Not particularly, because who's to say that "cop AI" doesn't screw over humanity in its attempts to shut down an AI it's deemed "suspicious"? The current usual

Not particularly, because who's to say that "cop AI" doesn't screw over humanity in its attempts to shut down an AI it's deemed "suspicious"? The current usual approach here is to train AIs without specific goals or optimization functions. Instead, they're rewarded by human reactions to their decisions. As such, they effectively don't learn "do this thing", they instead learn "do what the humans want, avoid what they don't" -- and through that, they indirectly learn to "do the thing" (because the humans want it) while avoiding dangerous outcomes (because the humans don't want behaviors that lead to those).