Skip to content
Robotic Brain Universe

Brain–computer interfaces

How implants like BrainGate and Neuralink read intention from motor cortex, the maths of decoding, and the race to restore speech and movement.

Intermediate · about 12 min · updated 2026-10-02 · awaiting clinical review

See it in 3D:

Illustrative simulation excitatory inhibitory

Brain–computer interfaces turn brain activity into commands. This reading covers directional tuning and the population vector (with an interactive decoder), Kalman filters and neural-network decoders, scalp, endovascular, surface and intracortical recording, writing touch back into the brain, speech neuroprostheses at conversational speeds, Neuralink's threads, and the failure modes and ethics that remain.

Contents
  1. Thoughts into actions
  2. What a brain–computer interface is
  3. Why build one
  4. How intention is read
  5. When the signals are there
  6. When it fails, and what can go wrong
  7. The mathematics of decoding
  8. The technology, from Utah arrays to threads
  9. Milestones
  10. Frontiers
  11. Check yourself

Thoughts into actions

In January 2024 Neuralink, Elon Musk's brain-implant company, implanted its brain–computer interface in its first human volunteer. According to the company, he soon used it to move a computer cursor by thought, playing online chess and the strategy game Civilization VI.[1,2,3]

Neuralink is one part of a field with decades of peer-reviewed results. A person with amyotrophic lateral sclerosis (ALS) who could no longer speak intelligibly had his attempted speech decoded to text at 62 words per minute; another man with ALS used a speech neuroprosthesis for self-paced conversation at about 32 words per minute, with 97.5% accuracy, over more than 8 months. A man paralysed by a spinal cord injury walked again through a wireless bridge between his brain and his spinal cord.[4,5,6]

This reading explains how a computer can read intention from the brain: what the motor cortex encodes, the mathematics of decoding, the different ways of recording, how the brain can be written to as well as read, and the hard problems that remain, from scar tissue to mental privacy.[7,8,9]

What a brain–computer interface is

A brain–computer interface (BCI) gives the brain a new, non-muscular channel for sending messages and commands to the outside world. It records brain activity, works out what the user intends, and turns that into commands for a computer display or another device. The user and the system must adapt to each other, both at first and continually.[7]

BCIs differ in where the sensors sit. Non-invasive systems record from the scalp (the electroencephalogram, EEG) or use functional MRI. Electrocorticography (ECoG) uses electrode grids on the brain's surface. Intracortical arrays push fine electrodes into the cortex to record the spikes of individual neurons. An endovascular device reaches the cortex from inside a blood vessel.[7,10,11,12,13]

Four ways to listen to the brain[1,11,12,13,14,15]
ApproachExampleWhat it recordsTrade-off
Scalp EEGP300 speller (1988)Summed activity of large populations through the skullNo surgery, but slow: 2.3 characters per minute in the first speller
EndovascularStentrodeECoG-like signals from inside a vein over motor cortexNo open brain surgery; signals coarser than spikes
Surface ECoGUCSF speech neuroprosthesisHigh-density grids on the speech cortexCraniotomy; large-vocabulary speech at 78 words per minute
IntracorticalBrainGate, NeuralinkSpikes of individual neuronsHighest resolution; electrodes inside brain tissue

Milestones in numbers

Information transfer rate of BCIs, 2002
up to 10–25 bits per minute[7]
Handwriting decoded from motor cortex
90 characters per minute, 94.1% raw accuracy[16]
Attempted speech decoded from intracortical arrays
62 words per minute[4]
Speed of natural conversation, for comparison
about 160 words per minute[4]
Neuralink N1 implant (company-reported)
1,024 electrodes on 64 threads[1]

Why build one

Spinal cord injury, brainstem stroke, ALS and other disorders can disconnect the brain from the body and take away voluntary movement, while the brain's motor areas still work. A BCI routes movement-related signals from the brain around the damaged parts of the nervous system to an external device.[12,17]

The most immediate goal is communication for people who are completely paralysed or 'locked in', so they can express their wishes to caregivers or use a computer. The longer-term goals are control of prosthetic limbs and, through stimulation, restoring sensation and movement.[6,7,18]

