The Atlantic created a searchable database of music used to train AI

Atlantic reporter Alex Reisner revealed recently four datasets music used in teaching Types of AI and he made them full search to people. Two of the sets are the largest at 12 million and 9 million. The other two are much smaller, but still represent the most educational data at over 100,000 songs each.

According to Reisner, these sets were downloaded many times and, although it is impossible to know who used them, Google and Stability they have all confirmed their presence in research papers. Some of the sources, such as Free Music Archive dataset, is free for personal use but requires a commercial license.

Although the datasets are freely available online in theory, using them as training data is not as simple as downloading a ZIP file and feeding it to an AI model. As Reisner explains:

Three of my findings are shared as a list of music links on YouTube or Spotify. AI developers stream real-time audio using tools that work, some of which allow creators to bypass ads, ads, and other channels that can get revenue or subscribers to creators. Such materials violate the platform’s terms of service.