With AI, copyrighted material is often not obvious to the end user, so I don't think it's fair to go after them.
I think it's non-trivial to make the AI company pay per use though, they'd need a way to calculate what percent of the response is from which source. Let them pay at training time with consent from the copyright holder, or just omit it.
The end user is 100% to blame for any copyright violations.
I suspect from purely logistics, AI training is better when it's free to injest all the content it can, and for that freedom it pays in some small royalty amount when that source is cited.
They'd simply pass that cost onto the customer. For universities or enterprises or lawfirms, or whatever, they would either include pre-existing agreements, or pay for blanket access. Whatever terms OpenAI, Anthropic, and Gemini sign with these entities, they can workout the royalties there.
These are all solved problems for every other technology middle man company.
The AI company is really more like a vector search company that brings you relevant content, kind of like Google, but that does not mean the user will use those results for commercial purposes. Google doesn’t pay royalties for displaying your website in search results.