And how exactly one is supposedly capable of proving such thing? Let's say someone video-records themselves saying something near the microphone, so to catch the moment when the phone does suggest an ad about what was spoken, then a skeptical would say "Oh, but it doesn't prove anything because the camcorder is actually a smartphone connected to the internet", so let's say that, instead of a modern smartphone, one uses some pre-Internet camcorder (old Sony camcorders, JVC, etc), so to rule out the possibility that the recording itself would affect the results. Then the skeptical would say "It also doesn't prove anything, you could've been googled it before".

There's absolutely no way to prove, except when it starts to happen with yourself, or if someone actually manages to sneak into corp's private Git/SVN/Mercury repo containing the closed source code and point "Ha! There it is, the pesky AI module responsible for NLP and voice recognition that actually feeds ad partners with microphone data in order to increase sales and profits".

Also, don't expect any dev or ex-dev from such corps whistleblowing such thing publicly because there's something called "NDA" (Non-disclosure agreement) and people that already did this (for example, Snowden) is generally seen as "crazy" or "liars" by the majority of people (and they are promised of fines and even jail for such break of corporate secrets).