There's no way this thing has the power to do adequate voice recognition on device for arbitray speakers and queries, even given a limited domain. It's sending everything to Amazon for processing. The only thing it probably recognizes by itself is "Alexa."

Right, so it sends the commands that you give it only after it recognizes a pre-programmed word. It will likely not send up casual conversation surrounding the command, as there would not be a non-nefarious reason to do that.

False positives... But yeah they're probably rare. The real risk is that it could be hacked or national security lettered to listen permanently. The FBI has form in this regard.

I'm guessing it's Android under the covers, and I believe that Google voice recognition is now processed on the device - so it might well be powerful enough to do the voice recognition without the cloud.

I believe it's only limited command processing, and less accurate than what can be done with more powerful computers.

Doesn't mean the NSA or Chinese equivalent won't figure out how to hack Echo to plant a little piece of stealthware on it that records all conversations of suspects in "terrorism related" investigations. Such stealthware might get detected once or twice (though not often-- it won't be ubiquitous), but such detections will easily covered up by forcing Amazon to announce a rare and obscure firmware bug backstory. Actually, come to think of it, with proper contextual targeting, such stealthware wouldn't even need to send full conversations. Just wait for detection of "the meeting" or "the rally" and boom spend off all the info you need to effectively disrupt pro-democracy rallies in Hong Kong.

The NSA could also do the same thing to a phone.

Or a laptop, or an iPad, or anything with a microphone on, really.

Phones, laptops & ipads don't have microphones designed specifically to pick up conversations from the other side of the room...

This very likely uses network speech recognition, so most things (except for the wake-up word) are being streamed up and recognized in Amazon's data center (Siri, SVoice, etc all do this also)

Exactly. Casual conversation won't be sent up; only commands.

SmartTV and other home-automation-with-voice-recognition device manufacturers would like a word with you.

Actually it's the other way around. We (as consumers) should have a word with them.