A single dataset, containing 12 million music tracks and reportedly downloaded thousands of times by AI developers, is just one of four massive collections now made searchable by The Atlantic. This new database offers clarity into the sheer scale of music used for AI model development.
AI models are being trained on millions of copyrighted songs to generate new music. Yet, the creators of these original works have largely been excluded from the process and its profits, sparking widespread concern among artists.
The database's publication will likely intensify legal challenges and regulatory scrutiny. It demands greater transparency and compensation mechanisms in AI model training, marking a critical juncture for intellectual property rights.
The Scale of AI's Music Library
The sheer volume of music ingested by AI models is staggering, yet details remain murky. Two compiled datasets reportedly contain 12 million and 9 million tracks, according to The Verge and Music Ally. However, Billboard Canada reports the two largest datasets each hold 12 million tracks. The discrepancy in reported dataset sizes underscores the chaotic and opaque nature of AI training data, making it nearly impossible for rights holders to accurately assess infringement or demand fair compensation.
The compilation reveals a systemic pattern of unauthorized data collection, from colossal datasets to smaller, significant collections. The scale of unauthorized data collection suggests that AI development has largely proceeded without regard for existing intellectual property frameworks, setting the stage for a fundamental clash over digital rights.
Who's Using the Data, and How Widely?
Alex Reisner, The Atlantic's reporter, found that massive song datasets circulate widely within AI development circles. One dataset holds 12 million tracks, another 9 million, according to MusicTech, while Billboard Canada reports the two largest datasets each hold 12 million tracks. The 12-million-track collection alone has been downloaded thousands of times. Even major tech players like Google and Stability AI have downloaded smaller datasets from the Free Music Archive for training, according to MusicTech.
