Technical Deep Dives

Brain-Computer Interfaces: How Neural Implants Work and Where They Stand

How neural implants read intention, why invasiveness buys resolution, and where the field stands after a 2026 speech BCI ran 3,800 hours in a participant's home at 99% word accuracy.

Gabriele Masetti · 2026-08-19

The basic problem a BCI has to solve

A brain-computer interface has one job: read electrical activity generated by neurons and turn it into a signal a machine can act on, or push a signal into the brain that neurons can interpret. Everything else — the surgery, the electronics, the machine learning — is in service of that single translation problem, and the translation is harder than it sounds because neurons don't produce anything resembling clean data.

A single cortical neuron fires a spike lasting about a millisecond, and the information a BCI needs is usually encoded in the statistical pattern of many neurons firing together, not in any one spike. Population activity in the motor cortex — the strip of tissue running over the top of the brain that plans and executes movement — correlates with things like intended arm direction, speed, and grip force.

Decode that population pattern accurately enough, in real time, and a paralyzed person can move a cursor, a robotic arm, or a cursor-based keyboard using intention alone. That correlation, first characterized rigorously in non-human primates in the 1980s and 1990s, is the scientific foundation nearly every modern motor BCI still builds on.

Invasive vs. non-invasive: a spatial resolution trade

The defining design choice in any BCI is how close the electrodes get to the neurons, because that distance sets a hard ceiling on signal quality.

Non-invasive EEG electrodes sit on the scalp and pick up voltage fluctuations that have already passed through skull, cerebrospinal fluid, and tissue layers that attenuate and blur the signal — a phenomenon sometimes called volume conduction. The result is signals in the range of roughly 2 to 100 microvolts, low spatial resolution (an electrode averages activity from millions of neurons at once), and heavy contamination from muscle and eye-movement artifacts.

EEG-based BCIs are genuinely useful — for detecting steady-state visual evoked potentials to drive a wheelchair interface, for instance, or for P300-based spellers — but they cannot resolve individual neurons or support the high-bandwidth control needed for fluent handwriting-speed typing or dexterous robotic-arm control.

Intracortical implants place microelectrodes directly into brain tissue, close enough to isolate spikes from individual or small clusters of neurons. That proximity is why the highest-performing BCIs to date — the ones that let a paralyzed person type at tens of words per minute or move a robotic arm with several degrees of freedom — are invasive. The cost is a craniotomy or comparable brain surgery, chronic risks (infection, hemorrhage, tissue scarring around electrodes that degrades signal over months to years), and a much slower, more expensive regulatory and manufacturing path.

Between these two extremes sit minimally invasive approaches: electrodes threaded through blood vessels or placed on the brain's surface (rather than penetrating it), trading some signal fidelity for a less traumatic procedure. Several of the companies below occupy exactly this middle ground.

The decoding pipeline: from spikes to intention

Once electrodes capture raw voltage traces, a BCI has to convert that into something usable, and this is where signal processing and machine learning do the heavy lifting.

The pipeline typically runs: spike detection and sorting (identifying when and from which electrode a neuron fired, and — in higher-fidelity systems — which neuron it was), feature extraction (often binning spike counts or extracting power in specific frequency bands over short time windows, e.g. 20–50 ms), and then a decoder that maps those features onto the variable of interest — cursor velocity, phoneme identity, intended word.

For decades the workhorse decoder was the Kalman filter, a linear recursive estimator that models how neural firing rates relate to movement kinematics and updates its estimate of cursor position or velocity on every new sample. Refined variants (such as the "ReFIT" Kalman filter, which retrains using intention-corrected training data) meaningfully improved on the naive version and were used in BrainGate consortium demonstrations of point-and-click cursor control and robotic-arm reaching.

More recently, groups have shown that nonlinear decoders — particularly recurrent neural networks such as LSTMs, trained on multiple days of multielectrode recordings — outperform Kalman filters on standard BCI benchmarks like bits-per-second in cursor tasks. The tasks that require the largest performance jump — decoding attempted speech, which involves rapidly changing, high-dimensional articulatory intent rather than a single smooth trajectory — have leaned even more heavily on deep learning, typically an RNN or transformer-style sequence model trained to map neural activity directly to phonemes or text, often combined with a language model to disambiguate likely words.

None of this is science fiction: it's supervised learning on a hard, noisy time-series problem, made tractable by dense multi-electrode arrays and enough labeled training data collected while a participant attempts to speak or move.

