Brain-Computer Interfaces and AI: The Science Behind Neural Prosthetics

On January 28, 2024, surgeons implanted a Neuralink device in a 29-year-old man named Noland Arbaugh, who had been paralyzed from the shoulders down after a diving accident. Within weeks he was moving a computer cursor by imagining the movement. The story made headlines, but it sits on top of three decades of academic work. Researchers at Brown University, Stanford, and the BrainGate consortium built the recording methods and the machine-learning decoders that make brain-computer interfaces (BCIs) usable today. About 5.4 million Americans live with some form of paralysis, according to a Christopher & Dana Reeve Foundation study from 2013, so the stakes are large.

Educational content, not medical advice. Talk to your doctor before changing any medication, treatment, or routine. Information here reflects current research but every person's situation differs.

Key takeaways

• Implanted BCIs have decoded imagined handwriting at 90 characters per minute and attempted speech at 62 to 78 words per minute, each in a single study participant.
• Machine-learning decoders do the heavy lifting: the electrodes only capture raw electrical activity, and a trained model turns it into letters, words, or cursor motion.
• Scalp EEG is safe and cheap but much blurrier. The distance between electrode and neuron largely sets how much information a BCI can extract.
• Chronic signal stability, meaning how long an implant keeps recording useful data, is the central unsolved engineering problem. Most published performance results come from between one and a handful of participants per study.

How neural signals become commands

Every voluntary movement starts as electrical activity in neurons. Cells in motor cortex fire when you plan to move your hand. Cells in a strip of cortex near the top of the brain's side control the lips, tongue, jaw, and larynx. A BCI records some of that activity and translates it into a command for an outside device, such as a cursor, a robotic arm, a speech synthesizer, or a set of electrodes that stimulate the person's own muscles.

The easiest way to picture the recording problem is a stadium during a match. Scalp EEG is like standing outside the building with your ear against the wall: you can tell when the crowd roars, but you cannot pick out a single voice. An electrode array pushed into the cortex is like placing microphones among the seats. You hear individual conversations, but only from the few rows near each microphone.

Researchers group recording methods by how close they get to the tissue. The trade-off is always signal quality against surgical risk.

Approach	Where electrodes sit	What it records	Main advantage	Main drawback
Scalp EEG	On the scalp	Blurred, population-level voltage	No surgery	Skull smears the signal; low bandwidth
Endovascular (Stentrode)	Inside a blood vessel next to motor cortex	Local field potentials through the vessel wall	No open-brain surgery	Lower resolution than penetrating electrodes
Electrocorticography (ECoG)	Surface of the brain, under the skull	Field potentials from patches of cortex	Better signal than EEG, less scarring than penetrating arrays	Requires craniotomy
Intracortical arrays	Penetrating the cortex	Spikes from single neurons or small groups	Highest-resolution signal	Surgery, scarring, signal loss over time

Hans Berger recorded the first human EEG in 1924, so the idea of reading electrical brain activity from outside the head is a century old. What changed in the last twenty years is the combination of implanted electrodes that listen to individual neurons and computing power that can learn what those neurons mean.

Why decoding is the hard part

Recording is only half the job. The relationship between neural firing and intended action is different in every person, and it drifts from day to day. Early decoders used linear filters such as the Kalman filter, a method borrowed from navigation systems that estimates a hidden state (here, intended cursor velocity) from noisy measurements. These worked well enough for moving a cursor on a screen, but they treat the brain as a simple mixing board where each neuron contributes a fixed share to each movement.

Real motor cortex behaves more like an orchestra than a mixing board. The meaning of one neuron's firing depends on what thousands of others are doing and on where the movement is in its plan. Modern decoders therefore use recurrent neural networks (RNNs), which keep a running memory of recent inputs, and in some labs transformer models. If you want the background on how those architectures work, the explainer on neural networks and deep learning covers the basics. The decoder learns the pattern of neural activity as a movement unfolds over a few hundred milliseconds and maps it to a continuous output such as velocity, or to a discrete one such as a letter or phoneme.

Phonemes are the smallest units of sound that distinguish words, such as the "b" and "p" in "bat" and "pat." Speech BCIs usually decode phonemes first and then use a language model to turn the phoneme stream into likely words. That step is conceptually similar to the autocomplete on a phone, and the same family of techniques described in how large language models work helps clean up the decoder's mistakes.

