But why is it tinfoil-tier?

It's basically common knowledge that if you turn on "hear me say OK Google" that it needs to always be listening to hear "OK Google". That's a green light to start parsing everything said all the time, which will be turned around for advertising. Because that's how Google makes money.

Expecting Google to use data to advertise isn't tinfoil?

The tinfoil part is the parsing.

Actual nlp parsing all audio, or even a tiny subset of devices, is far beyond Google's capabilities. It's not a trivial task to process.

Your phone locally has some extremely basic recognition for "ok google", after which selective actual nlp parsing takes place.

But detecting “oh, this is voice” is easy. Recording the time where that happens is also easy. Knowing when people are chatting around the phone, and roughly what the fundamental frequency of their voice is, could make $0.001 per person. At Google's scale, that's worth it.

So long as they don't get caught, anyway, because that's all sorts of illegal. Especially at Google's scale. (I don't think Google does this… probably.)

What's you're saying is possible and what we're talking about are two entirely different things though.

The fundamental frequency of the voice doesn't tell you what words they were saying.