What AI training datasets contain millions of copyrighted songs?

Question

Accepted Answer

LAION-DISCO and Sleeping-DISCO are among the datasets found to contain millions of copyrighted songs.

An investigation has uncovered four major datasets circulating within the AI development community that are widely used to train generative AI models, with LAION-DISCO and Sleeping-DISCO containing millions of tracks each. These datasets are instrumental in training music-generating AI systems designed to create new audio tracks from text prompts. The discovery revealed that over 21 million copyrighted songs have been identified across these datasets, raising concerns about intellectual property rights.
LAION-DISCO and Sleeping-DISCO Host Millions of Tracks

The largest of the identified datasets are LAION-DISCO, which comprises over 12 million tracks, and Sleeping-DISCO, containing more than nine million tracks. Two additional smaller datasets were also found, each holding over 100,000 tracks. These datasets are crucial for training sophisticated AI models that can analyse and replicate musical patterns.
Datasets Link to Songs on Popular Platforms and Include MP3 Files

LAION-DISCO and Sleeping-DISCO primarily contain web links directing to songs available on platforms such as YouTube and Spotify, rather than the audio files themselves. AI developers utilise scraping tools to automatically download audio from these links. In contrast, the two smaller datasets consist of actual MP3 files sourced directly from the Free Music Archive (FMA), a repository for independent music.
Major AI Companies May Have Utilised FMA Datasets

While definitive legal proof for every company is not established, public research papers have shown developers from Google and Stability AI acknowledging the use of FMA datasets for model training. Companies like Suno and Udio operate within data-sharing communities where these datasets have been downloaded extensively. Suno's executives have stated in legal filings that their AI model was trained on a vast amount of music files available on the open web.
AI Watchdog Tool Reveals Extensive Song Inclusion

Following the investigation, The Atlantic launched a search tool named The AI Watchdog, enabling users to examine the contents of these uncovered datasets. This tool has revealed the presence of over 21 million copyrighted songs, alongside millions of books, research articles, and videos. The tool's findings indicate that works by a wide range of artists, including those from Malaysia, are present in these collections, although inclusion does not definitively confirm usage in AI training.

Reading history

Hide →

This history is cleared when the page is closed.

What AI training datasets contain millions of copyrighted songs?

Loading

Was this article helpful?

Helpful

Not helpful

Report an issue

1

0

0

Share with friends