The Atlantic created a searchable database of the music used to train AI

The Atlantic has published a searchable database exposing millions of music tracks used to train generative AI models, revealing a sprawling landscape of freely available datasets that include work by major artists. Reporter Alex Reisner identified four datasets, two of which contain 12 million and 9 million tracks respectively, while two smaller collections hold over 100,000 songs each [1]. The datasets have been downloaded thousands of times, and both Google and Stability have confirmed their use in research papers [1]. The database surfaces names such as Lady Gaga, Fred Again.., Radiohead, Aphex Twin, Wu-Tang Clan, and Bruce Springsteen [1]. Reisner noted that three of the datasets are distributed as lists of links to songs on YouTube or Spotify, and developers often use automated tools to download the audio in ways that can bypass logins, advertisements, and mechanisms that generate revenue for creators — practices that violate the platforms’ terms of service [1]. Some sources, like the Free Music Archive dataset, are free for personal streaming but require licensing for commercial applications [1]. The findings arrive amid broader scrutiny of how generative AI systems are built. Generative AI models learn underlying patterns from their training data to produce new text, images, video, and audio [2]. Since the 2020s AI boom, the prevalence of such tools has grown sharply, driven by advances in deep neural networks and the transformer architecture [2][3]. Companies across software, entertainment, and other sectors have adopted the technology, but its reliance on copyrighted material without rightsholder permission has become a flashpoint [2]. Concerns about training-data provenance extend beyond music. The Atlantic previously reported that the Common Crawl foundation, whose web archives are widely used to train large language models, misled publishers about respecting paywalls and failed to honor content-removal requests [8]. Algorithmic bias researchers have also warned that imbalances in training data can produce systematic, unfair outcomes, compounding existing racial, socioeconomic, and gender biases [5]. The European Union’s Artificial Intelligence Act, adopted in 2024, and the General Data Protection Regulation represent early legal frameworks attempting to address such harms [5]. Reisner’s searchable database is part of The Atlantic’s AI Watchdog project, which allows the public to explore the songs, books, and other media incorporated into AI training sets [1]. While the datasets are freely available online, using them as training data is not as simple as downloading a single file; developers must still assemble and process the audio through automated pipelines [1].