Dataset ≠ Songs.
While investigating claims that AI developers trained on a 12-million-song dataset, I discovered something I had not expected:
The dataset does not contain 12 million songs.
It contains about 12 million links to songs, along with metadata:
Song titles, artist names, album name, YouTube URLs, duration, views.
It doesn't contain music.
Think of it as a giant list of songs, rather than a dataset of the songs themselves.
That changed my question to, “What was the pipeline?”
What happened between YouTube and the AI models? How were the actual recordings obtained and moved into AI training systems?
That part of the story seems to be less documented than the datasets some claim were used by AI to steal music.
Watch the video here.
Read the article.
Reader Responses
After publishing this short on platforms such as LinkedIn, YouTube, and Instagram, people shared questions, critiques, and alternative perspectives. I've selected a few excerpts and included my responses below. Names have been removed to keep the focus on the ideas rather than the individuals.
Reader: That's a great question! How did these AI companies get a A COPY (aka a mechanical) of these recordings!???
Your Thoughts?
I'm interested in hearing your perspective. Share your thoughts, questions, or alternative viewpoints in the comments below.