Why decoders need constant recalibration

Neural recordings are not stable. An electrode that picked up a clear cell on Monday may pick up a different cell, or nothing, on Thursday, because the brain moves slightly inside the skull and tissue around the electrode changes. Researchers call this non-stationarity. Imagine a radio that drifts off its station every few days: you would keep retuning it, and the cost of retuning is time spent not using the radio.

Several groups have tried to cut that cost. A 2020 paper in Nature Biomedical Engineering by Degenhart and colleagues showed that aligning low-dimensional summaries of neural activity (called neural manifolds, which capture the patterns of correlated activity across many neurons) kept a cursor under control in monkeys even after sudden, severe changes to the recordings. The method did not need to know what the animal intended, and it outperformed supervised recalibration. It is an animal result, but it points at a practical goal: a decoder that retunes itself without asking the user to repeat a calibration task.

The BrainGate story: decades of progress

BrainGate is a consortium that includes Brown University, Massachusetts General Hospital, Stanford, and other partners. It has run clinical trials of intracortical BCIs since the early 2000s, mostly using the Utah array, a small grid of silicon needles that penetrates the surface of the cortex. The arrays in the key papers below had 96 recording channels.

The milestones are worth laying out in order, because each one added a capability.

Year	Result	What it showed	Source
2006	Neural cursor in a man with tetraplegia	A 96-electrode array still recorded movement-related signals about three years after spinal cord injury; he opened simulated email and operated a TV	Hochberg et al., Nature
2012	Reach and grasp with a robotic arm	Two participants controlled a robotic arm in three dimensions; one, implanted about five years earlier, drank from a bottle	Hochberg et al., Nature
2017	Brain-controlled muscle stimulation	A man with a C4 spinal cord injury reached and grasped with his own arm, driven by cortical signals and electrical stimulation of his muscles	Ajiboye et al., The Lancet
2021	Handwriting BCI	90 characters per minute from imagined handwriting	Willett et al., Nature
2023	Interim safety report	14 adults implanted between 2004 and 2021; no device removals required	BrainGate feasibility study, Neurology

The 2012 result is the one most people remember: a participant using a robotic arm to bring a bottle to her mouth and drink. What the paper actually established is narrower and more important. Useful three-dimensional control was possible years after the injury, which meant motor cortex kept producing intention signals even when no movement followed.

From robotic limbs to the person's own arm

The 2017 Lancet paper took a different route. Instead of steering a robot, the system sent the decoded intention to 36 electrodes implanted in the participant's upper and lower arm, which stimulated his own paralyzed muscles. The participant was 53 years old, with a complete C4 spinal cord injury. He drank from a mug in 11 of 12 attempts in one session, roughly 15 months after the arrays were placed, and fed himself about two years after implantation.

This is the design philosophy behind many current projects: treat the BCI as a bridge across a damaged connection. The brain still plans the movement and the muscles still work. What is broken is the wiring in between, and the BCI plus muscle stimulation reroutes around the break.

The handwriting result

In 2021, Frank Willett and colleagues at Stanford published a different kind of demonstration. A participant with paralysis from spinal cord injury imagined writing letters by hand, and a recurrent neural network decoded the motor cortex activity into text. He reached 90 characters per minute with 94.1% raw online accuracy, rising above 99% offline with a general-purpose autocorrect. The authors argued that handwriting works well because each letter produces a distinctive, temporally rich pattern of neural activity, which is easier to tell apart than the point-to-point movements used in cursor control.

That insight links neatly to the rest of neuroscience. The brain's motor system encodes sequences more distinctively than static targets, which is one reason scientists who study brain mapping and neural circuits are interested in what these arrays reveal beyond their use as prostheses.

Speech BCIs: from tens to hundreds of words per minute

Speech restoration is where the field moved fastest between 2021 and 2025. People who cannot speak because of ALS (amyotrophic lateral sclerosis, a disease that destroys motor neurons) or brainstem stroke often rely on eye tracking or small residual movements to spell out words, which is far slower than talking.