How intention is read

The key discovery came from monkeys. Georgopoulos and colleagues recorded single neurons in motor cortex while monkeys moved a handle in eight directions. Three quarters of the arm-related cells they studied in detail (241 of 323) were directionally tuned: each fired most for one preferred direction and progressively less for directions further from it, and in most of them the firing followed a cosine curve.[19]

No single cell specifies a movement, because each is broadly tuned. But treat every cell as a vector pointing in its preferred direction, scale it by how much the cell's firing changes, and add them up: the resulting population vector points in the direction of the movement. This is the principle behind decoding intended movement from a group of neurons.[8]

Modern decoders go further. A Kalman filter treats hand motion as a hidden state that changes smoothly over time and the firing rates as noisy linear measurements of it, and combines both to estimate the motion in real time, with an estimate of its own uncertainty. Recent systems use recurrent neural networks and language models to turn neural activity into handwriting or words.[16,20,21]

One loop of an intracortical BCIIntentionthe user attempts a movement ora wordMotor cortexneurons change their firingElectrode arrayrecords spikes from many neuronsSignal processingspike detection, firing ratesDecoderKalman filter or neural networkDevicecursor, text, speech, roboticarmFeedbacksight, or touch by stimulationcloses the loop
One loop of an intracortical BCI. A BCI is a closed loop: the user sees (or, with stimulation of the sensory cortex, feels) what the device does and adjusts, while the decoder adapts to the user.[7,12,18,20]
Text version of the diagram
  1. Intention: the user attempts a movement or a word. Leads to Motor cortex.
  2. Motor cortex: neurons change their firing. Leads to Electrode array.
  3. Electrode array: records spikes from many neurons. Leads to Signal processing.
  4. Signal processing: spike detection, firing rates. Leads to Decoder.
  5. Decoder: Kalman filter or neural network. Leads to Device.
  6. Device: cursor, text, speech, robotic arm. Leads to Feedback.
  7. Feedback: sight, or touch by stimulation. Leads to Intention (closes the loop).

BCIs can also write to the brain. Microstimulation through electrodes in the hand area of the somatosensory cortex of a person with a long-standing spinal cord injury evoked touch sensations felt on the hand, many with a natural quality such as pressure, arranged in the expected body map. Feeding these sensations back while the participant controlled a robotic arm halved the time needed for a clinical upper-limb test, from a median of 20.9 to 10.2 seconds.[18,22]

When the signals are there

A BCI only works if the brain still produces movement signals long after the body stopped moving. It does: in the first BrainGate participant, intended hand motion modulated motor cortex spiking three years after a spinal cord injury, and a participant implanted five years earlier used a robotic arm to drink coffee from a bottle.[12,17]

Speech signals persist too. In a person with ALS, a detailed articulatory representation of speech sounds remained in motor cortex years after paralysis.[4]

Learning is getting faster. One speech neuroprosthesis reached 99.6% accuracy on a 50-word vocabulary on its first day of use, after 30 minutes of calibration recordings, and 90.2% on a 125,000-word vocabulary on the second day. A brain–spine interface for walking calibrates in a few minutes and stayed reliable for over a year, including at home.[5,6]

When it fails, and what can go wrong

The brain fights back. Arrays that work well in short recordings often fail to work reliably over years. A major failure mode is the reaction of brain tissue to the implant, which makes biocompatibility a primary concern in electrode design.[23]

Electrodes can move. Neuralink reported that in the weeks after its first implantation a number of threads retracted from the brain, reducing the number of effective electrodes and cursor-control performance; changes to the recording and decoding software restored performance.[2]

Surgery has costs. Implanting arrays through an open craniotomy can provoke inflammatory tissue responses, which motivated the endovascular stent-electrode, placed through a catheter in a vein over the motor cortex.[15]

Mental privacy. A decoder using fMRI and a language model reconstructed the meaning of speech that people heard or imagined. Its authors tested privacy directly and found that the person's cooperation was required both to train the decoder and to use it. Ethicists argue that neurotechnology and AI must respect privacy, identity, agency and equality.[9,10]

The mathematics of decoding

Decoding intention is a statistics problem: given noisy firing rates, estimate the movement most likely to have produced them.[20]

Cosine tuning of a motor cortex neuron[19]
f(θ)=b0+c1cos⁡(θ−θ0)f(\theta) = b_0 + c_1 \cos(\theta - \theta_0)

The firing rate of a directionally tuned neuron is highest when the movement direction θ\theta equals its preferred direction θ0\theta_0 and falls off smoothly on either side. Georgopoulos and colleagues found this sinusoidal relation in about three quarters of the tuned cells.

