When Zuck and big tech issues a denial you will notice they always say "we're not recording you" --- yeah cuz they are using on-device transcription to send back keywords!
That sounds a lot of computing power to record and use language detection on everything the device hears, not to mention storage and bandwidth, that many actions should leave a trail somewhere.
Not if it’s just passively listening device side and sending keywords as they come. Doesn’t necessarily have to record 24/7 and process
I'm sure this is true but I have been wondering, how did they find a way without microphone permission?