic assembled them after combing research papers and data-sharing forums. The files have been downloaded thousands of times.
Taylor Swift appears. So do the Beatles, Bad Bunny, Nirvana, Billie Eilish, Pearl Jam, Miles Davis and classical composers. Tens of thousands of lesser-known independent artists sit alongside the stars. New Radicals’ “You Get What You Give” shows up in two of the collections. The scale feels overwhelming. Yet it matches what companies have admitted using in the past.
Google trained models on more than 100,000 songs pulled from the Free Music Archive. That specific set, assembled in 2016 by Swiss researchers at the École Polytechnique Fédérale de Lausanne, contains 106,574 tracks. Most carry Creative Commons licenses. Those licenses demand credit for the artist and bar commercial projects. The Atlantic notes Google referenced the archive for its MusicLM system. Stability AI trained on a 13,874-track subset of the same collection.
Three of the newly highlighted datasets consist of links to YouTube and Spotify content. Developers bypass paywalls and logins with automated tools that violate platform rules. The fourth arrives with actual MP3 files. Companies repeatedly claim they train only on freely available material. These collections show how much supposedly protected music sits ready for download anyway.
Lawsuits tell part of the story. Major labels sued Suno and Udio. At least 12 cases target AI firms for using copyrighted music without permission. Some suits have settled. Universal Music Group reached terms with Udio last year and agreed to launch a licensed generative service in 2026. Warner Music struck similar deals. Independent artists filed class actions too. Country singer Tony Justice sued both platforms. David Woulard led another group alleging stream-ripping from YouTube and scraping of lyrics sites. Copyright Alliance tracked these cases into 2026.
Suno and Udio defend their approach as fair use. They say training does not harm original markets. Rachel Racusen, speaking for Suno, pointed to safeguards against impersonation. The company claims reproductions of training material should not occur. Yet labels presented dozens of generated tracks that sounded strikingly similar to protected hits. One Suno output echoed Michael Jackson’s “Thriller.” Another mimicked Chuck Berry. Courts have yet to deliver final rulings. Appeals loom. Uncertainty hangs over the entire sector.
OpenAI once scraped 1.2 million songs to build Jukebox. The company published its training method but omitted the song list. A spokesperson told The Atlantic the firm has always been transparent about that project. Google directed questions to a blog post stressing it uses only material it has rights to under its terms, partner deals and the law. Neither firm commented directly on the new datasets.
Hessel van Oorschot runs Tribe of Noise, which now operates the Free Music Archive. When he learned Google had used the collection, he sent a letter seeking discussion on consent and payment. The reply cited Google’s privacy policy and spoke of a vibrant content ecosystem. Van Oorschot described it as a clear dismissal. Suing across the Atlantic felt impractical. So he let the matter drop.
Some musicians changed behavior. Benn Jordan, active for more than 25 years, stopped uploading new work. He watched companies scrape his catalog, generate inferior versions and compete in the same market. Jordan built a tool that adds inaudible noise to audio files. The noise confuses AI models without affecting human listeners. Visual artists have tried similar poisoning techniques. Results vary. But the impulse reveals growing frustration.
Industry surveys echo the anger. One poll of music supervisors, filmmakers and advertisers found 97 percent want clear labeling of AI-generated tracks. Nearly half refuse to work with them at all. Concerns about training sources run deep. Respondents worried about stolen themes or phrases showing up in commercial spots. Record of the Day reported those findings late last year.
Legislators have taken notice. In January lawmakers introduced the Transparency and Responsibility for Artificial Intelligence Networks Act, known as the TRAIN Act. The bipartisan bill would let copyright holders access records to learn whether their works trained AI models. The Recording Industry Association of America, the Recording Academy and SAG-AFTRA back it. Supporters argue it ends the black box. Creators could finally check and, if needed, seek accountability. Congresswoman Madeleine Dean’s office announced the House version.
But transparency alone may not fix the economics. AI music generators now produce full tracks in seconds from text prompts. Listeners stream them on the same platforms that once paid artists. Sony discovered 135,000 AI-generated songs attributed to its roster on streaming services earlier this year. Revenue that once flowed to humans now risks evaporation. Settlements between labels and AI firms point toward licensed models. Those deals often favor big catalogs. Independent artists wonder where they fit.
The datasets The Atlantic published change the conversation. They make the abstract concrete. Search your own name. See whether your recordings helped build the machines now competing against you. Derek Clegg, one musician quoted in the report, put it plainly. “It just seems dishonest. It seems like theft. There’s going to have to be a reckoning.”
That reckoning unfolds in courtrooms, in Congress and in the choices musicians make about sharing their art. Some poison their files. Others sue. Many simply withdraw. Meanwhile the models improve. They absorb more data. They generate more songs. The volume keeps rising. And the original creators keep asking the same question. Who asked for any of this?
WebProNews is an iEntry Publication