Here's the short of it. 90% of the paranoia is unfounded conspiracy theory. These devices don't "listen" unless they are given the wake word. Yes, they are always listening in the sense that the microphone is on waiting for your command (though many devices can be set so that they can only be woken up with a button push). The question you want to ask is, "Do they eavesdrop?" and the answer to that would be an emphatic no.
I say 90% (and in reality, it's probably closer to 99%) because sometimes the devices will mistake what they hear for the wake word and start listening/recording.
People fear what they do not understand, and technology is one of the greatest misunderstood subjects. You want reassurance? Disconnect it from wifi. It can record all day long but can't transit. Or better yet, unplug it. No power, no recording.
But honestly, all of that is completely unnecessary. There have been a ton of experts like myself that have torn them down and analyzed every chip on their circuit boards. Not to mention run network packet sniffing tools to see what data (if any) is being sent out to the mothership.
The short of it is that they aren't up to anything nefarious, they don't eavesdrop on you, and the reason real people transcribe what you do say to it is so that they can improve the product. So for example - if you have a southern American accent, the way you say, "What's the weather for today?" is going to sound different from a New Yorker. AI can't tell the difference. They are looking for specific words (within a programmed margin of tolerance) and to improve on that, real people who CAN tell the difference between a NY and a southern accent transcribe what you say so it can go to the engineers who will feed it back to the AI. That way when they hear "weathur" versus "wethuh" the device will learn that both speakers are saying, "weather."
I hope that clears things up for you.