It may be freely available for non-commercial works, eg. Photos on Photobucket, internet archive free book archives, etc.
Most everything is on the internet these days, copyrighted or not. I’m sure if I googled enough I could find the entire text of Harry Potter for free. I still haven’t purchased it, and technically it’s not legally freely available. But in training these models I guarantee they didn’t care where the data came from, just that it was data.
I’m against piracy as well for the record, but pretty much everything is available through torrenting and pirate sites at this point, copyright be damned.