Brain-Computer Interfaces: What Electrodes Can and Cannot Read

The phrase "brain-computer interface" invites a specific misunderstanding: that an electrode placed in the brain gives access to thought the way a microphone gives access to speech. It does not. What an electrode measures is voltage fluctuation caused by ion movement across nearby cell membranes. Turning that into a cursor movement or a sentence requires statistics, calibration, and a great deal of care about which part of the brain you are listening to.

The clinical results of the past few years are nevertheless remarkable, and worth understanding precisely, because precision is what separates genuine progress from the version of this field that appears in headlines.

What is actually recorded

Neurons communicate with brief electrical events — action potentials, or spikes, lasting about a millisecond. An electrode positioned within roughly 100 micrometres of a cell body can resolve individual spikes from that cell. Slightly farther away, contributions blur into multi-unit activity: the summed spiking of an indistinct local population. Farther still, only slower field potentials remain, reflecting synchronised synaptic currents across many thousands of cells.

Each recording modality trades spatial resolution against invasiveness:

• Intracortical microelectrode arrays penetrate the cortex a millimetre or two and resolve single-neuron and multi-unit activity. Highest information rate, highest surgical risk, most severe long-term stability problems.
• Electrocorticography places an electrode sheet on the cortical surface, beneath or above the dura. It cannot resolve single neurons but captures local field activity with good stability, and is already familiar to neurosurgeons from epilepsy monitoring.
• Stereo-electroencephalography inserts depth electrodes along trajectories, sampling sparsely but reaching deep structures.
• Scalp electroencephalography records from outside the skull. Entirely non-invasive, millisecond timing, but the skull and scalp act as a spatial low-pass filter, smearing signals across centimetres.
• Functional near-infrared spectroscopy and functional MRI measure blood-oxygen changes, an indirect proxy for neural activity delayed by several seconds — useful for mapping, poor for real-time control.

The information-rate ceiling of each modality explains most of the field's structure. You cannot decode fluent speech from scalp EEG for the same reason you cannot read a book through frosted glass: the physical measurement discards the spatial detail the task requires.

A brain-computer interface does not translate thought. It learns a statistical mapping between recorded population activity and a behavioural variable the user is deliberately attempting to produce.

How a decoder works

Consider the canonical motor case. In the primary motor cortex, individual neurons are broadly tuned to movement direction: a given cell fires most strongly for movement toward some preferred direction and less strongly away from it. No single neuron specifies the movement, but the population does. Take the firing rates of a few hundred neurons, weight each by its preferred direction, and sum — the resulting population vector points approximately where the arm is being directed. This population-vector logic, established in animal studies in the 1980s, is the conceptual foundation of motor decoding.

Practical systems replace hand-built tuning models with fitted ones. A calibration session asks the user to attempt movements, or to imagine them, while the system records activity and learns a mapping from neural features to intended velocity — typically with a Kalman filter or a recurrent neural network. The user then closes the loop: they see a cursor move, adapt their own neural strategy to the decoder's behaviour, and the decoder adapts back. This co-adaptation is essential, and it is why performance improves over sessions rather than being fixed by algorithm quality alone.

Speech decoding follows the same logic with a different target variable. Rather than decoding words semantically, high-performing systems decode articulatory intent from sensorimotor cortex — the neural commands that would move lips, tongue, jaw, and larynx — and map those to phonemes. A language model then constrains phoneme sequences into plausible words and sentences, exactly as an acoustic speech recogniser does. Recent studies with participants who have lost the ability to speak have reported decoding at rates in the tens of words per minute from large vocabularies, with word error rates low enough for practical conversation. That is a clinical transformation for someone previously typing one letter at a time.

Note what the decoder relies on: the user actively attempting to speak. It is not extracting inner monologue, and systems designed to decode attempted speech generally do not produce output when the participant merely thinks in words without attempting articulation. Some research has probed inner speech specifically, and the ability to distinguish attempted from imagined speech is being studied partly as a privacy safeguard.

The stability problem

The unglamorous obstacle that has limited intracortical interfaces for two decades is that recordings degrade.

Several mechanisms contribute. Inserting a rigid silicon shank into soft tissue causes mechanical injury and disrupts the blood-brain barrier. Microglia and astrocytes respond, forming a glial sheath around the implant that physically displaces neurons and increases electrical impedance. Micromotion between a skull-anchored device and a brain that moves with pulse and respiration causes ongoing irritation. Electrode materials corrode; insulation delaminates; connectors fail.