Two papers published back to back in Nature in August 2023 set the benchmarks, and they are often confused with each other because both reported speed in words per minute.

Study	Recording method	Participant	Speed	Error rate
Willett et al., Stanford (Nature, 2023)	Microelectrode arrays in speech motor cortex	ALS	62 words per minute	9.1% on a 50-word vocabulary; 23.8% on a 125,000-word vocabulary
Metzger et al., UCSF (Nature, 2023)	High-density ECoG on the surface of speech cortex	Severe paralysis after brainstem stroke	Median 78 words per minute	25% median word error rate

The Stanford authors described their result as the first demonstration, to their knowledge, of large-vocabulary decoding of this type. The UCSF paper decoded the attempted silent speech into three outputs at once: text, synthesized speech personalized to the participant's pre-injury voice, and a facial avatar. Both teams reached high performance with a limited amount of training, in the UCSF case under two weeks.

The two figures are easy to swap. The 62-word-per-minute result comes from the Stanford microelectrode study, and the 78-word-per-minute median comes from the UCSF surface-recording study.

Accuracy that approaches daily use

A 2024 report on a man with ALS and severe dysarthria (slurred, effortful speech caused by weak muscles) pushed accuracy much further. According to the preprint version, the system reached 99.6% accuracy on a 50-word vocabulary after 30 minutes of training. After expanding to 125,000 words and about 1.4 more hours of training, accuracy was 90.2%. It later sustained 97.5% accuracy for more than eight months, and the participant used it in self-paced conversation for more than 248 hours. The peer-reviewed version appeared in the New England Journal of Medicine in August 2024.

Those numbers matter because the first speech systems were laboratory demonstrations in which an error rate of one word in four was acceptable. Sustained use over months at around 97% accuracy is a different kind of result: it is what a person needs to use the device at home.

Voice synthesis and inner speech

Two 2025 papers extended the idea in different directions. A Nature paper by Wairagkar and colleagues described a system that decodes activity from 256 microelectrodes in a man with ALS and produces synthesized speech with immediate audio feedback, so he hears his own voice as he speaks. The system also decodes paralinguistic features such as intonation, and the participant used it to change pitch and sing short melodies.

A Cell paper by Kunz and colleagues asked a more delicate question: can a BCI decode speech that is only imagined? Using recordings from four participants, the authors found that inner speech is clearly represented in motor cortex and that imagined sentences could be decoded in real time. Inner speech resembles attempted speech, but a distinct "motor-intent" signal separates the two. That distinction allowed the authors to describe methods that stop a speech BCI from decoding private inner speech unintentionally, which addresses a real worry about mental privacy.

The result does not mean these systems read minds in a general sense. They decode signals from a small patch of motor cortex that was trained on a specific person's attempts to speak specific sentences, and they need that person's cooperation to work.

Synchron's Stentrode: reaching the cortex through a vein

Any brain implant that requires opening the skull carries risks of bleeding, infection, and neurological injury. Synchron's Stentrode takes a different path. It is a mesh of electrodes folded into a catheter, delivered through the jugular vein to the superior sagittal sinus, a large vein that runs along the top of the brain beside motor cortex, and then expanded against the vessel wall like a cardiac stent.

The device records through the vein wall, which gives a blurrier signal than penetrating electrodes but avoids open-brain surgery. The first-in-human results were published in JAMA Neurology in March 2023 as the SWITCH study. Five people with severe bilateral upper-limb paralysis were enrolled, four with ALS and one with primary lateral sclerosis, and four were included in the analysis, all men with a mean age of 61. Over 12 months of follow-up the authors reported no serious adverse events, no vessel blockage, and no device migration. Mean signal bandwidth was 233 Hz and stayed stable. All four patients used the system to control a computer.

The trade-off is plain in those numbers. A bandwidth of a few hundred hertz is enough to detect a small set of attempted movements and use them as clicks, which is useful for texting or browsing. It is not enough to decode handwriting or fluent speech. Synchron's route is therefore best understood as a low-risk way to give a basic digital switch to people who would not accept a craniotomy, not a competitor to high-density arrays on raw performance.

