Privacy, training, and the household graph

Muse asks U.S. adults to put an agent on WhatsApp, with training opt-out and a Secure VM Meta says will be encrypted later. Reuters reported mishandled access in tests. Meta AI, a different product, already lives in the same family’s Instagram. Filing those as one privacy story is how you get a bad recommendation.

Labs train. Wearables record. Kitchen displays sync to a vendor cloud. Co-parenting apps are designed to be produced in court. None of that is automatically disqualifying. Pretending they are the same risk is.

Questions that belong on every record

Who is the user — a person, or a household with permissions? Can a child be a user, a subject, or neither? Is audio stored, and where? Is there a human review queue? Can the other adult see the log? If the vendor cannot answer those from a public page, the evidence grade should show it.

Apple is a ceiling, not a household

On-device processing is the cleanest story in the U.S. pocket. It still does not create a shared week. Privacy-max parents will pick Apple and then run Cozi or a paper calendar anyway. That pairing is the system. It is not a contradiction.