A brain-computer interface, or BCI, is a system that reads electrical or other signals produced by the brain, translates them using software, and converts them into commands that control an external device. The device might be a computer cursor, a robotic arm, or a speech synthesizer. What makes a BCI distinct from other assistive technologies is that it bypasses the body’s normal output pathways entirely: you do not need to move a muscle, press a button, or speak a word. The signal goes straight from neural activity to machine action, and the science behind that translation has moved from crude laboratory demonstrations to real-time sentence decoding and restored grip control in paralyzed people.
Reading the Brain’s Signals
Every BCI starts with the same problem: how do you listen in on what billions of neurons are doing? The answer depends on how close you are willing to get to the brain tissue itself, and that choice shapes everything about the system’s speed, accuracy, surgical risk, and longevity.
The least invasive option is electroencephalography, or EEG, where electrodes sit on the scalp. EEG picks up the summed electrical activity of large populations of neurons through skin, skull, and several layers of tissue, so the signal is blurry. It is good enough to detect broad patterns of intent, like imagining a left-hand squeeze versus a right-hand squeeze, but it struggles with fine-grained commands. Hybrid systems that pair EEG with other non-invasive sensors can sharpen the picture. One approach combines EEG with functional near-infrared spectroscopy, which measures blood-oxygen changes in the cortex. In motor imagery tasks, combining the two reached about 89% accuracy, a few percentage points better than either technology alone.1PubMed. A hybrid BCI based on EEG and fNIRS signals improves the performance of decoding motor imagery of both force and speed of hand clenching
At the other extreme, intracortical arrays are pushed directly into brain tissue. The most widely studied is the Utah array, a tiny silicon chip bristling with around 100 hair-thin electrodes. Because each electrode tip sits among a small cluster of neurons, it can pick up individual action potentials with a signal-to-noise ratio that early testing placed at roughly 6 to 1.2PubMed. The Utah intracortical Electrode Array: a recording structure for potential brain-computer interfaces That kind of clarity lets researchers track the firing of single cells and decode very specific intentions, like the direction and speed of a planned arm movement.
Sitting between scalp EEG and penetrating arrays is electrocorticography, or ECoG, where a thin electrode grid is laid on the brain’s surface without piercing it. ECoG picks up signals that are much cleaner than scalp EEG but avoids the tissue damage of penetrating electrodes. Researchers have shown that separable speech intentions can be captured from a cortical site less than a centimeter across using ECoG, meaning the surgical opening needed is far smaller than traditional craniotomy, which lowers risk for fragile patients.3PubMed Central. Using the Electrocorticographic Speech Network to Control a Brain-Computer Interface in Humans
Turning Neural Activity Into Commands
Raw brain signals are noisy. Even with the best hardware, the electrical trace from any electrode is a messy mix of the neural activity you care about, background firing from other neurons, electrical interference, and biological noise from muscle contractions or heartbeats. Before a BCI can figure out what you intend to do, it has to clean the signal, pull out the features that carry information, and feed those features into an algorithm that maps them to specific outputs.
For non-invasive systems, that processing chain is especially demanding because the signal has been filtered and smeared on its way through the skull. Feature extraction strategies and classification algorithms need to be tuned specifically to handle these low-amplitude, noisy recordings. For intracortical systems, the raw data is richer, but the volume is enormous: hundreds of channels, each streaming thousands of voltage samples per second.
The decoding step itself increasingly relies on deep learning. Rather than hand-crafting rules about which neural patterns correspond to which actions, modern BCIs train artificial neural networks on large amounts of recorded data. The network learns to recognize the relationship between patterns of brain activity and the intended action, and it can generalize to new situations the user has not explicitly practiced.4PubMed Central. Towards neural co-processors for the brain: combining decoding and encoding in brain-computer interfaces The result is a system that improves over time, both because the algorithm refines its model and because the user’s brain adapts to the interface.
What BCIs Can Do Right Now
The applications that get the most attention fall into two broad categories: restoring movement and restoring communication. Both have progressed from proof-of-concept demonstrations to systems that work in people’s homes, though none is yet as seamless as natural function.
Movement Restoration
For people with spinal cord injuries, BCIs can reconnect the brain’s movement commands to the limbs by routing them through a robotic arm or through electrical stimulators strapped to the person’s own muscles. A pooled analysis of studies using intracortical BCIs found that neuromuscular-stimulated systems, which activate the person’s paralyzed muscles directly, achieved roughly 84% accuracy on three-dimensional reach-and-grasp tasks, while robotic arm systems reached about 69%.5PubMed. Intra-cortical brain-machine interfaces for controlling upper-limb powered muscle and robotic systems in spinal cord injury The neuromuscular approach also tended to be faster, although there was wide variability between participants. Neither system is fast or accurate enough to replace able-bodied movement, but both represent a meaningful gain in independence for someone who cannot move their arms at all.
