> Essentially if you copy a lot of copyrighted material into a blob and then apply some sort of destructive compression to it. How destructive would that compression have to be for the copyright no longer to hold? My guess it would have to be a lot.

I imagine the goal is closer to "enough that no one notices we stole it", either in a way that it's not easily discoverable or even when directly analyzed there's enough plausible deniability to scrape by.