For the technology specifically covered in this article, all they are doing is reading the muscle activation of existing muscles, and then mapping to the prosthetic's actuators. The "twist" in this story is that sensing is done with ultrasound instead of EMG (better signal to noise if you can throw enough compute at it), and I believe some cool deep learning tricks to map sets of muscle actuations to prosthetic motion "better".

In order for this approach to work, a nervous signal has to travel out of the brain, and at least into the brain stem. In these cases, they actually have to get all the way to a giant amplifier (your muscle) in order to detect it.

When your talking about your brain imagining to flap your wings, unless your brain has mapped your swings to your lats, traps and delts, you're not sending any actual motor signals out of your brain and we can't read it. If it is mapped to your lats, traps and delts, then you're actually moving those muscles right now. If we used a system where we tried using your lats, traps and delts as points to control prothestic wings, then you run into the problem of what do you with your wings when you're not trying to use those muscles as a wing controller. In other words, you'll get a modal system... which are tricky. You'd basically have to switch off parts of 'baseline human' functionality at a time to get your 'human+' stuff.