Speech decoding: from lab demonstration to two years of home use

Two Nature papers from 2023, both peer-reviewed and coming from separate labs using different hardware, set the terms for everything that followed in speech BCIs.

Willett, Kunz, Fan et al. (Stanford, published in Nature, August 2023) reported a speech BCI using intracortical microelectrode arrays implanted in a participant with ALS-related paralysis. Their system decoded attempted speech directly into text at a rate the authors describe as a substantial advance over prior intracortical speech decoders, using a recurrent neural network decoder combined with a language model, and it represented the first demonstration of spike-based (rather than surface-based) intracortical recording used for speech-to-text decoding at that performance level. Frank Willett led the effort at Stanford, working within a group that has long included Krishna Shenoy and Jaimie Henderson.

Metzger et al. (UC San Francisco, published in Nature, 2023) took a different hardware approach — high-density electrocorticography (ECoG) arrays resting on the surface of the speech cortex, not penetrating it — in a participant with severe limb and voice paralysis from a brainstem stroke. Their system decoded three parallel outputs simultaneously: text (a median rate of 78 words per minute with a median word error rate around 25%), synthesized speech audio in a voice resynthesized to resemble the participant's pre-injury voice, and animation of a facial avatar for non-verbal communicative gestures such as expressions. It was the first system to drive a talking, expressive digital avatar directly from neural activity.

Both results were genuine milestones and both were supervised research demonstrations. Metzger's median word error rate of roughly one word in four was the figure people cited to argue that speech BCIs were not yet communication aids anyone could rely on without a researcher present, and decoder drift was the reason to doubt they soon would be: neural signals shift as electrode positions move and tuning properties change, so a decoder trained on one day's recordings degrades on later ones and needs recalibration.

That objection has now been answered for at least one system. Nicholas Card, David Brandman, Sergey Stavisky and colleagues at UC Davis, working within the BrainGate2 trial, reported in Nature Medicine in 2026 on an intracortical speech BCI used by a 45-year-old man with ALS-related paralysis and severe dysarthria for more than 3,800 hours at home with no researchers present.

Over that period he produced 183,060 sentences — 1,960,163 words — at an average rate of 56 words per minute, drawing on a 125,000-word vocabulary. In controlled copy tasks the system reached 99.2% word accuracy. During his own independent use, 92% of sentences were decoded at least mostly correctly, a lower figure because unscripted speech is harder than copying.

The hours matter more than the accuracy. Decoder stability is not solved as a general problem — every implanted array still drifts, and the literature on "plug-and-play" recalibration-free decoding still describes an open question. But a system has now stayed useful across roughly two years of daily life in a participant's home, which the field had not previously shown.

Where the companies actually stand

Neuralink received FDA investigational device exemption approval in May 2023 for its PRIME study, evaluating the N1 implant — 1,024 electrodes distributed across 64 flexible polymer threads, inserted by the company's R1 surgical robot — in adults with quadriplegia from spinal cord injury or ALS. The first participant, Noland Arbaugh, was implanted on January 28, 2024.

Early in the trial roughly 85% of his implanted threads reportedly retracted from their intended position, reducing signal channels; Neuralink says it addressed the degradation through software (signal-processing) changes rather than repeat surgery, and Arbaugh has continued using the device — reported as using it extensively for tasks including gaming, computer control, and, in later demonstrations, robotic-arm interaction — without a second implant.

On January 28, 2026, two years to the day after Arbaugh's surgery, Neuralink said it had 21 trial participants enrolled across three countries and two continents, at Barrow Neurological Institute, the Miami Project to Cure Paralysis, University Health Network in Toronto and Cleveland Clinic Abu Dhabi. The company reports no serious device-related adverse events across the cohort and says it has seen higher signal quality in 18 of the 20 implants that followed the first. It has also opened a separate trial, VOICE, aimed at restoring real-time speech for people with severe speech impairment from conditions such as ALS or stroke, with a stated target of conversational speed at 140 words per minute.

Synchron takes the minimally invasive route: its Stentrode is delivered through the jugular vein and expanded inside a blood vessel overlying the motor cortex, avoiding open brain surgery entirely. Its COMMAND study — the first FDA-approved IDE trial of a permanently implantable BCI — reported that all six enrolled patients met the primary safety endpoint (no device- or procedure-related serious adverse events causing death or increased disability) over a 12-month evaluation period, with successful vascular deployment and motor-cortex-region coverage in 100% of cases.

