Is the watch on the wrist of the person in front of me recording what I say?

You are sitting opposite somebody wearing a smartwatch. You are talking about something you would not want repeated. And at a certain point the thought arrives: is that watch recording?

The technical answer is yes, it can. The useful answer has more to it, and to my mind it ends up somewhere other than where the question would take you: the problem is not whether it can record, because the phone that person has had in their pocket for fifteen years can do that too. The problem is what happens next.

But the facts first, model by model, because there are more legends than information circulating about this.

What a watch’s microphone can do

Apple Watch. It has a native Voice Memos app: you start the recording with a button, you stop it with a button. While it is recording the screen shows a red stop button, the elapsed time and the word “Recording” at the top right. The maximum length is not stated in Apple’s documentation. There is also an orange microphone icon in the status area of the watch face, and Apple writes that it appears when the Apple Watch’s microphone is listening, meaning when you are talking to Siri, recording a memo or using an app that activates the microphone.

On that icon I have to be precise, because it is the point where a popular article risks reassuring you more than the sources allow. Apple documents it as an on-screen status icon, not as a hardware indicator. And the Apple platform security document, the one describing how app permissions are controlled, lists the microphone permission for iOS, iPadOS, macOS and visionOS: watchOS is not there. So the formulation I can defend is this: Apple documents an icon with that function; I have not found a document guaranteeing that it cannot be bypassed on the watch.

Pixel Watch. Native Recorder app, with a stated limit of 30 minutes per recording on the watch. The recording carries on even with the screen off. On first use you choose whether to sync them to the phone or keep them on the watch, and which option is preselected the documentation does not say. While it is recording, “Currently recording” appears among the watch’s notifications. No documented light indicator.

Galaxy Watch. Native Voice Recorder app from the Watch4 series onwards, with three modes, and you record by holding down the button. Maximum length not stated: the Samsung page only says that insufficient space can limit the recording. No recording indicator mentioned.

Garmin. There is a Voice Notes feature on several models, Venu 4, Venu X1, fēnix 9, D2 Air X15 and others, with the option of tagging the GPS position. Maximum length not stated in any of the manuals I read, and no indicator mentioned. Garmin does not publish a single list of the models with a microphone, so I cannot give you one.

Fitbit. The microphone is only on some smartwatches, not on the activity bands, and here the documentation says the clearest thing of all: there is no wake word, the assistant opens by opening the app, and the microphone switches off when the assistant replies, when you close it, or when the screen goes off. For bands like the Charge or the Inspire I have not found spec sheets declaring a microphone, but I have not opened them one by one, so I am leaving that as unverified.

And then there are the watches sold as recorders. Not smartwatches: devices with 8 or 16 gigabytes of internal memory and continuous or voice-activated recording, on sale on marketplaces and through surveillance equipment retailers. Here the sources I have are commercial pages, so they establish only that the category exists on the market. If somebody wants to record you secretly, they do not buy an Apple Watch.

The real hole: on Wear OS I do not know

This is the point at which the article was in danger of telling you something false, so I am setting it apart.

From version 12 onwards Android requires what are called privacy indicators, that is, the little microphone or camera icon in the status bar when an app is using them. It is the mechanism by which you notice on your phone. The Google document describing it, though, does not mention watches or any device form factor, and I have not found Google or Samsung documentation confirming that indicator is present on a Galaxy Watch or a Pixel Watch. From the other direction, Samsung documents the green dot only for phones and tablets.

So I cannot tell you either that an active microphone is visible on Wear OS or that it is not. It is a hole in the sources, and any article giving you a flat answer on this point has invented it. I am trying to verify it on a real watch and I will update.

But there is something that can be said already, and that answers the question the way an ordinary person asks it. Even where the indicator does exist, it is a dot on a watch face three and a half centimetres across, on somebody else’s wrist, turned towards them. At conversation distance you will not see it. The indicator is for the person wearing the watch, not for the person sitting opposite.

Continuous listening, told as it is

Here there is a contradiction that needs explaining rather than picking a side on, and the best source is the European Data Protection Board’s guidelines on voice assistants.

On one hand the Board writes that until the wake word is detected no audio leaves the device, and that the analysis to recognise it happens continuously but exclusively locally. On the other hand, two paragraphs later, it writes that the recognition is probabilistic, that developers tend to set a relatively low threshold so as not to miss intended activations, and that the consequence of a false acceptance can be the sending of audio recordings. And it adds that a second check of the pronunciation happens server-side, which is to say after the audio has already gone.

The two statements do not contradict each other: the first describes the design, the second how it actually works. Google confirms it on its own account, on its page about audio recordings: audio can be saved if the device wrongly detects an activation, a noise that sounds like “Hey Google” for example. And some devices include a few seconds before the activation, in order to capture the complete request.

