Is AI the final loss of Privacy?

Not only this; you can also use cloud hosted models that execute in a sealed enclave, a la confer.to or tinfoil.sh. I have Lots of Big Thoughts about the relative privacy/security tradeoffs of sealed computing vs local execution but too lazy to write them up. But assuming a reliable implementation, I think those are pretty good options (albeit open weight models and, as you say, a bit behind the frontier labs’).