Symbols in Cosine tuning of a motor cortex neuron
SymbolMeaningUnit
f(θ)f(\theta)firing rate for movement direction θspikes/s
b0b_0baseline ratespikes/s
c1c_1depth of modulationspikes/s
θ0\theta_0the cell's preferred direction°
Population vector[8]
P(θ)=∑i=1N(fi(θ)−bi) ci\mathbf{P}(\theta) = \sum_{i=1}^{N} \big(f_i(\theta) - b_i\big)\,\mathbf{c}_i

Each neuron votes for its preferred direction ci\mathbf{c}_i with a weight equal to its change in firing. The vector sum points in the direction of movement, even though every single neuron is only broadly tuned.

Symbols in Population vector
SymbolMeaningUnit
P\mathbf{P}population vector—
ci\mathbf{c}_iunit vector along neuron i's preferred direction—
fi,bif_i, b_ineuron i's firing rate and baselinespikes/s

Try it

grey: firing rates; dashed: true; solid: decoded
Neurons in the population (preferred direction)

Decoded direction 60°, error 0°. With all eight evenly spaced neurons the population vector recovers θ exactly; remove some and it drifts.

Illustrative values: baseline b = 20 Hz, modulation m = 15 Hz, no noise.

Set a movement direction and watch eight cosine-tuned neurons respond; the population vector (solid) recovers the direction (dashed). Switch neurons off and the estimate drifts: with fewer, unevenly spread neurons the population vector becomes biased.[8,19]
Kalman filter model[20]
xt=A xt−1+wt,zt=H xt+qt\mathbf{x}_t = A\,\mathbf{x}_{t-1} + \mathbf{w}_t, \qquad \mathbf{z}_t = H\,\mathbf{x}_t + \mathbf{q}_t

The first equation is a prior on how hand position and velocity evolve; the second says how firing rates depend on the hand's state. Both are linear with Gaussian noise, so the filter can update its estimate recursively at every time step, which makes it fast enough for real-time control.

Symbols in Kalman filter model
SymbolMeaningUnit
xt\mathbf{x}_thand state (position, velocity, acceleration) at time t—
zt\mathbf{z}_tfiring rates of the recorded neurons at time tspikes/s
A,HA, Hstate-transition and observation matrices, fitted from training data—
wt,qt\mathbf{w}_t, \mathbf{q}_tGaussian noise terms—
Bits per selection[7]
B=log⁡2N+Plog⁡2P+(1−P)log⁡21−PN−1B = \log_2 N + P\log_2 P + (1 - P)\log_2\frac{1 - P}{N - 1}

How much information each choice conveys when the user picks one of NN targets with accuracy PP. Multiplying by selections per minute gives the information transfer rate used to compare BCIs; with perfect accuracy B=log⁡2NB = \log_2 N.

Symbols in Bits per selection
SymbolMeaningUnit
NNnumber of possible targets—
PPprobability that a selection is correct—
BBbits per selection—
Word error rate[4,21]
WER=S+D+IN\mathrm{WER} = \frac{S + D + I}{N}

The standard accuracy measure for speech neuroprostheses: the number of word substitutions, deletions and insertions needed to turn the decoded sentence into the intended one, divided by the number of intended words. A 9.1% word error rate means about one word in eleven is wrong.

