← All privacy impact assessments

PIA: Knowing when someone has finished dictating

Status: Assessed before the code, 15 October 2026; implementation authorised. Nothing reaches the new check until a member has acknowledged the privacy policy version that describes it. Written: 15 October 2026 · Last revised: 15 October 2026 Nothing here is legal advice. Children's Code citations are to the exposure draft; section numbers may move on registration (due 10 Dec 2026).

1. Purpose and scope

When someone dictates into the assistant, into Add several (tasks or events), or into the box that describes a routine, the phone has to decide when they have stopped talking, because each of those boxes acts when they stop: the assistant sends the request, Add several sorts it, the routine box fills itself in. Until now the phone decided by a fixed wait after the last words, lengthened or shortened by a list of English words. People pause while they think, and a request that is cut off is sent half-finished.

This change lets the phone ask TypeSafe's Jev, at most twice per dictated message, whether the words so far sound finished. A confident "finished" lets the microphone stop sooner; a confident "not finished" makes it wait longer; anything else, or no answer, leaves the decision to the word list, exactly as before.

In scope: the words a member has dictated so far, sent while they are still dictating, in those four boxes on a member's own phone. Out of scope: group chat (its messages are encrypted for the group, and its dictation keeps the word list); the shared display (its accounts are never offered the check); the phone's own speech recognition, which is unchanged and described by the microphone's own notice; and what happens to the words after the member finishes, which each box's own assessment already covers.

2. Data involved

Kind Whose Children's? Sensitive? Free text?
The words dictated so far in one message, with group names replaced The member dictating; can mention anyone Can mention a child Can mention health, as anything a person says can Yes
What the member is doing: asking the assistant, listing things to add, or describing a routine The member No No No

Not sent: audio (it never leaves the phone or the platform's speech service), the group's records, its history, its roster, any identifier, the time, or anything another member wrote. Not stored by us: the words. Our server passes them to Jev and keeps nothing but counts and an outcome.

Strictly necessary? The words are the only thing the question is about. Without them there is nothing to judge; the word list is the alternative and stays the fallback.

3. Flows and locations

Data Device Our server (Australia) TypeSafe (United States)
Words so far Held in the box Received, names replaced, forwarded, not stored Judged once per question, answer returned
The answer (a probability) Used to set the wait Returned to the phone —
Usage — A log line with counts, outcome and latency; spend recorded against the member's allowance and the platform's ceiling —

The words leave the phone before the member has finished. In the boxes in scope, the same words, and usually more, would reach our server and the same processor moments later when the member stops. The new exposure is narrower than it sounds but real: words from a dictation the member abandons (cancels, types over, leaves the screen) may already have been sent once or twice. TypeSafe states it does not train on inputs. Logs carry no words, group or user identifiers.

4. Legal basis and consents

No new consent purpose. The check reuses the conditions each box already applies before any words reach a model: an adult or a 16- or 17-year-old who has said yes to AI features themselves, a group that is not Device-only, a person's own account (never the shared display or a profile account), and the assistant and overseas-processing choices the member already made. In addition, the member must have acknowledged the privacy policy version that says dictated words may be checked while they are still talking. Our server reads that acknowledgement on every request. A member who has not acknowledged it gets the word list and no request is made to TypeSafe on their behalf.

5. Risks and mitigations

Risk Who is harmed Mitigation Structural or conventional? Residual
A parent starts dictating something private, thinks better of it and cancels, but the first half already reached the United States The parent, anyone named Names replaced before sending; at most two checks per message; no storage by us; the policy says so Name replacement and the two-check limit are code; the disclosure is copy Accepted (§8)
A confident wrong "finished" sends half a request to the assistant or sorts half a list The member Measured before building: no confident wrong "finished" in 192 answers; the threshold is strict; the request is still reviewable where it was before Code threshold; measured once on invented words Accepted (§8)
A child's account dictates and its words reach TypeSafe A child Child accounts and 16- or 17-year-olds without their own yes are refused on our server before anything is sent Structural (server gate per request) None known
A Device-only group's words leave the household The group Refused on our server by the group's privacy level; the phone does not ask Structural None known
A member on an older notice is checked without being told The member Acknowledgement of the describing version is read on our server per request Structural None known

6. Structural versus conventional

Structural: the server gates (membership, child account, young member's own yes, managed account, Device-only, acknowledgement of the describing version, spend limits) run before any request to TypeSafe; names are replaced by our server from its own roster read, and a failed roster read sends nothing; the phone makes at most two checks per message; the answer is a number and carries no text back.

Conventional: that the four boxes, and only those, offer the check (a phone-side choice of which screens pass the check in); that the words are not stored (no code writes them; a test scans the route's log lines for content); that the policy describes the change (copy, pinned by tests on the policy text).

7. Best interests of the child

Children's accounts are refused before anything is sent, so the check never handles a child's own dictation. A parent's dictation can mention a child, as any request to these boxes already can; this change sends those words seconds earlier and, for an abandoned message, sends words that otherwise would not have gone. We judged that a microphone that stops sending half-requests is worth that narrow addition for adults, with names replaced and nothing kept.

8. Residual risk we accept

9. Review trigger

This assessment is re-read if the check is offered in group chat or on the shared display, if it ever sends anything beyond the words so far, if the number of checks per message rises, if the words are stored anywhere, or if the threshold is loosened.

10. Checklist against exposure draft s 38(2) (verified 2026-09-08)

s 38(2) Requires Answered in
(a) nature, scope, context, flow and purposes §§1–3
(b) why collection is strictly necessary §2
(c) lawful and fair means §4
(d) best interests of the child §7
(e) how we comply with the Code and the Act §§4, 6
(f) risk of harm to children §5

11. Sources