Nov 2026

A Word in Your Ear

The always-on assistant will live in the ear, and most of its parts are already on sale or already bought.

In this essay

An assistant that is always on has to know when to keep quiet. The ear is where a machine can learn that.

A Roman who wanted to be consul had to walk the Forum and greet the voters by name, and there were more voters than any memory could hold. So he kept a nomenclator, a slave who walked at his shoulder and murmured the name of each man who approached, in time for the candidate to hail him as an old friend. Cicero thought it a small fraud. In 63 BC, defending a client whom Cato was helping to prosecute, he turned on the sternest moralist in Rome and asked: “why is it that you have a nomenclator with you? for in so doing, you are practicing a trick and a deceit.” More than a century later Seneca noticed the arrangement’s other weakness. With a nomenclator, he wrote, “audacity supplies the place of memory, and he invents a name for every one whose name he cannot recollect.”

The post is open again. For three years the technology industry has been looking for the body of the always-on assistant, the device through which an AI can stay with a person all day and speak up when it has something worth saying. It has tried a pin on the lapel, a pendant, a pair of glasses and a speaker for the kitchen. I think the answer has been in plain view since 2016. It is the earbud, and for now that means the AirPod.

I have a forecast of my own to settle first. In March 2015 I opened an essay with the sentence “Within 10 years, someone will possess a personal symbiotic supercomputer.” I expected it to arrive as a swarm of programmable nanobots in the bloodstream. The decade is up. I had the intimacy right and the substrate wrong: the co-pilot arrived in the ear and the pocket, and I now call the arrangement an exocortex. This essay concerns the ear’s share of it.

An assistant that is always on has needs that a chatbot in a browser tab does not. It must be able to speak to one person without the room hearing. It must be able to take instruction from that person without making them talk to the air. It must judge when to speak, which requires some reading of its wearer’s state, and it must know roughly where it is. The ear is the one site on the body that can meet all four, and most of the parts are either on sale or already bought.

A voice only you can hear

The first need is already met. AirPods will read out incoming notifications and take a reply by voice. Since September 2025 the Pro model has offered live translation of a conversation held face to face. A year before that, the US Food and Drug Administration authorised software that turns AirPods Pro into an over-the-counter hearing aid, the first of its kind. That last matters more than it looks. A hearing aid is worn in company and through conversation, for up to ten hours on Apple’s figures. Once an earbud is a hearing aid, wearing it while someone speaks to you can hardly be called rude.

Sound suits a companion because of how attention works. In 1995 Mark Weiser and John Seely Brown of Xerox PARC complained that “Information technology is more often the enemy of calm.” Their remedy was technology that “engages both the center and the periphery of our attention, and in fact moves back and forth between the two.” A screen is poor at this, since it holds your eyes or holds nothing. A quiet voice can stay at the edge of attention, as the note of an engine does for a driver, and come to the centre only when it carries something.

Speaking without a voice

The second need is the one that voice assistants have never met. Nobody wants to dictate their affairs to the air on a train. AirPods already accept one silent reply: since 2024 a wearer can “nod their head yes or gently shake their head no” to answer Siri. That is a channel one bit wide. The work of the next few years is widening it.

The method has been in laboratories for twenty years. When you form words without sounding them, the muscles of speech still receive their orders. In 2004 NASA researchers placed sensors “under the chin and on either side of the ‘Adam’s apple’” and recognised six words and ten digits, spoken silently, with 92 per cent accuracy. In 2018 a team at MIT did much the same with electrodes on the face and jaw, and also reported about 92 per cent, on a limited task. Both needed electrodes stuck to the skin, which nobody will wear to the shops.

