on device "ai" capabilities actually probably makes this a lot more feasible

With modern tech and modern hardware it's feasible from a "maybe it can do a little bit of processing while the device is actually in use and use so little that it's not grinding everything else to a halt" but it would be immediately obvious to anyone sitting down and testing to see if it does anything weird and even just compiling a list showing how many times a few select keywords showed up to report back the next time it connects to a server anyways could still constitute an illegal wiretap.

Also, even if it manages to only take a persistent 3-4% processor load to run, what happens when every app that's managed to request microphone permissions is running this? You'd get the whole processor monopolized by spyware to the point nothing could run and it would become immediately obvious. That says that if a company were to try it, they'd basically have only a very brief window before legal hammers over the wiretapping came down or everyone else started doing it too and the manufacturers had to start cracking down because it was making their phones unusable.