From P. Ramlee to Taylor Swift, thousands of artists found in AI training datasets raising copyright concerns

KUALA LUMPUR, July 13 — More than 21 million copyrighted songs have been found across several massive datasets believed to be used for training generative Artificial Intelligence (AI) models.

The discovery was made in an investigation by The Atlantic journalist Alex Reisner, who uncovered four major music datasets currently circulating within the AI development community and widely used to train AI models.

The datasets contain millions of songs, including tracks by top global artists such as Taylor Swift, Bad Bunny and Ed Sheeran.

According to Reisner, the largest among the four datasets are LAION-DISCO, which contains more than 12 million tracks, and Sleeping-DISCO, which contains more than nine million tracks.

The other two smaller datasets each contain more than 100,000 tracks.

LAION, or Large-scale Artificial Intelligence Open Network, is a German non-profit organisation that creates large open datasets for AI research, while Sleeping-DISCO was created by Sleeping AI, a group of AI researchers that creates and publicly releases large training datasets for AI development.

How are these datasets used?

These datasets are used to train AI models, specifically music-generating AI systems designed to create completely new, fully produced audio tracks from descriptive text inputs, commonly known as prompts.

Instead of relying on human musicians or pre-made loops, these systems use deep-learning networks to analyse and replicate the underlying mathematical patterns of existing audio.

The two largest datasets, LAION and Sleeping, do not contain actual music files but instead include massive lists of web links pointing to songs on platforms such as YouTube and Spotify.

AI developers use specialised scraping tools to automatically download audio files from these links, bypassing standard security layers, user logins and advertisements.

The other two smaller datasets contain actual MP3 files sourced directly from the Free Music Archive (FMA), a popular repository for independent and royalty-free music.

Although these tracks are free to download, their Creative Commons licences prohibit commercial use. Using them to train commercial AI models would violate those terms.

Are major AI companies involved?

While the investigation does not provide definitive legal proof involving every company, Reisner named Google and Stability AI, citing public research papers where their own developers acknowledged using FMA datasets for model training.

AI companies such as Suno and Udio were also mentioned in the report. Although it is not possible to confirm whether the companies downloaded material from the four datasets, the report noted that developers from Suno and Udio operate within the same data-sharing communities and forums where the more than 21 million songs have been downloaded thousands of times.

Reisner also provided examples of tracks generated through Suno that produced almost identical versions of copyrighted songs, including Michael Jackson’s Thriller and Ed Sheeran’s Shape of You.

Meanwhile, in ongoing lawsuits filed by the Recording Industry Association of America (RIAA) against Suno, the company’s executives previously admitted in legal filings that they trained their AI model on “essentially all music files of reasonable quality” available on the open web.

The AI watchdog search engine

Following the investigation, The Atlantic launched a search tool called The AI Watchdog, allowing users to check what material is included in the datasets uncovered.

Apart from songs, the tool also includes more than 7.5 million books, 81 million research articles, 15 million YouTube videos, and writing from tens of thousands of films and television shows.

More datasets will continue to be added to the search engine.

A check by Malay Mail found that at least 1,000 songs by Malaysian artists also appeared in the revealed datasets.

The list includes legendary names such as the late Tan Sri P. Ramlee, Datuk Sudirman, as well as national songstress Datuk Seri Siti Nurhaliza.

Local indie acts and bands were also found in the datasets, including Carburetor Dung, Butterfingers, Hujan, Spooky Wet Dreams and Yuna.

However, the appearance of these names in the datasets does not confirm that their works were used to train AI models.

“The presence of a work in a dataset is not definitive proof that it was used. Companies often use multiple datasets in training, so the absence of a given work is also not proof that it hasn’t been used,” The Atlantic wrote on its AI Watchdog page.