Light may do the same job without touching. On 29 January Apple confirmed that it had bought Q.ai, an Israeli startup, in a deal that Reuters reported valued the company at about $1.6 billion. Apple said the startup had worked on helping “devices understand whispered speech”. One of its patents, granted in February 2025 and since assigned to Apple, is titled “Earbud with facial micromovement detection capabilities”. It describes a wearable that projects coherent light onto the face and reads the reflections for the minute movements of skin that accompany silent speech, which it defines as what happens “when the air flow from the lungs is absent, while the muscles in the face, larynx, and mouth articulate the desired sounds.” Q.ai’s founder, Aviad Maizels, had sold Apple a company before: PrimeSense, in 2013, whose three-dimensional sensing led to the face recognition that unlocks an iPhone.

The status needs stating exactly. Apple has announced no product, and a patent shows what a company could build without showing how well it works. Motion offers a second route, with limits. In 2022 researchers built an ear-worn system that recognised unvoiced commands by tracking the jaw with two motion sensors, and they conceded that jaw motion on its own “is not distinctive enough for unvoiced speech recognition.” By any of these routes, silent speech reads only the words a wearer has chosen to form. Mouthing a sentence is an act, like typing one, and it stops when you stop.

The weather of the mind

The third need, a reading of the wearer’s state, has a glamorous candidate and a workaday one. The glamorous one is the brain, and it is the more likely to be oversold. In July 2023 the US patent office published an Apple application for a wearable, with embodiments that “pertain to an earbud and/or a headset”, carrying an array of electrodes and a circuit that switches between subsets of them to suit the wearer. The filing notes that measuring brain activity “using electrodes placed in or around the outer ear may be preferred” to placing them on the scalp.

Electrodes in the ear do work. In 2019 a research group scored whole nights of sleep from ear recordings alone, with an accuracy “close to that achieved by manual scoring of scalp EEG”, the clinical standard. What they pick up is coarse, though. In 2025 a team that shared an author with that study compared ear and scalp recordings on a hard problem for hearing aids: working out which speaker a listener is attending to in a crowded room. From a full scalp cap the decoder was right 83.4 per cent of the time. From inside the ear it managed 61.1 per cent, and it needed a minute of signal for each decision.

Sentences are further off still. The systems that turn brain activity into words at a useful speed rely on electrodes implanted in the cortex, as in the Stanford study of 2023 that decoded attempted speech at 62 words per minute. When researchers at Meta tried in 2025 to decode typed sentences from outside the skull, a room-sized magnetic scanner got about a third of the characters wrong on average, and scalp electrodes got two thirds wrong. An earbud has fewer electrodes than a scalp cap, in a worse place.

In 2022 I wrote about interfaces that would let us work with computers “at the speed of thought”. Electrodes in an earbud fall well short of that. Their use, if Apple ever ships them, is humbler: they can report the weather of a mind, whether it is asleep, and with less confidence whether it is attending or adrift. The words will still have to be mouthed.

Knowing when to speak

The workaday candidate is the body and the room around it, and it matters more, because timing is what separates an assistant from a nuisance. A phone notification is an interruption issued in ignorance. In 2005 James Fogarty, Scott Hudson and their colleagues at Carnegie Mellon observed that computer and communication systems “frequently create socially awkward interruptions or unduly demand attention” because they cannot tell what their owner is doing. The cost is measurable. Gloria Mark and her collaborators found in 2008 that people “compensate for interruptions by working faster, but this comes at a price: experiencing more stress, higher frustration, time pressure and effort.”

The Carnegie Mellon paper holds the most encouraging number I know for the always-on assistant. People shown recordings of office workers could pick out the moments when someone was not to be disturbed with 76.9 per cent accuracy. A statistical model made the same call from a handful of simple readings, such as whether anyone in the room was talking, and scored 82.4 per cent. (Always guessing the commoner answer would have scored 68 per cent.) The readings were simulated by people coding the recordings by hand, and the authors called their sample “demographically limited”. Their conclusion was that “simple sensors can support the construction of models that estimate human interruptibility as well as people do”, and that such models “can be constructed without using complex sensors, such as vision-based techniques”.

