The Apple Watch Series 12 and Ultra 4 ship with a set of features Apple groups as Audio Intelligence: Sound Recognition, Music Recognition, Live Rewind and Siri Recap. Two of them require the microphone to be listening continuously.

Live Rewind lets the wearer double-press the Digital Crown and get a transcript of the previous 15 seconds. Siri Recap listens through the day and produces summaries intended, in Apple's framing, to jog the wearer's memory.

Apple has published an unusually specific account of what happens to the audio. The gap worth writing about is not in that account. It is in who the safeguards are for.

What Apple says it does with the sound

  • "Audio is sent to the Secure Exclave on Apple Watch and iPhone for processing and analysis", and "no audio is recorded or stored."
  • "Encrypted, condensed text versions of a conversation are sent to Private Cloud Compute. The text is less than half the length of the original transcript, and distills the conversation to its core meaning."
  • Summaries auto-delete after seven days unless saved to the Siri app, and saved ones can be deleted across synced devices.
  • Summaries do not identify and attribute speakers.
  • The features are opt in, not opt out, with scheduling controls.

Alongside that, the visible safeguards: activating the feature produces an audible chime even when the Watch is silenced or connected to headphones, and the display shows an animation and a microphone indicator.

Taken together this is a careful design. Audio does not leave the device. What leaves is a shortened transcript, encrypted, to infrastructure Apple built specifically so it cannot retain or inspect the contents.

Every one of those protections is aimed at the wearer

Read the list again from the position of the person sitting opposite.

They did not opt in. They cannot see the scheduling controls. The chime sounds on a watch that is probably on the far side of a table, in a room with other noise. The microphone indicator faces the wrist it is strapped to. The seven-day deletion applies to a summary of their words held in an account they do not control. And the assurance that speakers are not attributed is itself an acknowledgement that other people's speech is in there.

Apple's own guidance concedes the point and hands it to the user: "consider those around you where conversations might be private or sensitive."

That is a reasonable thing to ask of a wearer. It is not consent, and in some places consent is what the law requires.

What the law actually says

California Penal Code section 632 applies to anyone who intentionally, "without the consent of all parties to a confidential communication," uses an electronic amplifying or recording device "to eavesdrop upon or record" that communication.

Note the two verbs. A great deal of the commentary about these features turns on whether keeping no audio means there is no recording — but the statute reaches eavesdropping as well as recording, and a device that listens continuously, transcribes, and produces a durable summary is doing something the drafters would recognise even if no waveform survives.

Whether a transcript-only pipeline falls inside that language is a genuine open question, and it is not one Apple's engineering can answer. California is one of roughly a dozen US states requiring all-party consent, and several other countries are stricter still.

The normalisation problem

There is a second-order point worth stating plainly. Every ambient-listening feature shipped so far has come with an argument for why this particular one is fine: it is on-device, it is opt-in, it discards the audio, it deletes after a week. Each of those is true here.

The cumulative effect is still that the default assumption in a room shifts from "nobody is recording" to "somebody might be summarising". Once that shift happens, the per-feature safeguards do not undo it, because the person who has to act on the change is the one who was never asked.

What to do

  • If you wear one, learn the rule where you live, and treat all-party consent states as a reason to keep it off in meetings, clinics and anywhere confidential.
  • Say it out loud. The chime is not a disclosure.
  • Use the scheduling controls rather than leaving it on all day.
  • If you run an organisation, decide now whether always-listening wearables belong in rooms where privileged or regulated conversations happen, because the devices are already on people's wrists.

What is not established

  • Whether transcription without retained audio counts as recording or eavesdropping under all-party consent statutes. No court has ruled on this design.
  • Whether Apple will restrict the features by region.
  • How the Secure Exclave boundary is audited, or by whom.
  • What happens to a summary of a conversation with someone who objects after the fact.