Neuralink: more channels, a surgical robot, and limited peer review

Neuralink's design differs from the Utah array in density and in how it is installed. The N1 implant has 1,024 electrodes spread across 64 flexible threads, and a surgical robot inserts the threads into cortex while avoiding blood vessels. That is roughly ten times the channel count of a Utah array. Several of those details, such as the thread count, come from company materials rather than peer-reviewed papers.

The regulatory timeline is documented in news reports. The FDA rejected the company's first human trial application in 2022 and approved a first-in-human study in May 2023. The first implant was placed on January 28, 2024, and Noland Arbaugh, then 29, was introduced to the public in March 2024. He has a spinal cord injury at the C4-C5 level from a diving accident.

In May 2024 the company disclosed that a number of the implant's threads had pulled out of the brain tissue in the first participant, reducing the usable signal. Software changes to the decoding restored much of the lost performance. The episode is a good illustration of the stability problem discussed below. The brain moved more than the engineers had planned for, and the fix came from the decoder rather than from the hardware.

Direct comparison with BrainGate is difficult. The BrainGate results cited in this article come from peer-reviewed papers with described methods. The Neuralink figures that we cite come from company statements and press coverage, which are useful but not independent evidence. Readers who see a bits-per-second number for cursor control should check whether it comes from a published study, a company blog post, or a demonstration video, and whether the same task and metric were used in the comparison.

The company has also announced additional trial participants and expansion into other applications. For a conceptual overview of how the technology works and how it differs from other implants, see the separate explainer on Neuralink and brain-computer interfaces.

Sensation feedback: closing the loop

First-generation BCIs were open-loop. They decoded intention and sent a command, but the user received no touch information back through the neural interface. The user could watch the robotic hand, but could not feel what it was holding. Try picking up a paper cup while your fingertips are numb: you can see your hand, but you crush the cup or drop it, because the feedback loop that normally adjusts your grip is cut.

Bidirectional BCIs add a stimulation channel. In a 2016 study in Science Translational Medicine, Flesher and colleagues electrically stimulated the hand region of somatosensory cortex in a person with a long-standing spinal cord injury. The stimulation produced touch sensations that the participant could locate on the hand, in a layout that followed the brain's expected body map. Many of the sensations felt natural, including a sense of pressure. They occurred at low stimulation levels and stayed consistent for months, and changing the stimulation strength changed how intense the sensation felt.

Together with a motor decoder, that finding sets up closed-loop control: the participant intends to grasp, the decoder moves the hand, sensors on the robotic fingers measure contact, and stimulation delivers the contact back to the brain. Whether that improves practical tasks enough to justify two implants is a question for continuing clinical trials, and the field is still working out how to make the artificial sensations consistent and comfortable over years.

For readers curious about what stimulation does at the cellular level, the explainer on how neurotransmitters work describes the chemical signaling that electrical stimulation replaces or imitates, and the piece on neuroplasticity explains why the brain can learn to interpret an artificial touch signal at all.

The signal degradation problem

The most persistent engineering problem in implanted BCIs is that the brain treats an electrode as a foreign object. Tissue around a penetrating electrode reacts with glial cells (the brain's support cells), forming a scar that can distance the electrode from the neurons it is supposed to hear. Researchers believe glial scarring is a likely cause of electrode failure, though the mechanisms are still debated, and leakage of the blood-brain barrier at insertion has been proposed as a driver of the inflammation.

Recent measurements put numbers on the problem. A 2025 study in Acta Biomaterialia examined 980 microelectrodes from eleven arrays that had been implanted in three people with tetraplegia for between 956 and 2,130 days (about 2.6 to 5.8 years). It compared platinum-tipped and iridium-oxide-tipped electrodes and linked physical damage to recording and stimulation results. Iridium oxide electrodes were about twice as likely to record neural activity as platinum ones, and the authors suggested that erosion of the silicon shaft speeds up damage at the interface.

A separate 2025 modeling paper in the Journal of Neural Engineering estimated tissue strain caused by micromotion, the tiny movements of the brain relative to the skull, around Utah arrays. Edge and corner electrodes showed higher strain than interior ones, and strain correlated with electrode performance. The authors suggest that this micromotion may worsen scarring and degrade performance over time.

