AI Music Training Datasets Include Over 21 Million Copyrighted Songs

Recent findings have identified that millions of copyrighted songs are present in datasets used to train artificial intelligence music models. This development is notable due to the scale of music included and the involvement of multiple technology companies in AI research.

An investigation by The Atlantic found that four separate datasets used for AI music training collectively contain over 21 million copyrighted recordings. These datasets are being circulated among AI developers and include works from a wide range of artists.

The largest of these collections, LAION-DISCO-12M, holds about 12.6 million tracks sourced from.

YouTube links, while another, Sleeping-DISCO-9M, contains around 9 million tracks obtained from commercial music. Two additional datasets each contain more than 100,000 recordings, further expanding the scope of music available for AI training.

Music by internationally recognized artists such as Taylor Swift, Bad Bunny, Billie Eilish, Nirvana, Pearl Jam, the Beatles, Nick Cave, and Kylie Minogue is included in these datasets. The presence of such artists highlights the broad range of copyrighted material involved.

What the numbers show

• Over 21 million copyrighted songs are present in four datasets
• LAION-DISCO-12M contains approximately 12.6 million tracks
• Sleeping-DISCO-9M holds around 9 million tracks scraped from commercial music
• Two smaller datasets each have more than 100,000 recordings
• One musician found 71% of his discography included in the datasets

Google and Stability AI have confirmed in their published research that they used the Free Music Archive dataset, which is one of the collections identified in the investigation. However, the inclusion of a song in these datasets does not automatically confirm its use in AI model training, as companies may exclude certain works during the process.

The Atlantic has introduced a publicly accessible tool called the “AI Watchdog,” which allows artists to search for their songs within the four datasets. This tool provides musicians with a way to check if their work has been included in these large-scale collections.

One independent musician based in Nashville used the AI Watchdog tool and discovered that 48 of his songs, representing 71% of his total discography, appeared in the datasets. This example illustrates how individual artists can be affected by the inclusion of their music in AI training resources.

The datasets were compiled from various sources, including YouTube and commercial music platforms. The wide range of music in these collections demonstrates the extensive reach of AI training efforts in the music sector.

* This article is based on publicly available information at the time of writing.