In theory, couldn't the LLM access the content on your browser and it's cache, rather than interacting with the website directly? Browser automation directly related to user activity (prefetch etc) seems qualitatively different to me. Similarly, refusing to download content or modifying content after it's already in my browser is also qualitatively different. That all seems fair-use-y. I'm not sure there's a technical solution beyond the typical cat/mouse wars... but there is a smell when a datacenter pretends to be a person. That's not a browser.

It could be a personal knowledge management system, but it seems like knowledge management systems should be operating off of things you already have. The research library down the street isn't considered a "personal knowledge management system" in any sense of the term if you know what I mean. If you dispatch an army of minions to take notes on the library's contents that doesn't seem personal. Similarly if you dispatch the army of minions to a bookstore rather than a library. At the very least bring the item into your house/office first. (Libraries are a little different because they are designed for studying and taking notes, it's use of an army of minions aspect)