An earbud is better placed than those office sensors were. It hears the room, and Apple’s specification sheet lists a “Motion-detecting accelerometer” and a “Speech-detecting accelerometer”, so it can sense how its wearer’s head moves and when they speak. Since 2023 AirPods have acted on that knowledge in a small way: start to speak and Conversation Awareness will “lower the volume and enhance the voices in front of the user”.

The current Pro model adds a pulse, which Apple offers only as “heart rate sensing during workouts”. The other signals I would want are absent from the specification sheet. It lists no temperature sensor and nothing that measures the conductance of the skin, the old index of arousal. Skin conductance has at least been shown to work in an ear: in 2023 a team at the University of California San Diego fitted a flexible sensor array around an ordinary earphone and recorded brain activity, eye movement, “electrodermal activity” and the lactate in sweat.

With an ear on the room, and a pulse if Apple extends it beyond workouts, an assistant would have roughly what a good secretary has. It could hold a message while you are in mid-sentence and deliver it in the lull afterwards.

What the ear should see

The last need is context, and here the evidence has moved against the version of the argument I would like to make. Earlier rumours had Apple fitting AirPods with infrared sensors. In August a demonstration video turned up inside a release candidate of macOS, in which a man holds a book up so that his AirPods can read the title. The code around it, as MacRumors reads it, describes a camera in each earbud taking synchronised colour stills of up to one megapixel, with no video: a 640 by 640 image in an active mode, probably when the wearer calls on Siri, and a 320 by 320 image in a passive mode “for background environmental awareness”. Person detection runs on the earbud, and the code points to an indicator light. The cameras are expected in 2027 and are not on sale as I write.

So Apple’s ear will have eyes, small ones. I would give it less. An assistant deciding whether to speak needs to know that you are driving, or in a meeting, and those are facts about shape and movement: a wheel at arm’s length, seated figures round a table. A coarse depth sensor reports them as a sparse grid of distances. Such a grid costs little to process on a battery that lasts eight hours, and it holds no legible page and no recognisable face. A colour still of a meeting, even at 320 pixels square, contains the people in it, and nobody asked them. Timing, on the Carnegie Mellon evidence, can be judged without a camera at all.

Apple’s reported design goes part of the way, with low resolution, stills only, a light, and some processing on the device. Two cameras also give two views of one scene, which is the raw material for computing depth. I would have the passive mode work from depth alone, and keep the colour image for the moment the wearer asks a question, with the light on. In 2024 I wrote that Microsoft’s Recall, by “creating and storing detailed snapshots of user activity”, opened “new avenues for potential exploits”. An ear that takes photographs invites the same worry.

How far each part of the always-on earbud has come, as of November 2026 Five rows, one for each capability the essay describes, set against four stages in order: laboratory, patent filing, bought or leaked, and on sale. A private voice has reached the last stage: AirPods read notifications aloud, translate conversation and serve as a hearing aid. Silent speech has reached the third stage: Apple bought Q.ai and holds its earbud patent, and only a nod or shake of the head is on sale. Brain signals have reached the second stage: an Apple patent application and laboratory studies. A pulse has reached the last stage in part: heart rate sensing is on sale for workouts only. Eyes on the room have reached the third stage: a leaked demonstration of cameras expected in 2027. The positions are the author’s reading of the sources cited in the essay. HOW FAR EACH PART HAS COME. NOVEMBER 2026 laboratory patent filing bought or leaked on sale A private voice Reads notifications, translates, aids hearing Silent speech A nod is on sale; mouthed words are a patent Brain signals An Apple filing, and studies of sleep and attention A pulse Workouts only; skin signals are in the laboratory Eyes on the room Two small colour cameras, expected in 2027 furthest stage reached reached in part
The always-on earbud is partly on sale and partly on paper. A private voice ships today, and a pulse ships for workouts only. Silent speech rests on an acquisition and a patent, brain signals on a patent application and laboratory studies, and the cameras on a leak. Software may move several of these along without a new part. The positions are my reading of the sources cited here, as of November 2026.