The same European document also says the thing that concerns you, with no watch on your wrist at all: assistants can accidentally capture the audio of people who did not intend to use them, it is highly unlikely that an accidental activation could be interpreted as valid consent, and third parties’ voices can end up collected as background noise. The Garante per la protezione dei dati personali, the Italian data protection authority, reaches the same point in its briefing on digital assistants: the data can also concern people who happen to be in the same room, and it advises choosing an uncommon wake word, because everyday words cause unintended activations.

On the storage side, two useful and verified facts.

Google: the setting that saves voice recordings is off by default, and the official page says so. If it is off, Google still processes the audio in order to answer you but does not save it. And if you turn it off now, what is already saved is not deleted: that has to be deleted separately.

Apple: it states that it does not keep the audio of Siri interactions unless you explicitly opt in to the improvement programme, and at the same time states that transcripts and request data can be kept for up to two years, tied to an identifier generated by the device and not to your account. That is not Apple contradicting itself: audio and transcript are two different things, and keeping them apart avoids telling a falsehood in either direction.

On the American lawsuit, which gets quoted often and almost always badly: it is Lopez and others against Apple, the federal court for the Northern District of California, a fund of 95 million dollars, a class made up of owners of devices with Siri in the United States for the period from 17 September 2014 to 31 December 2024, and the Apple Watch is expressly on the list of devices. Up to 20 dollars per device, a maximum of five devices. In the official documents Apple denies all the allegations and states that the settlement is not an admission of liability. The non-monetary part is the more interesting one: Apple undertakes to confirm the permanent deletion of Siri audio recordings collected before October 2019, and to publish a page explaining how opting in to the improvement programme works. Note that: October 2019, while the class period runs to the end of 2024. For recordings after that date the settlement contains no deletion commitment at all.

The right way to tell it, then, is not “Apple was recording in secret” and not “nothing happened”: it is that Apple paid to close a lawsuit about unintended activations, while denying every allegation.

And the motion sensors?

There is a strand of research that gets told badly in practically every case, and it is worth cutting down to size with the figures from the papers, not with impressions.

The idea is this: the accelerometer and gyroscope of a device on your wrist measure tiny movements, and from those movements you could deduce what you are typing. You do not need the microphone permission, and historically you did not need any permission at all.

The most relevant work is MoLe, presented at MobiCom in 2015. Conditions: an Android Wear watch, left wrist only, sensors at 200 Hz, a laptop keyboard, eight volunteers, a dictionary of the 5,000 most frequent English words. The result is measured in positions in a list of guesses, and the real word lands on average in twenty-fourth place. And there are two limits that close the matter. The first: the detection rate for keys struck with the left hand is between 88 and 98 per cent, for right-hand keys between 2.6 and 6.2. The watch only hears the half of the keyboard its wrist is on. The second: lowering the sampling rate makes the result collapse, and at 20 Hz the average position of the real word goes from 24 to 1,218. It is the authors themselves who point to this as the remedy.

Another study from the same year, the one by Liu and colleagues presented at the CCS conference, gets to guessing a six-digit PIN on a payment terminal with high percentages, but only by supplying a list of candidates and with the classifier tuned on the same test data, and assuming you type with the hand wearing the watch. A 2018 paper published in IEEE Transactions on Mobile Computing, by Maiti and colleagues, measured on numeric keypads, gives 84 per cent when the system is trained on the same person, 70 when trained on other people, and down to between 52 and 61 with natural typing instead of dictated typing, collapsing to 23 per cent if the sensor rate drops to 10 Hz. The same work notes that on a QWERTY keyboard the per-letter accuracy is around 30 per cent, and that keeping the sensors at that rate used 31 per cent of the battery in an hour, which makes the attack none too invisible.

And then there is the paper that founded the gyroscope-as-microphone legend, Gyrophone, from 2014. It should be read for what it says: no smartwatches, these were smartphones on a table; no live human voice, the speech was played back through loudspeakers at around 75 decibels; a vocabulary of eleven words, the digits from zero to nine plus “oh”. Speaker-independent digit recognition reaches 17 per cent with the best method, against 9 per cent for pure chance. The authors write that the gyroscope signal is not intelligible to the human ear and that transcription of fluent speech was not demonstrated.

To sum up in a repeatable way: I have not found any research demonstrating speech recognition from the motion sensors of a wrist-worn device. What is documented is inferring what is being typed, in a laboratory, with a known keyboard, a stable posture, the right wrist, a high sampling rate and a list of candidates instead of an answer. The real risk sits with PINs typed on keypads, not with conversations.

