How are millions of copyrighted songs used to train AI?

Question

Accepted Answer

Millions of Copyrighted Songs Utilised in AI Training Datasets

Extensive datasets, containing over 21 million copyrighted songs, are being circulated and used within the artificial intelligence development community to train generative AI models. These models, particularly music-generating systems, are designed to create new audio tracks based on text prompts, by analysing and replicating patterns found in existing music. The investigation identified four primary music datasets, with the largest, LAION-DISCO and Sleeping-DISCO, holding over 12 million and nine million tracks respectively, alongside two smaller datasets each containing more than 100,000 tracks.
Data Acquisition Methods for AI Training

The largest datasets, LAION and Sleeping-DISCO, do not directly store music files but instead comprise vast lists of web links pointing to songs available on platforms such as YouTube and Spotify. AI developers employ specialised scraping tools to automatically download audio files from these links, a process that bypasses standard security measures and user authentication. In contrast, two smaller datasets consist of MP3 files directly sourced from the Free Music Archive (FMA). While these tracks are freely available for download, their associated Creative Commons licenses explicitly prohibit commercial use, raising concerns about potential copyright infringement when used for commercial AI model training.
Involvement of Major AI Companies and Platforms

While definitive legal proof is not provided for every entity, research papers indicate that developers from companies like Google and Stability AI have acknowledged using FMA datasets for model training. Furthermore, companies such as Suno and Udio have been mentioned in relation to these datasets. Although direct confirmation of their material sourcing is not available, developers associated with Suno and Udio operate within data-sharing communities where these datasets have been downloaded extensively. Executives from Suno, in legal filings, have admitted to training their AI model on a broad spectrum of music files of reasonable quality found on the open web.
Discovery of Malaysian Artists' Works in Datasets

An investigation by The Atlantic also revealed the presence of thousands of songs by Malaysian artists within these AI training datasets. This includes works from legendary figures such as the late Tan Sri P. Ramlee, Datuk Sudirman, and Datuk Seri Siti Nurhaliza, as well as contemporary indie acts like Carburetor Dung, Butterfingers, Hujan, Spooky Wet Dreams, and Yuna. However, the inclusion of an artist's work in a dataset does not definitively confirm its use in training AI models, as companies often utilise multiple datasets. Conversely, the absence of a specific work from a dataset also does not guarantee it has not been used in some capacity.

Reading history

Hide →

This history is cleared when the page is closed.

How are millions of copyrighted songs used to train AI?

Loading

Was this article helpful?

Helpful

Not helpful

Report an issue

4

0

0

Share with friends