Symbols in Word error rate
SymbolMeaningUnit
S,D,IS, D, Isubstituted, deleted and inserted words—
NNnumber of words in the intended sentence—

The technology, from Utah arrays to threads

Intracortical arrays. BrainGate's first participant had a 96-microelectrode array in primary motor cortex and used a 'neural cursor' to open e-mail and operate a television, even while talking. In Pittsburgh, two 96-channel arrays let a 52-year-old with tetraplegia control a prosthetic arm in seven dimensions, succeeding in 91.6% of reaching trials.[12,24]

Neuralink's threads. Neuralink's 2019 white paper described flexible electrode 'threads', up to 3,072 electrodes per array on 96 threads, inserted by a neurosurgical robot at up to six threads a minute while avoiding surface blood vessels. Its current N1 implant, the company says, records through 1,024 electrodes on 64 threads, each thinner than a human hair, placed by its R1 robot.[1,3]

Through a vein. The Stentrode is implanted with a catheter into the superior sagittal sinus next to the primary motor cortex. In its first two participants, both with ALS, it worked with an eye tracker to let them text, shop online and manage their finances at home.[13]

Speech from the surface. A high-density ECoG grid over the speech cortex of a participant with severe limb and vocal paralysis drove three outputs at once: text at a median 78 words per minute, synthesised speech personalised to the participant's pre-injury voice, and a talking facial avatar.[11]

Older cousins. The cochlear implant, which stimulates the auditory nerve, lets the average user follow predictable conversation in quiet with relative ease. Deep brain stimulation can both measure pathological brain activity and deliver adjustable therapeutic stimulation, and is among the most important clinical advances in neuroscience of the past two decades.[25,26]

Milestones

From one neuron to a spoken sentence

  1. 1982Motor cortex neurons are found to be cosine-tuned to movement direction.[19]
  2. 1986The population vector predicts movement direction from many neurons.[8]
  3. 1988The P300 speller lets people type with brain potentials recorded from the scalp.[14]
  4. 2006BrainGate: a man with tetraplegia controls a cursor and a prosthetic hand with a motor cortex array.[12]
  5. 2012People with tetraplegia reach and grasp with a robotic arm; one drinks coffee unaided.[17]
  6. 2013Seven-dimensional control of a prosthetic arm.[24]
  7. 2016Stimulating the sensory cortex evokes touch felt on the hand; a stent-electrode records cortex from inside a vein in sheep.[15,22]
  8. 2019Neuralink describes its thread-and-robot platform.[3]
  9. 2021Handwriting BCI reaches 90 characters per minute; first sentences decoded from a person with anarthria.[16,21]
  10. 2023Speech BCIs reach 62 and 78 words per minute; a brain–spine interface restores walking.[4,6,11]
  11. 2024Neuralink's first human implant; a speech neuroprosthesis sustains 97.5% accuracy for months.[1,5]

Frontiers

Conversation-speed speech. Intracortical speech decoding has gone from a 50-word vocabulary to a 125,000-word one; one system decodes at 62 words per minute, 3.4 times the previous record, and another has been used for more than 248 hours of self-paced conversation, voiced in a synthetic copy of the user's own pre-ALS voice.[4,5]

Walking again. The brain–spine interface links implanted cortical recordings to epidural stimulation of the spinal cord regions that produce walking. Its participant could stand, walk, climb stairs and cross complex terrain, and after rehabilitation with it he could walk with crutches even with the system switched off.[6]

Reading meaning without surgery. Non-invasive decoders were long limited to choosing among a few words. Combining fMRI with a language model, Tang and colleagues reconstructed intelligible word sequences capturing the meaning of perceived speech, imagined speech and even silent videos.[10]

Check yourself

Check yourself

  1. What did Georgopoulos and colleagues find about single motor cortex neurons?
    Show answer

    Most arm-related cells were directionally tuned: they fired most for a preferred direction and less for others, following a cosine curve.

  2. How does a population vector estimate movement direction?
    Show answer

    Each neuron is a vector along its preferred direction, weighted by its change in firing; the sum points in the direction of movement.

  3. Name the four main places a BCI can record from.
    Show answer

    The scalp (EEG), inside a blood vessel (endovascular), the surface of the cortex (ECoG) and inside the cortex (intracortical arrays).

  4. Why does a BCI user benefit from touch feedback?
    Show answer

    Vision gives limited information about grasping; adding touch through stimulation of the sensory cortex halved task times with a robotic arm.

  5. What is a major reason implanted arrays stop working over time?
    Show answer

    The brain tissue reaction to the implant.

  6. What does a word error rate of 9.1% mean?
    Show answer

    About one decoded word in eleven needs correcting (a substitution, deletion or insertion).

