Apple’s Watch Series 12 launch yesterday has one headline feature that’s generating a lot more discussion than the usual battery/screen/chip incrementalism: Siri can now listen continuously, in the background, and produce recaps and transcriptions of conversations that happened nearby. You don’t have to say “Hey Siri” first. You don’t have to open an app. The watch just… hears things, and later on you can ask it what was said.

This isn’t a wholly novel idea. Plaud and a handful of other AI-pendant/pin startups have been selling “ambient recording” hardware for a couple of years now, and Apple absorbing the feature into a device tens of millions of people already own is the obvious next step. But it does change the calculus considerably, because a Plaud device is a thing you have to deliberately buy and clip to your shirt. An Apple Watch is a thing that’s already on 100 million-plus wrists, and now, with a software update, all of them potentially become ambient recording devices.

Interestingly, and as usual, there are a bunch of start-ups sat squarely in the path of this particular juggernaut. BlueDot being an obvious example; when they launched on the Apple Watch recently, one of the proposed benefits was not needing additional hardware and marking yourself out as recording. They also said:

“The smartphone made everyone a photographer. Ambient AI is going to make everyone an archivist of their own life.”

Their own life, and all those around them. I don’t think the example of photographs is as relevant: as others around me have noted, this is closer to Meta’s “smart glasses”. The idea of constantly recording activity is a dystopian trope.

The bystander problem

The interesting privacy question here was never really about the wearer. If you want to record yourself, that’s largely your own business. The interesting question is about everyone else in the room, who did not buy this watch, did not install this software update, and who quite possibly have no idea it exists.

Plaud’s own influencer marketing guidelines explicitly rule out showing covert recording in promotional content: no demos of someone wearing the device to secretly capture a conversation, no “gotcha” framing. That’s a company whose entire product is ambient audio capture! They’ve worked out that showing it being used covertly is a reputational grenade. It just doesn’t necessarily stop the line from being crossed in practice, because the hardware doesn’t know or care what its marketing guidelines say, but it tells us how they want this to be shown in normal use.

Apple’s device doesn’t appear to have any notification or visible status indicator when recording. So although you can turn the Recap feature on and off, there is no indication for those people around you that you’re recording. And while it doesn’t generate a full transcription - you get a headline plus summary/actions - I’m not really sure this changes the fundamental privacy angle very much.

The clinic scenario

The example that’s been on my mind most is a UK NHS clinician wearing one of these in clinic. Picture a GP or a specialist nurse who’s adopted the watch because the ambient recap genuinely helps them — more on that below — and now every consultation is, by default, being listened to and transcribed by a third-party AI system.

Special category health data is about as sensitive as data gets under UK GDPR, and clinics already have fairly strict rules about what gets recorded and how. The MHRA’s recent position is that ambient transcription tools of this kind don’t count as a medical device in their own right — they’re not diagnosing anything or making a clinical decision, so they sit outside that specific regulatory regime.

I’ve seen a lot of pushback against ambient AI in this setting online. I think the MHRA categorisation is right, but the question is then, how do you assess the use of this system? In practice, any technology like this is supposed to be assessed under the DCB0160 standard (on which the NHS is currently consulting). Clinical risk occurs in any IT technology, and ambient transcription has already generated notable errors.

Of course, practitioners make errors as well. The issue isn’t so much the error - it’s how you detect it, remediate it, and seek to lower the incidence in the future. With a full reporting AI, such as Nuance in radiology, it’s part of a medical device and you can yellow card the software via the MHRA.

There’s also the more basic question of whose data this even is. A patient discloses something in confidence to a clinician, under a professional duty of confidentiality that both parties understand. They did not consent to that disclosure being processed by an AI model, stored, and potentially used to improve some vendor’s transcription accuracy.

Where this is genuinely useful

Of course, there are real, unglamorous cases for this technology. It’s easy to forget this when focussing in on the bystander-consent argument.

