Industries / Frontier science / Brain-computer interfaces
Wearable devices
This work reads brain signals from outside the skull, with sensors worn on the head or in the ear and no surgery.
7
Milestones achieved, from 1929 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1929
Brain waves through the scalp
Hans Berger, a psychiatrist at the University of Jena, took up recording electrical activity from the human head again in 1924 and in 1929 published his first report on the human electroencephalogram (EEG), a trace of the brain's electrical activity taken from the scalp. By 1931 he had 1,133 recordings from 76 people and had described the alpha and beta waves.
Lessonfew believed him until the physiologist Edgar Douglas Adrian and Brian Matthews reproduced his recordings in 1934, five years after the first report, by the account of a 2023 appraisal in Arquivos de Neuro-Psiquiatria.
Credit
Hans Berger
1988
Spelling by attention
Lawrence Farwell and Emanuel Donchin of the University of Illinois showed a grid of letters on a screen and flashed its rows and columns. When the row or column holding the letter a person was attending to flashed, the brain produced a P300, a voltage bump about 300 milliseconds after a noticed event, and the computer picked it out of the EEG. Healthy volunteers communicated at about 0.20 bits a second, or 2.3 characters a minute.
Credit
L. A. Farwell and E. Donchin
1991
A cursor moved by rhythm
Jonathan Wolpaw and colleagues at the Wadsworth Center in Albany, New York, trained healthy volunteers to raise or lower the mu rhythm, an 8 to 12 Hz wave over the part of the brain that controls movement, and turned its size into up or down movement of a cursor. Over several weeks the volunteers learned to reach a target at the top or bottom of the screen, typically in about 3 seconds.
Lessonthe skill took weeks of training, and the authors named their way of setting the cursor's response from the measured spread of each person's rhythm sizes as "the key factor in its success".
Credit
J. R. Wolpaw, D. J. McFarland, G. W. Neat and C. A. Forneris
1999
Spelling for locked-in people
Niels Birbaumer and colleagues used slow cortical potentials, gradual voltage shifts in the EEG that people can learn to control, to drive an electronic spelling device for people completely paralysed by amyotrophic lateral sclerosis (ALS), a disease that destroys the nerves that move the muscles. It gave people who could no longer operate any muscle-driven aid a way to communicate.
Credit
N. Birbaumer and colleagues, in Nature
2013
A gaming headset tested
Nicholas Badcock and colleagues at Macquarie University in Sydney recorded 21 adults with a low-cost gaming headset, the Emotiv EPOC, and a laboratory EEG system at the same time while they heard tones. The size and timing of the strong brain responses to sound matched between the two systems. A weaker response, the mismatch negativity, matched for only 11 of the 21 people.
Lessona cheap headset could stand in for lab equipment for strong signals, but not yet for weak ones.
Credit
N. A. Badcock and colleagues, in PeerJ
2018
Brain scanning in a helmet
Magnetoencephalography (MEG) measures the tiny magnetic fields the brain gives off, and normally needs a fixed scanner cooled to near absolute zero with the head held still. Elena Boto, Matthew Brookes and colleagues at the University of Nottingham built a helmet of quantum sensors that need no such cooling, with coils that cancel the background magnetic field. It recorded brain activity at millisecond resolution while people nodded, stretched, drank and played a ball game, with results that compared well with a conventional scanner.
Credit
E. Boto, M. J. Brookes and colleagues, in Nature
2023
Heard speech decoded without surgery
Alexandre Défossez, Jean-Rémi King and colleagues at Meta AI trained one model on four public datasets of 175 volunteers recorded with MEG or EEG while they listened to stories and sentences. Given 3 seconds of MEG, the model ranked the correct stretch of speech first out of more than 1,000 up to 41.3% of the time, averaged over participants, and in its top ten up to 70.7% of the time. EEG reached only 17.7% and 25.7% top-ten accuracy on its two datasets.
Lessonthis decodes speech a person hears, not speech they mean to say, and MEG did far better than EEG.
Credit
A. Défossez and colleagues (Meta AI), in Nature Machine Intelligence, October 2023
Watch for
Typed sentences read by MEG
Meta says its Brain2Qwerty v2 rebuilds sentences that volunteers type from their MEG signals with 61% average word accuracy, and 78% for the best participant. It was trained on about 22,000 sentences typed by nine volunteers, each about 10 hours in the scanner; the character error rate was 29% with MEG and 65% with EEG. Meta says it has released the code and data. The system needs a room-sized scanner, works only after a typing session ends, and learns from typing, which its intended users cannot do.
Achieved whenan outside group reproduces the accuracy, on the released data or on new volunteers, in a peer-reviewed paper.
