Did some research on this a bit ago -- I believe the emerging CS term for this is "undocumented wake words." Phrases like "hey Google" are wake words (they activate your device's listening function), and the ones you know about are documented. Naturally, undocumented wake words would be ways of activating that function that you don't know about. I don't think there's a lot of literature on them, and no direct evidence, despite some papers on false activations, how much of a conversation is recorded when listening is deliberately activated, etc. This research was supposed to culminate in an experiment of my own on the topic, but my partner fucked it up, and I'm totally not mad and totally don't think Google got to him.
I could see a few ways undocumented wake words could work (actively recording literally all the time doesn't seem likely):
It triggers on brand names.
It triggers on topic keywords ("plumber," "bike," "dog").
It doesn't trigger the listen/record/analyze function at all (like documented wake words). Instead, it counts how often keywords in one of the above categories are uttered and then sends the count back in lieu of a recording. This would be sloppier, but probably easier to design and hide, and if someone is constantly saying words you think cat owners use, you can guess they're a cat owner.
There are at least three other explanations for the phenomenon of getting an ad/video recommendation/etc. seemingly right after you speak about it:
You're getting a hyper-targeted ad based on other things you do, not your conversations being recorded. See the famous example of Target sending an underage woman an ad for maternity items before she even knew she was pregnant. Had she told a friend she thought she was pregnant immediately before receiving that ad, she might have concluded her phone was listening.
You searched for XYZ a few months ago (or emailed or texted about it), forgot, and now mistakenly believe you had only spoke about XYZ once and that must be how Google knows you're interested in it.
Coincidence, maybe made more probable by some broader interest of yours. If you watch a few NFL highlight videos and your location data puts you in Seattle, you're probably going to get a Seahawks ad at some point, and it might be right after you talk about the Seahawks for the first time with a friend.
Did some research on this a bit ago -- I believe the emerging CS term for this is "undocumented wake words." Phrases like "hey Google" are wake words (they activate your device's listening function), and the ones you know about are documented. Naturally, undocumented wake words would be ways of activating that function that you don't know about. I don't think there's a lot of literature on them, and no direct evidence, despite some papers on false activations, how much of a conversation is recorded when listening is deliberately activated, etc. This research was supposed to culminate in an experiment of my own on the topic, but my partner fucked it up, and I'm totally not mad and totally don't think Google got to him.
I could see a few ways undocumented wake words could work (actively recording literally all the time doesn't seem likely):
There are at least three other explanations for the phenomenon of getting an ad/video recommendation/etc. seemingly right after you speak about it:
deleted by creator
If you're using Chrome they're for sure getting everything.
Minor nitpick but “underage woman” is a weird phrase. She’s a child, not a woman
I'd add the location of the user/device as a fourth explanation. If the feature is enabled, it can promote "relevant" suggestions.