What the safety record shows

A 2023 interim safety report in Neurology covered the BrainGate feasibility study. It included 14 adults implanted between 2004 and 2021, with an average of 872 days of implant time per participant and 12,203 total days. There were 68 device-related adverse events, including six device-related serious adverse events, but none required removal of the device, and there were no intracranial infections or deaths attributed to the device. The most common problem was skin irritation around the connector on the head. The authors concluded that the safety profile was comparable to other chronically implanted medical devices.

That is encouraging, but it describes 14 people. It cannot rule out rarer complications, and the system tested had a connector that passes through the skin, which most designs for home use aim to avoid.

Different engineering routes

Several approaches are being pursued in parallel, and none has been shown to solve the problem across years in large groups of people.

• Surface and fully implanted ECoG. A 2025 Journal of Neural Engineering paper followed one person with spinal cord injury who used a fully implanted ECoG-based BCI at home. The data covered 54 months, 40 of them in home or community settings, with average use of 38 minutes a day and a decoder AUROC (a measure of how well it separates movement intent from rest) of 0.959.
• Flexible electrodes. Probes that match brain tissue's stiffness aim to reduce the mechanical irritation that drives scarring. Most of the supporting data is in animals.
• Better decoders. Methods such as manifold alignment, described above, try to keep control stable as individual recordings shift.
• Endovascular delivery. The Stentrode avoids penetrating the brain at all, at the cost of resolution.

The ECoG result is a useful counterweight to the intracortical story: lower resolution, but stable decoding for years in one person at home.

What people get wrong about BCIs

Misconception 1: BCIs read thoughts. Current systems decode a narrow set of signals from a specific brain region after being trained on one person. A handwriting decoder knows nothing about what the participant is thinking in general. The Kunz paper on inner speech is the closest challenge to this point, and its authors responded by building ways for the user to keep inner speech private.

Misconception 2: consumer EEG headsets are weak versions of the same thing. They share the physics of recording electrical activity, but the performance gap is large. Scalp electrodes record blurred population signals through bone and skin, so consumer devices detect broad states such as relaxation or a small number of trained mental commands. They are not a path to typing or speech.

Misconception 3: a high bits-per-second figure means a better device. Throughput depends on the task, the participant, the decoder, and whether the figure was measured in an independent study. Comparing a cursor-control result with a speech result, or a lab session with home use, is like comparing a sprint with a marathon.

Misconception 4: results from one participant generalize. Nearly every study cited here has between one and a handful of participants, all highly selected and motivated. Performance varies, and the first person to try a system is not a sample of the population that might eventually use it.

What this means for the next decade

The pattern across the research is consistent. Decoders improve fast, because they are software and benefit from every advance in machine learning. Hardware improves slowly, because it has to survive in living tissue for years. The practical bottleneck has therefore moved from "can we decode intention?" to "can we keep recording it?"

Open questions include how well a decoder trained on one person's attempts to speak transfers to other people, how to design consent for a device whose function changes with software updates, and who maintains an implant if its manufacturer stops supporting it. A 2026 analysis in the journal Device argued that conventional consent frameworks do not fit implantable BCIs well, because these systems combine neurosurgery, continuous neural data capture, adaptive decoding, and functions that change over time.

FAQ

Q: What is a BCI and how does it work at a basic level? A: A brain-computer interface records electrical signals from neurons, either through implanted electrodes or scalp EEG. Software then decodes those signals to infer intended actions and converts them into commands for a cursor, a robotic arm, a speech synthesizer, or electrodes that stimulate muscles.

Q: What is the difference between invasive and non-invasive BCIs? A: Invasive BCIs place electrodes in or on the brain, which gives much better signals but requires surgery. Non-invasive systems such as EEG use scalp electrodes and avoid surgery, but the skull blurs the signal and limits performance. Endovascular devices such as the Stentrode sit between the two.

Q: How fast can current BCIs decode speech? A: In two Nature papers from August 2023, a Stanford group reported 62 words per minute using microelectrode arrays in a person with ALS, and a UCSF group reported a median of 78 words per minute using surface ECoG in a person with brainstem stroke. Both rates are well below the pace of ordinary conversation, so a gap remains.

Q: Is it safe to have a BCI implanted? A: The best long-term data come from the BrainGate feasibility study: among 14 participants there were 68 device-related adverse events and six serious ones, and none required removal of the device. That is a small sample, and any implant carries surgical risks that a physician should discuss with a patient.

