Taking a guess: ML is just too good. Don't forget, it's also a signal to leave surveillance areas, eg "last seen driving into the desert". Any kind of outlier behavior is PART of the detection model ... that's where probabilistic models gain leverage, in fact ...
I think if “he left surveilled areas for no registered reason” is something that law enforcement observes and considers evidence of criminal activity, we’ll have well and truly lost any semblance of being a free society.
I don't see how parallel construction applies here.
You are vastly overestimating your adversary here.
This kind of statement is hard to argue, as incentive to conceal technological ability is likely baked into the game. So then what's your threat model?