Apple already ships hearing aid functionality on AirPods Pro in several markets — a feature that quietly turned a consumer earbud into an FDA-cleared hearing aid alternative for a fraction of the price of traditional devices. Ambient transcription is a fairly natural extension of that: for someone who is hard of hearing, having a device that can show them what was just said, or replay the last thirty seconds of a fast-moving conversation, is straightforwardly assistive technology, not a novelty.

Additionally, this new AI can listen for various background sounds, and alert the user to something they may have missed hearing. Even if someone is not using assistive hearing technology, these additional alerts could genuinely be life-saving in the right circumstances. Equally, if you are profoundly unable to hear, knowing that you do not need to rely on visual alerts to get the same warnings as everyone else is going to be very comforting.

The same goes for people with ADD/ADHD, or various memory conditions, where the problem isn’t hearing the conversation, it’s holding onto it. Being able to ask “what did my manager just ask me to do” ten minutes after a conversation you were nominally present for but not fully tracking is a real quality-of-life improvement, and for some people a genuine accessibility need rather than a convenience. It’s very easy to think of this technology as being assistive. In that frame, it’s actually a legal obligation for employers to at least consider the use of these tools - and, if well-proven, there may even be an obligation to provide the technology in the first place.

This is the part of the debate that I think gets flattened too easily into “always-on recording bad.” The same capability that makes this creepy in a clinic or a pub is what makes it genuinely liberating for someone who currently has to ask a colleague to repeat themselves three times, or who loses entire meetings to their own working memory. This could be a game-changer for people who struggle with executive function.

A few more scenarios I’m thinking about

A handful of other situations that I think are worth working through, because they don’t all resolve the same way:

  • The pub table. Six friends, one of whom is wearing the watch. The other five never agreed to anything, and there’s no realistic social mechanism for them to object without it being enormously awkward — “can you turn off your always-listening watch” is not a sentence most people have the confidence to say out loud. Is the watch even on? Is it crucial that the watch captures the latest thoughts on that ludicrous display last night?
  • The workplace meeting. Employers are going to have to write policy on this faster than they’d like, in the same way they eventually had to write policy on smartphone cameras a decade ago. Do you ban them from the building? Trust the chime? Assume every meeting is now potentially a transcript sitting on somebody’s personal device, outside any corporate data governance? Is it the same for every employer? Many places ban phones - trading floors being an obvious case.
  • The journalist’s source. Recording laws in the UK are already a slightly awkward patchwork — you can generally record your own conversations without telling the other party, but publishing or processing that recording is where data protection law and professional ethics start to bite. An ambient AI watch doesn’t ask “is this a story” before it starts transcribing; it just runs.
  • The family kitchen. Children in a household where a parent wears one of these are, by definition, growing up with more of their unscripted conversation captured and potentially retained than any generation before them, without anything resembling their own informed consent. In UK law, the covert recording of children has been typically frowned upon - to put it mildly.

Google Glass became a cultural punchline within about eighteen months of launch, largely because the camera was visible and people reacted badly to being potentially recorded by someone’s face-mounted computer in a bar. It’s worth asking what a watch doing something arguably more invasive — capturing audio rather than video, and doing it by default rather than requiring an obvious point-and-shoot gesture — is likely to do.

I think Apple have done a pretty good job with this feature, and their stance on privacy is typically good. I suppose my main worry is not so much what will happen with this feature: the more I think about it, the less I worry about the privacy angle in the general case, although this is highly contextual. Maybe, a more accurate way of putting this is, “I don’t think I have a privacy concern with this technology, but I strongly expect other people I’m around will do”.

But other people will copy Apple, having done much less of the ground work. Is that a good thing? Covert recording is nothing new, the legal situation isn’t really changing, and the processing is (in many ways) nothing new. But like the frogs in the pot, the temperature is slowly rising, and we’re realising suddenly things feel hot.