Claimed by
Meta, June 2026
EEG in everyday headphones
Neurable One headphones, priced at USD 499, carry 12 EEG sensors in the ear pads. Neurable says they record passively while worn as ordinary headphones and report "focus, cognitive load, cognitive speed, and brain battery".
Achieved whenan independent study shows these scores follow brain activity, and not muscle or eye-movement noise, when checked against a research-grade EEG system.
Claimed by
Neurable, October 2026
Open
01
One phantom, many headsets
A public scorecard of consumer and open-hardware EEG devices (headbands, headsets, EEG headphones and earbuds), each tested on the same phantom head: a gel or 3D-printed model that conducts electricity like the scalp, fed known signals from a signal generator. Each device is scored on its noise level, how well it recovers a known 10 Hz signal of set size, timing jitter against a light-triggered marker, and dropped data.
The briefCredit · unclaimed02
Testing focus scores on phantoms
A bench test of the scores that consumer headsets and EEG headphones show their users, such as focus, calm or "brain battery". Each device sits on a phantom head that is fed four kinds of signal in turn: brain-like rhythms that change in a set pattern, muscle-like high-frequency noise, blink-like pulses, and a flat signal. The test records what the app reports for each.
The briefCredit · unclaimed
Open 01 · the brief
One phantom, many headsets
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scorecard of consumer and open-hardware EEG devices (headbands, headsets, EEG headphones and earbuds), each tested on the same phantom head: a gel or 3D-printed model that conducts electricity like the scalp, fed known signals from a signal generator. Each device is scored on its noise level, how well it recovers a known 10 Hz signal of set size, timing jitter against a light-triggered marker, and dropped data.
Why it is still open
Phantom tests so far compare one or two systems at a time. Maynooth University's gel phantom, driven by a solar cell, compared the timing of a medical-grade g.Tec system with an Emotiv EPOC+. The authors of a 2024 ear-EEG test kit write that "in EEG, a standardized phantom does not exist". A 2025 benchmark of the Emotiv EPOC X and Muse S tested 24 people, not a phantom. A 2025 Heriot-Watt University team built a 3D-printed conductive phantom head for £48.10, but checked it with a single OpenBCI board. No multi-device phantom scorecard was found.
mural.maynoothuniversity.iepmc.ncbi.nlm.nih.govebi.ac.ukpmc.ncbi.nlm.nih.gov
Built with what exists
A conductive gel or agar head phantom, or the 3D-printed conductive phantom described by the Heriot-Watt team; a two-channel signal generator or a microcontroller with a digital-to-analog converter; a photodiode or small solar cell for timing markers, as in the Maynooth method; Lab Streaming Layer to record every device on one clock; MNE-Python for analysis; an OpenBCI Cyton board as a low-cost reference. The advance is cheap 24-bit EEG amplifier chips such as Texas Instruments' ADS1299, used on OpenBCI boards, which a 2016 side-by-side test by Jérémy Frey found could stand in for a medical-grade g.tec amplifier in BCI tasks.
First step
Build a gel phantom, drive it with a 10 Hz sine wave of about 50 microvolts, and record it with an OpenBCI Cyton and one consumer headband at the same time.
Done looks like
At least 6 devices, each tested in at least 10 sessions on the same phantom with the same 5 test signals; a published table of noise (microvolts, root mean square), amplitude error (%), timing jitter (milliseconds) and data loss (%); raw files, phantom build files and code on Zenodo.
Open 02 · the brief
Testing focus scores on phantoms
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A bench test of the scores that consumer headsets and EEG headphones show their users, such as focus, calm or "brain battery". Each device sits on a phantom head that is fed four kinds of signal in turn: brain-like rhythms that change in a set pattern, muscle-like high-frequency noise, blink-like pulses, and a flat signal. The test records what the app reports for each.
Why it is still open
The independent comparisons found tested people and computed their own measures from the raw signal, not the vendors' scores. A 2025 study found that measures from the Emotiv EPOC X and Muse S were "less sensitive to task demands" than those from a research-grade system. No bench test of app scores against known inputs was found. Meanwhile EEG scores are now sold in everyday headphones (see Watch for).
Built with what exists
The phantom and signal generator from the item above; a phone running each vendor's app; the vendor's developer kit where one exists, or screen recording with text recognition, to log the scores; Python for the tally.
First step
On the phantom, log one headband app's focus score through ten minutes of muscle-like noise (20 to 100 Hz) and ten minutes of flat signal, and compare.
Done looks like
At least 3 products, each with 20 runs per signal type, and a published tally of whether each score moved with the brain-like rhythm and whether it moved with muscle or blink noise alone.