Q: What is the biggest obstacle to BCIs becoming widely available? A: Chronic stability is the most-cited problem. Tissue reacts to implanted electrodes and recordings drift, so devices lose signal or need recalibration. Other obstacles include surgical risk, cost, the small size of trials, and the need to tailor decoders to each person.

Q: Can a BCI read private thoughts? A: No general-purpose system exists. Current decoders work on narrow signals, such as attempted handwriting or attempted speech, after training on one individual. A 2025 Cell study found that inner speech can be decoded from motor cortex in some conditions, and described methods to prevent unintended decoding of private inner speech.

Q: Are there consumer BCIs? A: Yes, but they use scalp EEG, which has far lower bandwidth than implants. They suit simple applications such as relaxation feedback or a few trained commands. They cannot reproduce the typing, speech, or limb control reported in implant trials.

The takeaway

Brain-computer interfaces work because motor cortex keeps producing clear intention signals long after the body stops responding, and machine-learning decoders have become good at reading those signals. The remaining problems are physical and organizational: electrodes that degrade in living tissue, small trials, and few independent comparisons between devices. When you read a new BCI headline, check three things: how many people were studied, whether the result was sustained over months, and whether it was published with methods other researchers can scrutinize. Those details predict whether a result will turn into a clinical product better than any speed number does.

Sources

• Hochberg et al., "Neuronal ensemble control of prosthetic devices by a human with tetraplegia," Nature, 2006: doi.org/10.1038/nature04970
• Hochberg et al., "Reach and grasp by people with tetraplegia using a neurally controlled robotic arm," Nature, 2012: doi.org/10.1038/nature11076
• Ajiboye et al., "Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia," The Lancet, 2017: doi.org/10.1016/s0140-6736(17)30601-3
• Willett et al., "High-performance brain-to-text communication via handwriting," Nature, 2021: doi.org/10.1038/s41586-021-03506-2
• Willett et al., "A high-performance speech neuroprosthesis," Nature, 2023: doi.org/10.1038/s41586-023-06377-x
• Metzger et al., "A high-performance neuroprosthesis for speech decoding and avatar control," Nature, 2023: doi.org/10.1038/s41586-023-06443-4
• "An accurate and rapidly calibrating speech neuroprosthesis," medRxiv preprint, 2023: doi.org/10.1101/2023.12.26.23300110
• Wairagkar et al., "An instantaneous voice-synthesis neuroprosthesis," Nature, 2025: doi.org/10.1038/s41586-025-09127-3
• Kunz et al., "Inner speech in motor cortex and implications for speech neuroprostheses," Cell, 2025: doi.org/10.1016/j.cell.2025.06.015
• Mitchell et al., SWITCH study of the Stentrode, JAMA Neurology, 2023: doi.org/10.1001/jamaneurol.2022.4847
• Flesher et al., "Intracortical microstimulation of human somatosensory cortex," Science Translational Medicine, 2016: doi.org/10.1126/scitranslmed.aaf8083
• Degenhart et al., "Stabilization of a brain-computer interface via the alignment of low-dimensional spaces of neural activity," Nature Biomedical Engineering, 2020: doi.org/10.1038/s41551-020-0542-9
• "Interim Safety Profile From the Feasibility Study of the BrainGate Neural Interface System," Neurology, 2023: doi.org/10.1212/WNL.0000000000201707
• "Degradation of 980 intracortical microelectrodes in three humans," Acta Biomaterialia, 2025: doi.org/10.1016/j.actbio.2025.02.030
• "5-year follow-up of a fully implanted BCI in a spinal cord injury patient," Journal of Neural Engineering, 2025: doi.org/10.1088/1741-2552/adc48c
• Christopher & Dana Reeve Foundation, paralysis statistics: christopherreeve.org
• Neuralink timeline and company-reported details: Wikipedia, Neuralink

---

Read next:

• Neural Networks Decoding the Brain: How AI Reads Neural Activity
• Brain-Computer Interfaces: How Neuralink and BCIs Work
• AI Against Superbugs: Machine Learning's Role in Antibiotic Discovery