It's relatively straightforward to immediately suspend test takers for a semester and expel them on a second infraction.
Administrations won't allow it because they just don't care enough. It's a pain dealing with complaining parents and students.
In any case, cheating has existed since forever. There is nothing much new about cheating in in-class exams now with AI than before without.
I am amenable to this argument (am GP), however I do think this is unique.
AI is transformative here, in toto, in the total effect on cheating, because its the first time you can feasibly "transfer" the question in with a couple muscle-memory taps. I'm no expert, but I assume there's a substantive difference between IDing someone doing 200 thumb taps for ~40 word question versus 2.
(part I was missing personally was that they can easily have a second phone. same principle as my little bathroom-break-cheatsheet in 2005 - can't find what's undeclared, they're going to be averse to patting kids down)