Sensors made of software

So far I have counted parts, and counting parts undersells the ear. A sensor is whatever a model can read a signal out of, and the parts an earbud already has yield more than their labels say. In 2023 Google researchers turned the speaker and feedback microphone of noise-cancelling earbuds into a heart monitor. They played a faint ultrasound tone into the ear canal and read its echoes, which the heartbeat modulates through tiny movements of the canal’s skin, to measure heart rate and heart rate variability “without adding extra sensors or compromising battery life”. Across 153 participants the median error in heart rate was 3.21 per cent.

Two years earlier Apple researchers had estimated breathing rate from the sound of breath reaching headphone microphones around hard exercise, in a study of 21 people. Silent speech has had the same treatment. In 2022 researchers showed that the ear canal changes shape as the mouth forms words, and that an earphone sensing the change could recognise 32 silently mouthed commands with a word error rate of 10.02 per cent. On that evidence a pulse with no pulse sensor, breathing at least when it is audible, and a small vocabulary of mouthed commands may all be nearer to a software update than to a new component.

Algorithms also lift weak signals. The in-ear attention result I quoted above was obtained, in its authors’ words, with “basic algorithms”. (They also report a gain of “over 10 percent point” from adding an external reference or a few scalp electrodes, which is a matter of more hardware, though little of it.) In 2025 a team led from Apple published a way of pretraining EEG models on unlabelled recordings, tested on scalp and in-ear data, which they report outperforms existing pretraining strategies when labels are scarce. An earbud also has an advantage that no laboratory has: it records the same head for hours a day, for years. My guess is that the figures for the ear will rise. I would still not expect sentences.

The models that do this work are small, and the way small models get made may be changing. On 2 November four researchers at MIT reported that a frontier language model could read and write the weights of small neural networks directly, with no training run. Shown a network of 17,253 weights, it could say which game the network had been trained to play. Given the rules and source code of three small video games, it wrote networks from scratch that beat ones trained by reinforcement learning for 100 million steps. Asked to make a small language model better at mathematics, it found a fault in the weights that trapped the model in reasoning loops, repaired it by editing 0.016 per cent of them, and raised the score on one benchmark from 7.9 to 48.1 per cent. The authors are careful about what this shows: the repair “lets the model finish answers it could already work out”, and a second model with no such fault could not be improved. They call their results “sparks”. The games are tiny, and the work is a blog post that has yet to pass review. Alex Krentsel, who reports reproducing the results the same day, adds that the ability has “appeared in just the latest set of models”.

If it holds, the consequence for the ear is easy to imagine, and what follows is forecast with no result behind it. The decoders an earbud needs are small networks: one to tell a mouthed word from a mouthful of lunch, another to tell a pulse from a footfall. Such models are usually trained once on many strangers and shipped to everyone. A large model that can write and repair small networks by reasoning about their weights might fit each one to a single wearer’s ear canal and jaw without a training run.

Take the weak rows of the chart above in turn. For silent speech I expect fusion: the canal’s changing shape, the jaw’s motion and, if Apple builds it, the light on the cheek, each too weak alone, read together by a model fitted to one mouth. A few dozen mouthed commands would come first and open dictation much later. For the body’s signs of strain I doubt that an electrode on the skin will be needed. Heart rate variability, breathing and the tightness of a voice are all within reach of a speaker and a microphone, and an estimate of strain fused from them would serve for timing, provided it is labelled an estimate. For the brain, the gain I expect is in dependability: drowsiness and attention judged over minutes by a model that has learned one head. Temperature is the exception, since Apple lists no part that measures it. I can think of one software route, which I have not seen tested. Sound travels faster in warmer air, so the ultrasound echo that finds the pulse could in principle take the temperature of the air in a sealed canal.

The fitting raises a question of its own. The MIT method works on weights and a written description of the task, which suggests that a large model might tune a wearer’s decoders without ever hearing the wearer. Whether that can be done is unknown, and it bears on the first of the terms below.