  7. Is a company blog post the same as a clinical trial result?
    Show answer

    No. Company updates are not peer reviewed; trial results are judged when published in peer-reviewed journals.

Glossary[4,7,8,20,22]

Brain–computer interface
A system that turns recorded brain activity into commands for a device, without using muscles.
Intracortical array
A set of fine electrodes inserted into the cortex to record the spikes of individual neurons.
Electrocorticography (ECoG)
Recording from electrodes placed on the surface of the brain.
Endovascular BCI
A BCI whose electrodes reach the brain from inside a blood vessel, delivered by catheter.
Directional tuning
A neuron's firing varying with the direction of a movement, peaking at its preferred direction.
Population vector
The sum of neurons' preferred-direction vectors weighted by their activity, which predicts movement direction.
Decoder
The algorithm that estimates intended movement or speech from neural activity.
Kalman filter
A recursive estimator that combines a model of how a state changes with noisy measurements of it.
Intracortical microstimulation
Passing small currents through cortical electrodes to evoke sensations.
Word error rate
Substitutions, deletions and insertions needed to correct a decoded sentence, divided by its length.
Locked-in
Conscious but almost completely paralysed, often unable to speak.

References

  1. Neuralink. PRIME Study Progress Update. neuralink.com (company blog, not peer reviewed) 2024. https://neuralink.com/blog/prime-study-progress-update/
  2. Neuralink. PRIME Study Progress Update — User Experience. neuralink.com (company blog, not peer reviewed) 2024. https://neuralink.com/blog/prime-study-progress-update-user-experience/
  3. Musk E, Neuralink. An integrated brain-machine interface platform with thousands of channels. Journal of Medical Internet Research 2019;21(10):e16194. doi:10.2196/16194
  4. Willett FR, Kunz EM, Fan C, Avansino DT, Wilson GH, Choi EY, et al.. A high-performance speech neuroprosthesis. Nature 2023;620(7976):1031-1036. doi:10.1038/s41586-023-06377-x
  5. Card NS, Wairagkar M, Iacobacci C, Hou X, Singer-Clark T, Willett FR, et al.. An accurate and rapidly calibrating speech neuroprosthesis. New England Journal of Medicine 2024;391(7):609-618. doi:10.1056/NEJMoa2314132
  6. Lorach H, Galvez A, Spagnolo V, Martel F, Karakas S, Intering N, et al.. Walking naturally after spinal cord injury using a brain–spine interface. Nature 2023;618(7963):126-133. doi:10.1038/s41586-023-06094-5
  7. Wolpaw JR, Birbaumer N, McFarland DJ, Pfurtscheller G, Vaughan TM. Brain–computer interfaces for communication and control. Clinical Neurophysiology 2002;113(6):767-791. doi:10.1016/S1388-2457(02)00057-3
  8. Georgopoulos AP, Schwartz AB, Kettner RE. Neuronal population coding of movement direction. Science 1986;233(4771):1416-1419. doi:10.1126/science.3749885
  9. Yuste R, Goering S, Agüera y Arcas B, Bi G, Carmena JM, Carter A, et al.. Four ethical priorities for neurotechnologies and AI. Nature 2017;551(7679):159-163. doi:10.1038/551159a
  10. Tang J, LeBel A, Jain S, Huth AG. Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience 2023;26(5):858-866. doi:10.1038/s41593-023-01304-9
  11. Metzger SL, Littlejohn KT, Silva AB, Moses DA, Seaton MP, Wang R, et al.. A high-performance neuroprosthesis for speech decoding and avatar control. Nature 2023;620(7976):1037-1046. doi:10.1038/s41586-023-06443-4
  12. Hochberg LR, Serruya MD, Friehs GM, Mukand JA, Saleh M, Caplan AH, Branner A, Chen D, Penn RD, Donoghue JP. Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature 2006;442(7099):164-171. doi:10.1038/nature04970
  13. Oxley TJ, Yoo PE, Rind GS, Ronayne SM, Lee CMS, Bird C, et al.. Motor neuroprosthesis implanted with neurointerventional surgery improves capacity for activities of daily living tasks in severe paralysis: first in-human experience. Journal of NeuroInterventional Surgery 2021;13(2):102-108. doi:10.1136/neurintsurg-2020-016862
