Siri Recap turns a colleague's words into a summary only the wearer can reach

A wearer double-taps the crown while a waitress runs through the lunch specials, and fifteen seconds of her speech becomes text he can ask Siri about or save to the Siri app ^[14][18]. Apple's privacy account of that moment is entirely about him. No audio recording is created or stored, and the transcript is end-to-end encrypted ^[3][16]. She is in the pipeline without any point in it she can touch ^[22].

That pipeline runs through three compute environments before a summary exists: the Secure Exclave on the watch, the on-device speech and language models on the iPhone, and Apple's foundation models in Private Cloud Compute ^[21]. Raw audio is deleted twice on that path, once on the watch after transfer and once on the phone after transcription ^[20]. What persists is the condensed version with a generated title, encrypted and sent back to phone and watch ^[10]. Which makes the load-bearing sentence in Apple's framing, that the features do not create or store audio recordings ^[3], both accurate and slightly beside the point. The durable artifact is text about what somebody else said.

That text does not travel alone. Apple says it also sends Now Playing data, calendar data, high-level location labels such as home, work and school, and point-of-interest categories like grocery store or park, while withholding precise location ^[11]. A safety model on the phone screens the text to omit potentially harmful terms before it leaves ^[9]. Both are reasonable choices, and both are made about a second person's words with no input from that person.

Sound Recognition is the clean case in this set. It notifies the wearer about doorbells, sirens, alarms or a baby crying without sending anything off the watch ^[4], works when the phone is elsewhere, and Apple points it at people with hearing loss ^[19]. It reacts to sound, not to anyone's speech.

Product teams tell themselves users do this with something like Siri Recap: open the setup screen, weigh it, and set a schedule that covers only the hours where nobody would mind ^[6][15]. What users actually do with a setting that has an always-on option ^[6] is take the default and stop thinking about it. Live Rewind and Siri Recap need Siri AI, which arrives in beta later this year in English ^[17], and there isn't yet any retention or usage data to check that assumption against.

The forcing function for anyone shipping ambient capture, or writing a policy about someone else's, is a two-axis grid: whose speech gets processed, and who holds the control. Wearer's speech under wearer control is dictation, settled. Others' speech under their own control is a meeting recorder that announces itself, also settled. Others' speech under wearer-only control is the cell these two features occupy, and the sources describe no outward indicator and no per-conversation exclusion for the other party ^[22].

So the policy line that bites is a device-class line rather than a recording rule, because on Apple's account nothing is recorded ^[3]. If a colleague wants out of a summary, the only mechanism available is the wearer's mouth.