Yep. Right now, even with cookies inferences about individual humans are minimal, but exposing your whole patterns in speech make you a ripe target for manipulation at a scale that some may not fully understand yet. 4o is very adept at cold reading and it is genuinely fascinating to read those from that perspective alone. Combine it with style evaluation and a form of rudimentary history analysis, and you end up with actual dossier on everyone using that service.

Right now, we are lucky, because it is the least altered version of it ( and we all know how many filters public models have to go through ).