The phrase "brain-computer interface" gets used loosely. Depending on who's talking, it might mean Neuralink's implanted electrode arrays, a consumer EEG headset for meditation apps, or a research rig at a university neuroscience lab that lets a paralysed patient type text by imagining hand movements. These are all BCIs in a technical sense, but they sit at completely different points on the capability-invasiveness spectrum, and they work through fundamentally different mechanisms.
Having spent years working on neural signal processing and BCI-adjacent health-tech projects, I find myself explaining the same gaps repeatedly — gaps between what investors expect, what press coverage describes, and what the actual engineering constraints allow. This piece is an attempt to close some of those gaps from a technical standpoint.
What a BCI Is Actually Measuring
The brain generates electrical activity continuously. Neurons communicate through electrochemical signals — when a neuron fires, there's a rapid change in ion concentrations across its membrane, producing a measurable voltage spike called an action potential. The spike itself lasts roughly one millisecond. In a region like motor cortex, hundreds to thousands of neurons may fire in coordination when you intend a movement.
A BCI records some version of this electrical activity and translates it into an output — a cursor position, a character typed, a robotic arm trajectory, a prosthetic limb command. The quality of that recording, and therefore the quality of the eventual decoded output, depends almost entirely on how close your electrode is to the neurons you care about.
This is the fundamental physical constraint that every BCI system has to work around. Neural signals attenuate extremely rapidly with distance. An electrode sitting a few micrometres from a neuron's cell body can record that cell's action potentials clearly. Move the electrode one millimetre away and the signal from that same neuron is buried in noise from everything else in the vicinity.
The Electrode Spectrum: From Scalp to Cortex
This distance constraint explains why BCI methods range so dramatically in what they can record.
EEG — Electroencephalography
EEG places electrodes on the scalp surface. The signal it records is not individual neuron activity — it's the summed electrical field from large populations of synchronised neurons, filtered through the skull, dura mater, and various tissue layers. By the time a neural signal reaches a scalp electrode, it's been averaged across perhaps millions of cells and attenuated to the microvolt range (typically 10–100 µV for the signals you care about).
What EEG gains in non-invasiveness it pays for in spatial resolution and bandwidth. You can detect broad frequency patterns — theta, alpha, beta, gamma bands — and use them to infer mental states. Motor imagery BCIs based on EEG work by detecting changes in the 8–30 Hz range (mu and beta rhythms) that occur when a person imagines moving a limb. These systems can achieve single-axis cursor control and simple command selection, but they're slow, require significant training, and their performance degrades quickly in noisy environments.
ECoG — Electrocorticography
ECoG places electrode grids directly on the cortical surface, beneath the skull but above the brain tissue. This is a neurosurgical procedure and is currently used mainly in epilepsy patients who are already having craniotomies for other reasons — the implant is temporary during their monitoring period. The proximity improvement over EEG is substantial. ECoG can resolve signals from patches of cortex a few millimetres wide, can follow frequencies into the high-gamma range (70–200 Hz), and shows much better signal stability over time than scalp electrodes.
Research groups have used ECoG to decode speech from imagined vocalisation, reconstruct continuous hand trajectories, and achieve typing rates that approach practical usability. The limitation is obvious: it requires opening the skull.
Intracortical Arrays — Utah Array, Neuropixels, Thread Electrodes
At the invasive end, penetrating electrode arrays go into the brain tissue itself. The Utah array — 96 electrodes arranged on a 10×10 grid, each penetrating roughly 1.5mm into cortex — is the most clinically used design and the one deployed in the BrainGate trials and early Neuralink human studies. At this depth, individual neurons can be isolated and recorded — a capability called single-unit recording.
Single-unit data is qualitatively different from EEG or ECoG. You're watching the spike trains of identified cells, not population averages. This allows decoding of intended movements with enough precision to control a robotic arm in three dimensions, including grip, at speeds and dexterities that approach natural movement in idealised conditions.
The closer the electrode to the neuron, the better the signal — but also the greater the immune response, tissue displacement, and long-term instability. This tradeoff is not solved. It is the central challenge of implantable BCI engineering.
Signal Processing: From Voltage to Intent
Recording the signal is only half the problem. Converting it into a useful output requires a processing pipeline with several distinct stages, and each stage introduces potential failure modes.
Amplification and Filtering
Neural signals are small — in the microvolt to low-millivolt range — and sit alongside noise sources that can be orders of magnitude larger: powerline interference at 50/60 Hz, movement artifacts from electrode cables, electromyographic (EMG) noise from nearby muscles, and electronic noise from the amplifier hardware itself. Front-end amplifiers need very high input impedance and very low noise figures; the analog front-end design is a specialised discipline in its own right.
