LLMs work on clear text inputs

But the setup is that Apple doesn't know which cleartexts currently being processed are associated with which user meaning that even live surveillance can't work without surveilling everybody, making any such program very quickly discoverable. Read the section about non-targetability in the link.

Apple deserves credit for correctly analyzing their threat model and designing their system accordingly.

I’m willing to concede that Apple’s system is the best designed of the bunch. I’m not willing to call it “private”, however, if it processes unencrypted inputs in the jurisdiction of a nation state with pervasive government surveillance.

I agree from a technical perspective actually. What this system does to counteract that is use a social strategy. The government couldn't break "this specific system" without a lot of people knowing, and that makes it very likely that they would be discovered. And not in a vague way but in a very specific one.

Would this work in some authoritarian countries? No and Apple doesn't care. Would it work in a western one? Maybe.