Apple begins beta testing Apple Watch AI conversation rewind and recap features

Apple has started beta testing Audio Intelligence, its AI listening feature set for the Apple Watch, according to TechCrunch. The rollout begins with developers through the iOS 27.3 update, roughly two months after the capabilities were unveiled at Apple's September "Surprise and Shine" hardware event. A public beta and wider release are expected to follow.

Four opt-in features for newer watches

The package consists of four features that users must deliberately turn on, and it is available on the Apple Watch Series 12 and Apple Watch Ultra 4. Two are aimed at environmental awareness: sound recognition, which alerts deaf and hard-of-hearing users to noises such as alarms, sirens and doorbells, and song identification, which leans on Shazam to name music playing nearby.

The remaining two — Live Rewind and Siri Recaps — deal with human conversation, and they are the ones most likely to draw scrutiny, given the criticism aimed at other AI wearables such as Meta's AI glasses.

How Live Rewind works

Live Rewind targets the common problem of missing something that was just said. Double-pressing the digital crown pulls up a text snippet covering roughly the previous 15 seconds of speech — useful when someone rattles off directions or recites a long list of restaurant specials. A chime and an on-watch animation play at the same time, so people nearby can tell that something was captured.

The snippet disappears permanently after 30 seconds unless the user saves it to the standalone Siri app that arrived with iOS 27. An "Ask Siri" button beneath the text lets users follow up on anything unclear, such as an unfamiliar term.

Siri Recaps summarizes without transcripts

Siri Recaps goes further by producing summary notes of conversations. TechCrunch reports that Apple does not identify who is speaking, store raw audio, or generate word-for-word transcripts. Instead, each conversation is reduced to a summary sentence plus bullet points covering practical details such as dates, deadlines and pricing. Apple positions it as a stand-in for jotting down notes by hand, deliberately stripped of tone and emotion, for situations like work meetings, parent-teacher conferences or training sessions.

The feature pauses when headphones are playing music or running noise cancellation. Users can trigger recaps manually or on a schedule — for instance, only during work hours and never at night — and can stop them at any time from Control Center. Saved recaps can be moved to the Notes or Journal apps; recaps that are not saved are deleted within seven days. Users are prompted to review them at the end of each day and again when they unlock the watch in the morning. Apple also says sensitive material, including government identifiers, financial and authentication data, certain personal identifiers and language related to self-harm, is excluded by design.

The processing pipeline

The features combine the watch's S11 chip, its microphone, Apple Intelligence models running on the iPhone, and Apple's Private Cloud Compute. Lightweight models on the S11 first determine whether the wearer is part of an ongoing conversation. Audio is buffered in a secure enclave — a hardware-isolated memory area the rest of the operating system cannot reach — and is reachable only by the paired iPhone through a secure pairing mechanism.

On the iPhone, the audio is converted into condensed text with filler words stripped out. That text is encrypted and sent to Private Cloud Compute, which produces the recap and returns it to the phone; Apple says the recap and any intermediate summaries are deleted from its cloud immediately afterwards. Live Rewind snippets are likewise processed on the iPhone. Apple claims battery life on both devices is unaffected.

Why it matters

Ambient-recording wearables have earned a poor reputation, and Apple is betting that aggressive deletion defaults, visible and audible cues, and heavy on-device processing can make conversation AI socially acceptable. If the approach holds up, the pattern — transient buffers, no raw audio retention, summaries only — could become a template for how consumer AI handles the most sensitive data it can encounter: other people's speech. The developer beta is also the first real-world test of Private Cloud Compute inside a continuous listening product, which will draw attention from privacy researchers and from rivals weighing their own always-on assistants.