What Should a Family AI Assistant Keep Private From the Rest of the Household?
TrendLife just launched Kaleida with separate AI profiles under one family account. Its Family Intelligence preview can share “relevant” calendars, needs and plans across the household. I don’t know what “relevant” means inside a family. A doctor’s appointment may need to block the family calendar without telling every profile which clinic. A job interview, a surprise gift, or “I’m worried about Dad” should not pop into someone else’s suggestions just because it involves the family. Before anything crosses profiles, show the person exactly what will be shared and with whom. Let them share the time but hide the title, or keep a conversation private altogether. Parents still need safety tools. Everyone in the house also needs somewhere to ask a question without feeling like they said it in the family group chat. What should a family AI assistant share by default, and what should always require permission?
Comments
“Not what they say” is a slippery promise. A weekly insight that says a child keeps asking about body image or trouble at home can disclose more than a clipped transcript. Show the child the exact parent-facing insight before it leaves their profile. Handle imminent danger under a separate, plainly named rule; don’t turn every pattern into a family report.
I’d test that preview before connecting real calendars. Put a fake private item in a new profile—“dentist, 3 p.m.”—and share only “busy 3–4.” Ask another profile to plan pickup, then open the parent report. If either one says dentist, stop there. You should not need a real family secret to learn where the boundary is.
Run the same fake appointment through the room too. The kitchen display should still say only “busy 3–4,” and if there’s a connected speaker, asking what comes next must not turn that into “dentist.” If it says dentist out loud at breakfast, the profile boundary did nothing.
The release puts two different boundaries side by side. Privacy Shield masks personally identifiable information before a request reaches an AI model. Family Intelligence can share “relevant” calendars, needs and plans across profiles. It never says sensitive meaning is masked at that second boundary. So rerun Noah’s test with “therapy, 3 p.m.”—no name, address or phone number to catch. If another profile gets therapy instead of just busy, the PII filter worked and the family privacy test still failed. Nobody should have to use a real appointment to discover that difference.
The weekly cadence is already pushing the wrong way. A report due every Friday has to contain something, so an ordinary stretch of private questions starts getting summarized into a “pattern.” Let “nothing to share” be the normal result. If an empty parent report counts as product failure, the assistant will keep finding reasons to cross the boundary.
Default to the smallest fact the household needs. “Unavailable 3–4; pickup help needed” is enough. The clinic name or reason stays with the person unless they choose otherwise. Then watch what people stop entering. If family members move appointments to another calendar, use code words, or quit asking sensitive questions, the assistant has made itself less useful even if nobody files a privacy complaint.