  14. Farwell LA, Donchin E. Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials. Electroencephalography and Clinical Neurophysiology 1988;70(6):510-523. doi:10.1016/0013-4694(88)90149-6
  15. Oxley TJ, Opie NL, John SE, Rind GS, Ronayne SM, Wheeler TL, et al.. Minimally invasive endovascular stent-electrode array for high-fidelity, chronic recordings of cortical neural activity. Nature Biotechnology 2016;34(3):320-327. doi:10.1038/nbt.3428
  16. Willett FR, Avansino DT, Hochberg LR, Henderson JM, Shenoy KV. High-performance brain-to-text communication via handwriting. Nature 2021;593(7858):249-254. doi:10.1038/s41586-021-03506-2
  17. Hochberg LR, Bacher D, Jarosiewicz B, Masse NY, Simeral JD, Vogel J, et al.. Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature 2012;485(7398):372-375. doi:10.1038/nature11076
  18. Flesher SN, Downey JE, Weiss JM, Hughes CL, Herrera AJ, Tyler-Kabara EC, et al.. A brain-computer interface that evokes tactile sensations improves robotic arm control. Science 2021;372(6544):831-836. doi:10.1126/science.abd0380
  19. Georgopoulos AP, Kalaska JF, Caminiti R, Massey JT. On the relations between the direction of two-dimensional arm movements and cell discharge in primate motor cortex. The Journal of Neuroscience 1982;2(11):1527-1537. doi:10.1523/JNEUROSCI.02-11-01527.1982
  20. Wu W, Gao Y, Bienenstock E, Donoghue JP, Black MJ. Bayesian population decoding of motor cortical activity using a Kalman filter. Neural Computation 2006;18(1):80-118. doi:10.1162/089976606774841585
  21. Moses DA, Metzger SL, Liu JR, Anumanchipalli GK, Makin JG, Sun PF, et al.. Neuroprosthesis for decoding speech in a paralyzed person with anarthria. New England Journal of Medicine 2021;385(3):217-227. doi:10.1056/NEJMoa2027540
  22. Flesher SN, Collinger JL, Foldes ST, Weiss JM, Downey JE, Tyler-Kabara EC, et al.. Intracortical microstimulation of human somatosensory cortex. Science Translational Medicine 2016;8(361):361ra141. doi:10.1126/scitranslmed.aaf8083
  23. Polikov VS, Tresco PA, Reichert WM. Response of brain tissue to chronically implanted neural electrodes. Journal of Neuroscience Methods 2005;148(1):1-18. doi:10.1016/j.jneumeth.2005.08.015
  24. Collinger JL, Wodlinger B, Downey JE, Wang W, Tyler-Kabara EC, Weber DJ, et al.. High-performance neuroprosthetic control by an individual with tetraplegia. The Lancet 2013;381(9866):557-564. doi:10.1016/S0140-6736(12)61816-9
  25. Wilson BS, Dorman MF. Cochlear implants: a remarkable past and a brilliant future. Hearing Research 2008;242(1-2):3-21. doi:10.1016/j.heares.2008.06.005
  26. Lozano AM, Lipsman N, Bergman H, Brown P, Chabardes S, Chang JW, et al.. Deep brain stimulation: current challenges and future directions. Nature Reviews Neurology 2019;15(3):148-160. doi:10.1038/s41582-018-0128-2

Related readings

  • Neurons and electrical signals

    How a nerve cell makes a 100-millivolt spike, the equations that describe it, and why one neuron may be a deep network in disguise.

    Introductory

  • Neural networks, biological and artificial

    How brains and machines learn from experience, the maths both share, and why AI and neuroscience keep borrowing from each other.

    Intermediate

  • Hearing and the temporal lobe

    How the cochlea analyses sound, how auditory cortex maps frequency, speech and music, what goes wrong, and the implants and gene therapies that restore hearing.

    Intermediate

  • Touch and the parietal lobe

    How fingertips, the somatosensory map and the parietal cortex turn touch and space into action, and how touch is being restored artificially.

    Intermediate

Template anatomy for education. Not patient-specific. Not for clinical decision-making.