What is the threat model where this matters? You have to trust the recipient with Signal. The only one I can think of is the case where your recipient is using a compromised fork and is unaware. In this case, talking about the tool and checking with them about what they are using is really the only countermeasure.

Yes, all while he’d have a private chef and a staff that keep him safe.

As a thought experiment: what would have happened if instead of a public health regulation approach, we dealt with restaurant safety by providing a few safe places and advocating everyone go there if they don’t want salmonella or e-coli poisoning. We’d have people ignorant going to the dangerous places, others misinformed or in denial, and a flood of misinformation that food poisoning is either “fine” or there’s no avoiding it anyway so best not to worry.

There’s a difference between a wake word and general purpose speech recognition. A simple wake word can be done in simple hardware on the device, while general purpose speech processing either requires heavy, relatively constant CPU usage, or heavy network traffic to pipe the audio to a server for processing.

Yes, that’s how I use it. It has access to a read only bind mount of my photo directory. The ML doesn’t write exif data to the images, just keeps that in its database.