That's a necessary component of how it works: these devices don't do the speech recognition locally. They upload it to their cloud and run it through their mach

That's a necessary component of how it works: these devices don't do the speech recognition locally. They upload it to their cloud and run it through their machine learning network. They're not supposed to upload "everything" you say, only the stuff you say after the wake phrase. The issue is that they're known to wake up and start recording when they're not supposed to, because they're not perfect about recognizing the wake phrase. There's also the fact that they have to be _listening_ all of the time in order to hear the wake phrase, and the security fear is that, due to a bug or malicious software, they could end up recording and uploading everything they hear as well. That's not a problem of the stated intent of the device, it's a problem of the latent capabilities being used improperly. The best way to secure the device is to remove those capabilities.