The problem is with the LLM then training on that data _once_ and then storing it forever and regurgitating it N times in the future without ever crediting the original author.

So far, humans themselves did this, but only for relatively simple information (ratio of rice and water in specific $recipe). You're not gonna send a link to your friend just to see the ratio, you probably remember it off the top of your head.

Unfortunately, the top of an LLMs head is pretty big, and they are fitting almost the entire website's content in there for most websites.

The threshold beyond which it becomes irreproducible for human consumers, and therefore, copyrightable (lot of copyright law has "reasonable" term which refers to this same concept) has now shifted up many many times higher.

Now, IMO:

So far, for stuff that won't fit in someone's head, people were using citations (academia, for example). LLMs should also use citations. That solves the ethical problem pretty much. That the ad ecosystem chose views as the monetisation point and is thus hurt by this is not anyone else's problem. The ad ecosystem can innovate and adjust to the new reality in their own time and with their own effort. I promise most people won't be waiting. Maybe google can charge per LLM citation. Cost Per Citation, you even maintain the acronym :)