Synchron has also demonstrated Stentrode users operating consumer software, including a publicly reported case of a participant controlling an Amazon Alexa device through the BCI. Synchron's approach trades some signal resolution (it reads population-level activity through a blood vessel wall rather than penetrating cortex) for materially lower surgical risk. A $200 million Series D announced on November 6, 2025 funds a 2026 pivotal trial — the study Synchron must clear before filing the first premarket approval application for an implantable BCI.

Precision Neuroscience received FDA 510(k) clearance on April 17, 2025 for its Layer 7 Cortical Interface, a film thinner than a human hair carrying 1,024 microelectrodes that rests on the brain's surface without penetrating it, inserted through a sub-millimeter incision. The clearance covers temporary use (up to 30 days), primarily for applications like intraoperative brain mapping during tumor surgery; the company reports testing the device in dozens of patients undergoing brain surgery for other reasons, distinct from a chronic, at-home BCI product. Precision Neuroscience positions itself explicitly as a less-invasive, easier-to-remove alternative to penetrating electrode arrays.

Blackrock Neurotech makes the Utah Array (also called the NeuroPort array), the 96–100-channel penetrating microelectrode grid that has been the workhorse research tool of intracortical BCI science since it was first implanted in a human, Matt Nagle, in 2004 as part of the original BrainGate trial run out of Brown University.

It remains the only FDA-cleared high-channel-count microelectrode platform capable of single-neuron recording, and it underlies most of the peer-reviewed academic BCI literature, including the BrainGate consortium's ongoing multi-site trials and Stanford's speech-decoding work. Blackrock's MoveAgain system received FDA breakthrough device designation in 2021 for translating cortical activity into external device control for people with paralysis.

What "state of the art" actually means today

Strip away the marketing and the state of the field is this: intracortical BCIs can give a person with paralysis real-time, continuous cursor and text-entry control at rates competitive with or approaching typical computer use for short, supervised sessions; can decode attempted speech into text, synthesized audio, and avatar animation at good but clearly imperfect accuracy; and can do this with a safety record — across the trials named above — that has so far avoided serious device-related harm in the small cohorts studied.

Minimally invasive and surface approaches (Synchron, Precision Neuroscience) are trading some of that performance ceiling for meaningfully lower surgical risk and, in Precision's case, reversibility.

What remains unresolved is durability and scale, not proof of concept, and the durability question is now better answered than it was. The UC Davis result shows one intracortical speech system holding up across thousands of hours of unsupervised home use. That is a single participant on a single device, which is the correct amount of weight to put on it. Electrode arrays still degrade — through scar tissue formation, thread migration, or signal drift — over a timescale of months to a few years, which is why decoder recalibration and hardware longevity remain active engineering problems.

None of the human trials described here has yet reached the patient numbers, follow-up duration, or regulatory pathway (a full FDA premarket approval for a chronic implant, as opposed to an investigational exemption or a 510(k) for temporary use) that would constitute an approved, commercially available medical product for home use. The honest summary is that BCIs have crossed convincingly from "theoretically possible" to "demonstrably works in carefully monitored clinical trials" — a major and hard-won threshold — while remaining well short of a mature, at-scale medical technology.

The engineering constraints that still bind

Three constraints shape nearly every design decision in the field. First, bandwidth versus safety: more electrodes and denser sampling generally improve decoding accuracy, but denser, more invasive arrays also raise surgical and chronic tissue-response risk, which is exactly why the field now spans a spectrum from EEG to Utah arrays rather than converging on one approach.

Second, wireless power and data: a clinically viable implant has to transmit large amounts of neural data (or a compressed, decoded version of it) through the skull without a percutaneous cable, which limits either bandwidth or battery life. Third, decoder generalization: a decoder trained on one day's recordings degrades on subsequent days as electrode positions shift and neural tuning properties drift, which is why "plug-and-play" stability — a decoder that keeps working without daily retraining — has been treated in the literature as a distinct research problem rather than an incidental detail. The 2026 home-use result is the strongest evidence yet that it is tractable on a specific array and a specific task, which is not the same as being solved across hardware, users and applications.

Progress on any one of these axes tends to trade against the others, which is a large part of why the field is populated by several companies making different bets — penetrating versus surface versus endovascular electrodes, wired versus wireless telemetry, linear versus deep-learning decoders — rather than a single dominant design.