No surprise about SD being trained on copyrighted, watermarked images. Brought this up recently [0] and months ago [1] with the use of Stable Diffusion.
Once again, the generative AI grift continues with companies using it to scrape artists images without their permission, attribution, and especially without their credit and Stability AI was already aware of the implications of doing the equivalent on training on copyrighted music with Dance Diffusion which I said earlier in [0]
They wouldn't dare; hence used public domain music.