I’m a lifelong musician and producer who has been experimenting a lot with Suno for music creation this year.

It’s been interesting to notice that Suno-generated tracks have all kinds of subtle tempo drift, and the tempo maps look very similar to the irregularities shown in the article, at least from a numerical perspective. I have actually spent a lot of time quantizing AI-generated tracks to the grid (to aid chopping/editing), even creating automations that do it for me now when I ingest Suno tracks into Ableton for follow-up production work.

Unknown to me whether this is A) an emergent “imperfection” fundamental to the model and its training dataset, or B) something Suno intentionally injects to “humanize” their output, but knowing that low-effort Suno-generated tracks are currently flooding music distributors, I think it’s ironic that the same factor the author describes as exemplifying the very humanity of recorded music may also be one of the biggest tells that a song was AI-generated, for music being released today.