The part that counts: recording and publishing are not the same thing

An honest preamble is needed here: I am not a lawyer, these are legal references and not legal advice, and the case summaries I quote I took from legal digests, not from the full text of the judgments. If this genuinely affects you, you need the references verified in a database and a person who does this for a living.

The starting principle, in Italy, has been settled since 2003, when the joint criminal divisions of the Cassazione, the Italian supreme court, held that the recording of a conversation by one of the people taking part in it, even without the other’s knowledge, is not an interception, because whoever is capturing it is not a third party, and that it constitutes documentary evidence. So: somebody taking part in a conversation can record it.

Article 615-bis of the criminal code, the one on unlawful interference with private life, punishes anyone who improperly obtains news or images of another person’s private life using visual or sound recording instruments. It has two limits and both need saying. The first: it applies to the private dwellings referred to in article 614, so not to a bar, a street, an office open to the public. The second: the prevailing case law applies it to the outside third party, not to somebody taking part in the private moment. On the inside of a car, as a curiosity, the decisions diverge.

Now the useful part, which is the second paragraph of the same article: it punishes, separately, anyone who reveals or disseminates that news through any means of communication to the public. These are two distinct forms of conduct, and in Italian practice it is the second one that gets hit.

The most recent example comes from the Garante privacy. In August 2025 it imposed an urgent limitation on a YouTube channel, ordering the removal of an audio recording and prohibiting its reproduction in any form; in September 2026 it concluded by holding the publication of the audio and of extracts from a private chat unlawful, with a fine of 4,950 euro, justified by the purpose of making a spectacle of it, by the damage to reputation including professional reputation, and by the aggravating factor of online dissemination. In neither decision does the Garante address how that audio had been obtained: it hits the publication.

And there is a further limit that cuts down “you can always record”. In July 2025 the Cassazione held, in an employment case, that the covert recording of colleagues can be lawful only if it is relevant, necessary, not replaceable by less intrusive means, and timely in relation to pending or imminent proceedings. Recording in order to build up evidence in general is not enough, and in that particular case a recording of almost two hours was held to be exploratory, with the disciplinary sanction judged proportionate. The burden of proving the defensive purpose falls on whoever did the recording.

There is also a criminal provision called the dissemination of fraudulent recordings and filming, brought in in 2018, which concerns exactly the publishing of a recording made fraudulently in order to cause harm. I have not verified it in the text, so I stop here and flag it as the thing anyone wanting to go deeper has to read.

What you can do, if the watch is on your own wrist

Half of this question, when you get down to it, is about your watch.

1. Apple Watch, take away the voice activations: Impostazioni (settings), Siri, and turn off separately "Alza per parlare" (raise to speak), "Rileva Siri o Ehi Siri" (listen for Siri or Hey Siri) and "Premi la Digital Crown" (press the Digital Crown). The interface is in Italian; these are three distinct switches and they go off one at a time.
2. Apple Watch, clear the history: Impostazioni, Siri, "Cronologia Siri" (Siri history), "Elimina cronologia di Siri" (delete Siri history). And on the iPhone, to get out of audio retention and any human review, Impostazioni, Privacy e sicurezza (privacy and security), Analisi e migliorie (analytics and improvements).
3. Wear OS, take away the wake word: on the watch, swipe down from the top, Settings, Google, digital assistant, and turn off "Hey Google" detection. The Google page exists only in English, so these are the names from the English documentation and what you read on your own screen may differ.
4. Google, stop the retention: Account Google, Dati e privacy (data and privacy), Impostazioni cronologia (history settings), Attività web e app (web and app activity), and untick voice and audio activity. Remember that it is already off by default and that unticking it now does not delete what is already saved: for that you have to go through activity.google.com.
5. Check the apps' microphone permissions, and here I have to tell you that for the watch I did not find a documented route. For Apple the documentation says permissions are managed from the iPhone, and the microphone is not even named among the shared ones; for Wear OS I found nothing. The documented fallback is from the phone: on iPhone Impostazioni, Privacy e sicurezza, Microfono; on Android Impostazioni, Privacy, permission management, Microfono. It is the route the Garante gives too.

On the Pixel Watch there is a detail worth knowing: the wake word only works while the watch’s screen is on. On the more recent models there is instead a mode that is triggered by raising your wrist and speaking close to your mouth, with no wake word, and which signals that it is available with a small glow at the bottom of the watch face. How visible that glow is from a metre away, sitting opposite somebody, is exactly one of the things I want to try before writing it down.

To understand what sits underneath

The subject of permissions, meaning which apps can use what and where you check it, keeps coming back: I wrote about it for location in Does Google know where I have been? The three switches to check, and the reasoning on the microphone is identical. If what interests you is the opposite side, meaning what a wearable device can do when it is designed to watch and listen, there is But what are AI glasses actually for?.