Speech Decoding
Speech BCIs aim to let people who have lost the ability to speak produce words or sentences from brain activity alone. One recent system used a 256-channel micro-ECoG array to decode the full set of Mandarin Chinese syllables, a task requiring the system to distinguish among 394 distinct sounds. It achieved a median syllable identification accuracy of about 71% in a single-character reading task and demonstrated real-time sentence decoding.3PubMed Central. Using the Electrocorticographic Speech Network to Control a Brain-Computer Interface in Humans That error rate is still too high for casual conversation, but combined with language-model correction (the same kind of predictive text your phone uses), it gets much closer to usable fluency.
Stroke Rehabilitation
BCIs are also being used not as permanent prostheses but as temporary training tools to help stroke survivors regain movement. The idea is that when a patient imagines moving a paralyzed hand and the BCI detects that intention, it triggers a robotic brace or electrical stimulator to complete the movement. Over weeks of practice, that repeated pairing of mental effort and physical result can drive the brain to rebuild damaged motor pathways. A meta-analysis of controlled trials found that BCI-based training produced a medium-to-large effect on upper limb motor recovery compared to conventional therapy, with a standardized mean difference of 0.79.6PubMed Central. Brain‐computer interfaces for post‐stroke motor rehabilitation: a meta‐analysis Several of the BCI groups crossed the threshold considered clinically meaningful on standard motor assessments, while fewer control groups did.7PubMed Central. Brain-Computer Interfaces for Stroke Motor Rehabilitation
Two-Way Traffic and the Sense of Touch
Early BCIs were strictly one-directional: brain signals went out, commands came back as visual feedback on a screen. But controlling a robotic hand by watching it on a monitor is a bit like trying to tie your shoes while wearing thick gloves. You can see what your fingers are doing, but without the feel of the lace between them, every movement is slow and imprecise.
Bidirectional BCIs solve this by sending information back into the brain. A second set of electrodes, placed in the somatosensory cortex (the region that processes touch), delivers tiny electrical pulses that the person perceives as sensations on their hand or fingers. In one study, adding this tactile feedback to a robotic arm controlled via motor cortex recordings improved the user’s ability to control grasp force compared to vision alone.8PubMed Central. Intracortical Microstimulation Feedback Improves Grasp Force Accuracy in a Human Using a Brain-Computer Interface Another demonstrated that supplementing vision with evoked tactile percepts improved overall robotic arm control.9PubMed Central. A brain-computer interface that evokes tactile sensations improves robotic arm control The sensations are not identical to natural touch, but they are informative enough that the brain can use them to fine-tune grip strength and object handling.
How the Brain and the Machine Learn Each Other
One of the more surprising aspects of BCI research is that performance does not depend solely on better algorithms or better hardware. The brain itself changes in response to using the interface. Neurons that were not originally involved in a task can be recruited, and firing patterns shift to become more distinguishable for the decoder. Meanwhile, the decoder can be retrained on the fly to keep up with the brain’s evolving signals. This mutual adjustment, where both the biological and artificial sides adapt, is called co-adaptation, and it turns out to be critical for long-term performance.
Adaptive closed-loop BCIs take this further by dynamically adjusting stimulation or decoding parameters in real time based on the user’s current brain state. Machine learning algorithms continuously monitor the incoming signals and tweak the system’s behavior, which helps sustain performance even as neural patterns drift over hours or days.10PubMed Central. Electroencephalogram-based adaptive closed-loop brain-computer interface in neurorehabilitation: a review Research on ECoG-based BCIs has confirmed that the combination of neural adaptation and machine learning together outperforms either alone.11Journal of Neural Engineering. Spatial co-adaptation of cortical control columns in a micro-ECoG brain–computer interface In rehabilitation settings, this co-adaptation is especially valuable because it actively promotes the kind of brain reorganization that supports recovery.
The Hardware Longevity Problem
A BCI that works brilliantly for six months but degrades by year two is not much help to someone with a lifelong condition. Long-term reliability is one of the field’s most stubborn challenges. The body treats any implant as a foreign object and mounts an immune response. Around penetrating electrodes, glial cells form scar tissue that gradually insulates the electrode tips from nearby neurons, weakening the signal over time.
Data from the BrainGate clinical trial, which tracked 20 Utah arrays across 14 participants over a span of roughly two decades, found that the average proportion of electrodes still detecting spiking activity was about 36%, with only a 7% average decline over each participant’s enrollment period.12PubMed Central. Long-term performance of intracortical microelectrode arrays in 14 BrainGate clinical trial participants That is a gradual degradation, not a cliff, and many participants maintained enough working channels to keep using their BCIs for years. But it does mean the system slowly loses resolution, and replacement surgery carries its own risks.
One promising direction is making electrodes so thin and flexible that the brain essentially ignores them. Researchers have developed nanoelectronic thread electrodes with dimensions smaller than a single cell. In animal studies, these threads formed seamless integration with surrounding tissue: capillaries grew back around them with an intact blood-brain barrier, and there was no chronic neuronal loss or glial scarring.13PubMed Central. Ultraflexible nanoelectronic probes form reliable, glial scar-free neural integration Signal quality remained stable over months, a stark contrast to rigid silicon arrays. The challenge is scaling these delicate structures for human use and proving they can last for years, not just months.
