Voice assistants broadcast private content to the whole room

What it is

A voice assistant's replies are broadcast to the room: is the calendar free, the message content, the medication reminder, the order status — the answer is spoken "to" the asker but reaches everyone within earshot. Sound has no direction and ears have no lids; this is bystander overhearing: everyone inside the output radius is a recipient, and the system has no model of who is listening.

Why it happens

Physics decides half: screens have viewing angles and distance falloff — a glance takes deliberate orientation; sound fills the room and hearing takes no action at all. A medium you cannot look away from gives you no freedom to avoid it.

The product layer decides the other half: voice-first design funnels services that could be private into the open speaker channel — message read-outs, calendar queries, purchase confirmations all happen on an open acoustic channel. Two engineering gaps amplify it. First, speaker recognition is unreliable: A's question gets answered from B's account (A's query met with B's calendar), so account boundaries are porous at the voice layer to begin with. Second, scheduled announcements ignore who is present: the household morning briefing recites the family's schedule on the morning an overnight guest is at the table, who is thereby informed of where everyone will be.

The secondary effect mirrors shared screens but runs hotter: users learn to "wait until alone", and assistant use withers into timers and weather — the voice channel's personal functions exist in name only.

Studying it

• Voice-assistant privacy research (described generically): surveys and audits of smart-speaker users consistently rank "family and housemates overhearing" as the top privacy concern; suppression behaviours — waiting until alone, disabling message read-out — are reliably observed.
• Acoustic measurement: the intelligibility radius of typical home speakers — at what distance and background noise the content remains intelligible. It turns "the overhearing range" from intuition into a measurable quantity, often used as a product evaluation item.
• Variables: the cross of content category with presence composition (which private categories get spoken while bystanders are present), user suppression behaviour after broadcasts, long-term usage shrinkage.

Where it stops holding

• Broadcast is not the original sin: sounds the household wants to share — timers, music, household reminders — are exactly the speaker's value. The harm is account-scoped private content flowing into the open channel; the criterion is "was the audience chosen by the content's owner", the same test as for shared surfaces.
• Deaf and hard-of-hearing households experience the problem inverted: broadcast becomes exclusion rather than leakage — the information reaches the ears but not the person. Voice-channel privacy design must hold two lines at once: "give bystanders less" and "don't drop the intended listener".
• Open-plan homes offer no acoustic room to retreat to: "say it in the other room" is physically unavailable in a studio layout; the spatial structure closes that exit.

Applying it

• Split by audience scope: household-level content may be spoken; account-level content (messages, calendars, orders) is not read aloud by default — either push it as text to the asker's phone, or confirm explicitly before speaking ("read it out?").
• Check who is present before broadcasting: only release personal content when paired devices or voice prints indicate the account owner is alone; otherwise demote automatically to a silent notification.
• Sanitise scheduled announcements: the morning briefing says "something on this morning" at most, details staying on personal devices; never combine a person's name with a specific event in a scheduled announcement.
• How to check: have the system log every event where "spoken content was account-scoped and listeners present exceeded the asker" — each such event is an over-disclosure; compare the event rate before and after deployment.

Related

• Same group: Z6.03.1 Personal information must not appear on shared devices · Z6.03.3 Monitoring inside the home is power-asymmetric
• Nearby: Z6.04 Non-users in the space · Z4.10 Voice, app, and physical control coexisting
• Search terms: overhearing · voice assistant privacy · bystander · smart speaker disclosure