The result is a slow attrition of usable channels. Some long-term participants have retained useful control for many years, which proves durability is achievable, but yield across implants is variable in a way no clinical product can tolerate.

Three engineering directions address this. First, mechanical compliance: flexible polymer probes with bending stiffness far closer to tissue, inserted with temporary stiffeners or by robotic threading, provoke a smaller chronic response. Second, size reduction: thinner probes displace less tissue and appear to attract less encapsulation. Third, avoiding penetration entirely — surface arrays and endovascular electrodes, the latter delivered through a blood vessel adjacent to motor cortex using catheter techniques borrowed from stroke treatment, accepting lower signal resolution in exchange for a far less invasive procedure.

Wireless power and telemetry matter as much as electrodes. Percutaneous connectors — a physical pedestal through the skin — are a persistent infection risk and a practical barrier to living normally with an implant. Fully implanted systems with inductive power and radio links remove that risk but constrain the power budget, which in turn constrains how many channels can be sampled and how much processing happens on the implant.

What the technology is genuinely for

The strongest clinical case is restoration of communication and control for people with severe motor impairment: late-stage amyotrophic lateral sclerosis, brainstem stroke, high cervical spinal cord injury. For these users, the comparison is not against a keyboard but against eye-tracking at a few words per minute, or nothing.

Adjacent applications are progressing at different rates. Sensory feedback — stimulating somatosensory cortex or peripheral nerves to produce touch sensation, closing the loop on prosthetic control — measurably improves grasp performance and, according to participants, the sense that the limb is theirs. Spinal cord stimulation paired with intention decoding has enabled volitional walking in participants with spinal injury, effectively bridging the lesion electronically. Closed-loop deep brain stimulation, which senses a biomarker of pathological activity and stimulates only when needed, is already used in some movement-disorder and epilepsy devices and represents the most commercially mature form of bidirectional neural interface.

Cognitive enhancement in healthy people, by contrast, remains speculative. There is no demonstrated method for writing structured information into cortex. Stimulation can produce crude percepts — flashes of light from visual cortex, tingles from somatosensory cortex — but arbitrary high-bandwidth input would require knowing the neural code for the content being written, which we do not.

Honest limits

Several constraints deserve to be stated flatly, because they bound what is plausible for the coming decade.

Sampling is minuscule. A high-channel-count array records from hundreds to a few thousand sites in a cortex containing billions of neurons. Decoding works because motor and speech areas have low-dimensional, redundantly encoded population structure. Higher cognitive functions do not obviously share that property.

Localisation matters more than channel count. An array in the wrong place decodes nothing useful regardless of resolution, and functional anatomy varies between individuals, which is why implantation is typically preceded by mapping.

Calibration drifts. Neural populations recorded on a given day are not identical to yesterday's. Decoders need recalibration, and reducing that burden — through unsupervised alignment of neural manifolds across sessions — is an active and practically important research area.

Regulatory timelines are long. Implanted active devices in the brain occupy the highest risk class. Multi-year trials with small cohorts are the norm, and correctly so.

The ethical questions that are actually load-bearing

Public debate tends toward mind-reading scenarios that current physics does not support. The nearer-term issues are more mundane and more pressing.

Neural data recorded for decoding also contains information the user did not intend to share — markers of fatigue, attention, mood, and in some cases seizure risk. Data governance for continuous intracortical recordings has no established precedent comparable to, say, medical imaging.

Device longevity raises a duty-of-care question with no clean answer: if a company ceases to support an implanted system, a user may lose a capability they have integrated into daily life, and explantation carries its own risk. Several participants in early trials have described precisely this anxiety.

Agency and attribution become genuinely complicated when a decoder contributes to the output. If a language model completes a sentence the user did not fully specify, whose sentence is it? For speech neuroprostheses this is not a thought experiment; language models are essential to current accuracy.

Finally, access. These systems are expensive, surgically delivered, and require specialist support. A technology that restores communication only to those near a handful of academic centres has solved a smaller problem than it appears to.

Where progress will show up first

Watch for three specific developments rather than dramatic demonstrations. First, multi-year channel-yield data from flexible and thin-film arrays — durability, not peak performance, is the gate to a product. Second, decoders that maintain accuracy for weeks without recalibration, which is what turns a laboratory session into a usable device. Third, fully implanted wireless systems used at home rather than in a laboratory, which is where the real constraints on daily use become visible.

The field's trajectory is best understood as neuroprosthetics rather than telepathy: careful, incremental restoration of specific lost functions, achieved by measuring a small, well-chosen part of the brain very well. That is a narrower ambition than the popular framing. It is also, for the people it helps, considerably more valuable.