Actually, just consider that illegality is defined by other humans, often arbitrarily.

A system that protects the privacy of users who have "American" free speech is simultaneously a system that protects "criminals" who violate blasphemy, royal-insult, obscenity, or any number of other laws in other countries.

I'm not sure there's any way to formalize it really. It follows pretty much in one step from the axiom "your computer doesn't know if you're scum or not". And if we had the strong AI needed to change that, we'd just be letting a computer invade our privacy and judge us instead of a human or government.

Every time the AI checks and finds out that you are not a scum, its memory could be wiped so that it retains no knowledge of anything private it learned about you during the check.

That only solves one minor issue with loss of privacy.