Uncovering Copyrighted Songs in AI Training Datasets: Implications and Ethical Considerations
A significant number of copyrighted songs, totaling over 21 million, have been discovered in several large datasets used for training generative Artificial Intelligence (AI) models. These datasets, identified by journalist Alex Reisner, contain millions of tracks, including songs by popular artists like Taylor Swift, Bad Bunny, and Ed Sheeran. The largest datasets, LAION-DISCO and Sleeping-DISCO, consist of more than 12 million and nine million tracks, respectively, while two smaller datasets each contain over 100,000 tracks. These datasets are utilized to train AI models, particularly music-generating AI systems that create new audio tracks based on text prompts using deep-learning networks.
LAION and Sleeping datasets do not contain actual music files but instead include web links to songs on platforms like YouTube and Spotify. AI developers use scraping tools to download audio files from these links. The other two datasets contain MP3 files sourced from the Free Music Archive (FMA), which have Creative Commons licenses prohibiting commercial use. While the investigation does not provide concrete evidence of major AI companies' involvement, Google and Stability AI were mentioned in research papers acknowledging the use of FMA datasets for model training. Suno and Udio were also cited in the report, operating within the same data-sharing communities where the 21 million songs have been downloaded.
The Atlantic launched The AI Watchdog search tool following the investigation, enabling users to check the content of the uncovered datasets, which include songs, books, research articles, YouTube videos, and writing from films and TV shows. The tool will continue to add more datasets for search. A check revealed that Malaysian artists, including P. Ramlee, Sudirman, Siti Nurhaliza, and indie acts like Carburetor Dung and Yuna, appeared in the datasets. However, the presence of these works in the datasets does not confirm their use in training AI models. The Atlantic noted that companies often use multiple datasets for training, so the absence of a work does not indicate it was not used.
In conclusion, the discovery of copyrighted songs in AI training datasets raises concerns about potential copyright violations and the ethical use of AI technology. The investigation sheds light on the practices of AI developers and the need for transparency and accountability in the development and training of AI models. The AI Watchdog tool provides a valuable resource for monitoring the content used in AI training datasets and promoting responsible AI development practices.