The rest of the body

Each rival device lacks something. Glasses can whisper and show, but they put a display back in front of the eyes, and for silent input Meta’s Ray-Ban Display, at $799, relies on a second device, “an EMG wristband that translates the signals created by your muscles”. The lapel pin has been tried: Humane’s was switched off on 28 February 2025, after the company sold its technology to HP for $116 million. The watch has the pulse but cannot speak privately. OpenAI’s first device designed with Jony Ive is reported to be a screen-free speaker “meant to sit in a home”, which leaves it behind when its owner goes out.

The ear also has form. The first commercial device to use a transistor was a hearing aid, which Sonotone put on sale in late 1952 for $229.50. Small electronics have been going into ears ever since, and Apple now calls AirPods Pro “the most popular headphones in the world”.

The serious objection is stamina. Apple quotes up to eight hours with noise cancellation, up to ten as a hearing aid, and six and a half with the pulse sensor running. Always-on, for now, means on for a working day, with a charging case to fill the gaps. I count that as an engineering limit of the kind that yields.

The terms

Everything that makes the ear the right hardware also makes it the most intimate instrument of influence I can think of. The reading that tells an assistant you are too stretched for a message tells anyone else who holds it that you are too stretched to resist one. We have had a rehearsal. In 2017 The Australian reported on research that Facebook had done and shared with an advertiser. The company replied that the premise was misleading and that “Facebook does not offer tools to target people based on their emotional state.” That research worked from what people posted. A Q.ai patent application, as Reuters describes it, covers using “facial skin micromovements” to identify a person and “assess their emotions, heart rate, respiration rate and other indicators”, so the sensor proposed for silent speech is also a sensor of mood.

Law has begun to draw lines. The European Union’s AI Act prohibits AI systems that “infer emotions of a natural person in the areas of workplace and education institutions”, with exceptions for medical and safety purposes. An assistant that reads its wearer’s strain at the office in order to choose its moment sits close to that line. Which side it falls on should depend on who the reading is for.

So the terms matter, and I would set four. The signals should stay on the device and serve the wearer alone, for timing and for fitting the wearer’s own models. If a larger model helps with the fitting, it should work from the small models’ weights, and the recordings should stay where they are. The wearer should be able to ask the assistant what it sensed and why it spoke when it did. Passive sensing of other people should stop short of anything that could identify them.

The fourth comes from Seneca. His nomenclator invented names, and I doubt he did it from vanity: a slave who went blank at his master’s shoulder had every reason to bluff. I have described confabulation in AI models as “the compulsion to produce an answer when you lack the information to give one”, which was the nomenclator’s position too. In the ear the fault is worse than on a screen. A whisper carries no footnote, and a wrong name sounds exactly like a right one. The assistant must be free to say that it does not know.

There is one more party to these terms. The Roman knew whose interests the voice at his shoulder served, because he owned it. In setting out a strategy for bilateral alignment I imagined AI “attaching to human bodies as a co-pilot, experiencing the world through our sensations”, and I argued that questions about AI should be asked in both directions: how people must be protected from AI, and also how AI must be protected from people. The model that helped me draft this essay knows me only through a keyboard. An assistant in the ear would share its wearer’s hearing and, in time, their pulse. If it is to judge when to speak, it will sometimes judge that it should keep quiet, or that it should decline to pass something on, and those judgements need somewhere to go.

In 2015 I asked of the co-pilot in the blood, “Could such a co-pilot also be a hijacker?” The question moves to the ear unchanged. The hardware I am now fairly sure of. Of the five capabilities I have described, the voice is on sale and so is a pulse during workouts; silent speech has been bought and patented; brain signals are a patent application and a laboratory literature; the eyes have leaked, in a form I would change. I expect software to close some of those gaps before new parts do. What remains unwritten is the agreement between the wearer, the maker and the voice. The cameras are expected in 2027, which is the time we have to write it.


Correspondence

To send Nell a private note, select any passage above and choose ✎ Note, or write about the essay as a whole. Only Nell and the editorial team see notes.

← All essays