Bandpass filtering separates the frequency ranges of interest. Spike activity sits above ~300 Hz; local field potentials are below ~300 Hz; EEG features are typically below 100 Hz. The choice of filter cutoffs, filter types (Butterworth vs. Chebyshev vs. Bessel), and filter order all affect the downstream signal quality and introduce phase delays that matter for real-time applications.
Spike Detection and Sorting
For intracortical recordings, the next step is identifying when individual neurons fire. Spike detection usually works by threshold crossing — when the filtered signal exceeds some multiple of the noise floor, a spike event is recorded. The threshold is typically set at 3–5 times the RMS noise, which catches most real spikes while keeping false positive rates manageable.
Spike sorting is the harder problem. Multiple neurons near a single electrode produce spikes with different shapes — different amplitudes, widths, and waveform morphologies depending on cell size and electrode-cell geometry. Sorting algorithms try to cluster these waveforms to identify which spikes came from which unit. Traditional methods used principle component analysis and manual curation; newer methods use template matching and variational inference. Both have failure modes, and ground truth for sorting quality is essentially impossible to obtain in vivo.
Decoding: From Neural Activity to Commands
The decode step is where machine learning enters the picture. Given a stream of neural activity — whether spike rates, LFP power in frequency bands, or raw waveforms — you want to predict a continuous state (cursor position, limb velocity) or discrete intent (which character to type, which direction to move).
For continuous decode, the most widely used method is population vector decoding and its variants, including the Kalman filter and Wiener filter. These linear methods work surprisingly well for low-dimensional targets like 2D cursor position. For higher-dimensional targets — a full arm with seven degrees of freedom — nonlinear methods and recurrent neural networks have shown improvements, but at the cost of interpretability and training data requirements.
The recalibration problem is severe and underappreciated. Neural representations drift. The same neuron that strongly encodes rightward reaches on Monday may shift its tuning curve by Wednesday. Chronic implants see gradual changes over months as the electrode-tissue interface evolves. Decoders trained on one session degrade on subsequent sessions unless periodically recalibrated — and recalibration requires the user to perform training tasks, which is not always feasible for severely paralysed patients.
Latency, Bandwidth, and Real-World Performance
End-to-end latency in a BCI system accumulates across several steps: signal acquisition, analog-to-digital conversion, transmission, spike detection, decoding, and output rendering. State-of-the-art research systems achieve round-trip latencies of 15–50 milliseconds, which is sufficient for many applications but still longer than the natural sensorimotor loop (typically under 100ms for practiced movements, but much faster for reflexes).
Information throughput is measured in bits per second. The practical upper bound for current clinical intracortical BCIs is roughly 4–5 bits/second for communication tasks — enough for text entry at maybe 40 characters per minute under optimal conditions. For comparison, a typist using a keyboard works at 200–400 characters per minute. The gap is real, and closing it requires either denser electrode arrays (more independent neural channels), better decoding algorithms, or both.
The Clinical Translation Problem
Laboratory demonstrations often look more capable than they are, because research environments optimise for maximum performance at the expense of real-world robustness. A demo recorded in a controlled lab setting, with a well-rested expert user, a recently recalibrated decoder, and a research team on hand to adjust parameters in real time, will outperform the same system in a home environment.
Clinical translation requires solving several additional problems: device reliability over years (not weeks), wireless power and data transmission that doesn't require transcutaneous connectors, biocompatibility of chronic implants, regulatory pathways for medical devices, user training protocols that work for patients with limited motor or cognitive capacity, and maintenance protocols that don't require hospital visits.
The companies and institutions working at this layer — BrainGate consortium, Neuralink, Synchron, Blackrock Neurotech, and several university groups — are doing genuinely hard engineering. The progress is real. The timelines in press coverage are usually optimistic.
Where We Are Now and What Comes Next
The current state of BCI technology is roughly this: intracortical systems can restore basic communication and limited motor function in people with severe paralysis. Non-invasive systems are practical for very simple commands and biofeedback applications. The gap between these poles is the active research frontier.
Several developments are worth watching. Higher-density electrode arrays are moving from research tools to near-clinical devices — Neuropixels, originally developed for rodent neuroscience, is now being adapted for human use and can record from hundreds of neurons simultaneously. Fully wireless, battery-free implants are moving through early human studies. Closed-loop neurostimulation, where the BCI both reads and writes to neural circuits, is showing promise for both motor rehabilitation and psychiatric applications.
The field is moving, sometimes faster than the press coverage, sometimes slower. The honest answer about when fully natural-dexterity prosthetic control or fluent speech decode will be clinical realities is: we don't know yet. But the scientific foundation is deeper and more solid than it was ten years ago, and the engineering teams working on the hard problems are among the most skilled in medical device history.
At Neurivvy Intelligenx, we work at the signal processing and software end of this stack — neural signal decoding pipelines, EEG acquisition systems for non-invasive monitoring, and clinical data infrastructure. If you're building in this space and want to talk through what the engineering constraints actually look like, we're happy to have that conversation.