Getting to the Brain Without Open Surgery
The biggest barrier to wider adoption of high-performance BCIs is that the best ones require opening the skull. Endovascular BCIs offer a radically different approach: threading a small electrode-tipped stent through a blood vessel into the brain, much the way cardiologists place coronary stents. The Stentrode, the first such device tested in humans, was implanted in four patients with severe paralysis. The final safety data from that trial confirmed that recording neural signals from inside a blood vessel is feasible and that the procedure avoids the risks of open-brain surgery.14PubMed Central. Assessment of Safety of a Fully Implanted Endovascular Brain-Computer Interface for Severe Paralysis in 4 Patients
The signals from an endovascular device are not as detailed as those from an intracortical array. The electrode is separated from neurons by the blood vessel wall, so it picks up field potentials rather than single-unit spikes. But for many practical applications, like selecting items on a screen or controlling a powered wheelchair, that level of detail may be sufficient, and a procedure performed through a vein in the neck is far less daunting than craniotomy.
Functional ultrasound is another technology being explored as a way to read brain activity without electrodes at all. Rather than measuring electrical signals, it images tiny changes in blood flow within the brain at high spatial and temporal resolution. The approach is still in early development, but researchers see potential for a fully non-invasive BCI that offers spatial precision closer to what implanted devices achieve.
Privacy and the Uniqueness of Neural Data
As BCIs move from research labs into clinical and eventually consumer settings, a set of ethical questions has grown harder to ignore. Neural data is not like other medical data. An EEG trace or an intracortical recording can reveal information the user did not intend to share: emotional states, cognitive load, attention patterns, even implicit biases embedded in neural responses. That data is uniquely sensitive precisely because of its intimate nature.15PubMed Central. Regulating neural data processing in the age of BCIs: Ethical concerns and legal approaches
Security is a related concern. Because BCIs create a direct technical communication channel with the brain, they raise novel questions about unauthorized access and data interception that do not have close parallels in other medical devices.16arXiv. A Framework for Preserving Privacy and Cybersecurity in Brain-Computer Interfacing Applications A hacked pacemaker is dangerous because it controls the heart. A hacked BCI is dangerous because it could, in principle, read or even influence brain activity. Current research-grade systems are not connected to the internet and use closed communication links, but consumer-facing devices will eventually need robust cybersecurity frameworks that today’s regulatory landscape has barely begun to address.
Existing medical device regulations were written for hardware that measures or stimulates; they were not designed for devices that continuously generate rich, interpretable streams of cognitive data. Several legal scholars and neuroethicists have argued that neural data deserves its own category of protection, separate from and stronger than the rules governing ordinary health information. Whether that protection takes the form of new legislation, expanded data-privacy law, or device-specific regulation is still an open debate.
How Non-Invasive BCIs Combine Multiple Signals
For users who do not need or want surgery, one strategy for squeezing more information out of scalp-level recordings is to detect multiple types of brain response simultaneously. A hybrid approach tested the combination of two well-known EEG signals: the P300, a voltage deflection that appears when you notice something unexpected, and the steady-state visually evoked potential, a rhythmic response driven by a flickering stimulus. Researchers found that detecting both at the same time was feasible without one interfering with the other; the P300 response was unaffected by the flickering stimulation, and the steady-state response was unaffected by the oddball task that triggers the P300.17PLoS ONE. Simultaneous Detection of P300 and Steady-State Visually Evoked Potentials for Hybrid Brain-Computer Interface Stacking independent signals like this effectively gives a non-invasive BCI more “channels” of information without adding more electrodes.
These hybrid and multimodal systems are unlikely to match the performance of intracortical arrays for high-speed control tasks. But they represent a practical sweet spot for applications where moderate accuracy is acceptable and surgery is not. Think of neurofeedback for attention training, accessible gaming interfaces, or communication devices for people who can still use some voluntary movement but want a hands-free backup. The gap between invasive and non-invasive performance is narrowing, and for many everyday uses, the non-invasive side may already be good enough.
The Regulatory Path for Implanted Devices
Bringing an implanted BCI to market is not just a technical challenge; it is a regulatory one. In the United States, implanted BCIs are classified as medical devices and must go through the FDA’s approval process. Because the technology is genuinely new, several companies have pursued the Breakthrough Device designation, which provides more intensive FDA interaction and can speed up review without lowering the safety bar. The broader industrialization path for implanted neural regulation devices has been mapped out in regulatory roadmaps that trace the route from fundamental neuroscience discovery through clinical trial to market access.12PubMed Central. Long-term performance of intracortical microelectrode arrays in 14 BrainGate clinical trial participants
One complication is that a BCI is not just a medical device in the traditional sense. It is also, in some configurations, a data-collecting platform that continuously monitors brain activity. Regulators are still figuring out how to handle the overlap between device safety oversight (does the implant harm the brain?) and data governance (who owns the neural recordings, and what can be done with them?). The answers will shape not only which products reach patients but how much control those patients retain over the information their own brains generate.