Of course the meta data is visible. Its probably more useful than the actual content of the conversation too. I mean from an ML perspective how would you even make features out of conversation that help with CTR ? That too without creeping the users out. I'd imagine its the same reason why meta doesnt (likely) listen in on mobile mics. Why go through the whole shebang of running always on transcription when simple features like who talked to who and at what times are more useful at establishing user similarities.

I'm not making a stance on things, just clarifying the previous comment