I'll try to go quickly.
First off, I don't see what is taking place as theft—it's not being stolen—and it is not necessarily the case that it can be used freely. What's taking place, certainly in an AI context, is an interest in the underlying data itself. It's not being republished. It's not being commercialized in any way where people are taking that original source and trying to make that original available to them. They're interested in the underlying data itself. To the extent to which we see even these systems learning and then generating results that are similar, musicians and creators have been doing that forever. Many of their perspectives and their approach is based on seeing what others are doing and trying to adapt and put their own unique spin on it.
In that sense, this is not theft in the way we would conventionally think of theft, and it is not free use. The uses—at least what we've seen from the courts so far—are consistent with what fair use or fair dealing would be, and frankly, because you mentioned the EU, the EU has a text and data mining exception that specifically seeks to ensure there are certain kinds of uses that are appropriate for that kind of informational analysis purpose.
That's not to say we don't need to be thinking about copyright in this context, but I'm not wholly convinced the starting point here is to say that what's happening is that content is being stolen and then used with no limits at all. There are clear limits in the law right now under fair dealing and fair use, and the kinds of uses are not what we would conventionally see when we think of that.