As I think about it, I'd ask the same regarding the voice-controllable phone in my pocket.
FWIW, I don't trust the devices, because their operations are out of any consumer's ability to control them, but maybe I'm missing something about their technical limitations that would work in my favor.
Practically, it would be unwieldy to have 10,000,000 devices sending audio to 10,000,000 audio decoding and processes 24 hours a day nonstop.
You'd also see a huge bandwidth hit at your ISP, which would certainly kill future adoptions if the word got out.
More information here: http://www.rowetel.com/?page_id=452
This is of course the product of tinfoilhattery at its finest level, until someone does it for real.
Companies that interact (peer) with them would likely see something though, but possibly not as easily as it seems. The average home internet connection probably downloads far more than 12.6 MB of content from AWS hosted services every day. The only question is whether the upload amount would trigger any alarms. I think in most cases not, as it would probably just go a very small amount towards evening those peering connections out, which are likely very heavy in the other direction.
There are many reasons why it doesn't make sense for them to do this, but this isn't one of them.
Edit: To clarify, and put this in perspective, 12 Gbps is 1.5 GB per second, which is less than 127 terabytes a day. Amazon, through AWS in multiple regions, is entirely capable of adding 127 terabytes of storage a day, and already transfers MUCH more than 12 Gbps. This is not impossible, just very improbable.
As given by squarefoot, you can record human voice at 1200 bit/sec = 150 bytes/sec. A day is 86400 seconds, assume people are talking (generously) 10% of the day, so 8640 seconds * 150 bytes = 1.3 megabytes per day uploaded to Amazon.
Does anyone doubt that _the company that runs AWS_ is incapable of dealing with barely a megabyte per device per day?
Also, the definition of "unmodified" and "configured" is very broad in this context. There's no need for a screwdriver or messing with a JTAG port when a well hidden magic packet coming from the Internet can trigger a watchword=off function.
The device has to continuously process audio to listen for the wake word (upon which it does start sending everything to amazon) and a few seconds of the buffer before that so you can just bark out commands rather than waiting for it to acknowledge you.
You can ask Alexa, "Are you always listening?" to get details about it. It's possible that the device doesn't have anything interesting anyway and they were just resisting the disclosure for PR reasons.
True, I can ask, but I don't know if I'd always believe the answer.
I saw no evidence to prove or disprove this and there's nothing to say that the device couldn't be altered after installation.
[edited: grammar]
Even if one didn't know that, the fact that the trigger word is easily changed by the user should indicate that everything is in software.