there's nothing stopping from Alexa recording content locally and transmitting it with a traditional request once activated.
Also, there's no reason for Amazon to want a constant stream of Alexa data hitting their cloud services. The goal is to mine and extract valuable data about users to build profiles, ideally you'd want to do that on-device and send up the most valuable bits.
These are also capabilities they can silently introduce at anytime for any device.
You can store around 1000 hours of recorded speech in a gigabyte of flash memory that costs $0.25. For most people, that could hold their entire conversational activity for a year. Storage and off-line upload (perhaps during a software update when packets are flying around anyway) isn't difficult.
They can not silently introduce new hardware to these devices. You should look into what kind of hardware these devices actually contain. Try to set up an intelligent, personalized data miner with those constraints.
Steve was not thinking this through. I can batch compress chunks of audio and lazy upload them when I think nobody is around. A listening device could probably even guess when you are not going to be using your internet connection.
Your assertion that "nobody can prove that" is rather absurd.