@stepheneb
On your first par: This is roughly the approach taken by the HathiTrust. As I said, I admire it. But it's limited to the works in its own collection and to the US public domain.

On your second par: Yes, it's fairly important to get this right. I should have said that AI tools can be useful even if they only do the first 80% of a job (say), and leave the rest for humans, or if their results are subject to human review. We already use AI tools this way to create metadata for new publications. We could take the same approach with a new AI tool for determining copyright status.