Brain-Computer Interfaces: How Neural Signals Control Technology
Author: Ukrainian Psychological Hub · Published: October 12, 2026 · Editorial Policy
Brain-computer interfaces (BCIs) are systems that turn selected patterns of brain activity into commands for an external device. A working interface measures a physical neural signal, processes that measurement, infers a task-relevant variable, and produces an output such as a cursor movement, selected letter, synthesized word, or assistive action. Feedback often lets the person and decoder adapt to one another. These steps form a new channel for communication or control when ordinary muscular pathways are limited.
The most consequential applications are investigational systems that help people with severe paralysis communicate or control assistive technology. Wearable EEG systems and implanted neural recorders share an engineering principle, but they differ in signal, training, precision, risk, and evidence of clinical benefit. Wolpaw's foundational account and a recent broad noninvasive review describe the basic approach.
What is a brain-computer interface?
A BCI acquires signals associated with nervous-system activity and transforms them into computer-interpretable output without relying on the usual movement of the muscles that would perform the task. It may restore or supplement communication, assist movement, or create a new experimental interaction with technology. Brain–machine interface is a largely overlapping term. Neural interface is broader, sometimes including peripheral nerve devices and stimulation systems that do not provide brain-driven computer control.
There is no single BCI product or signal. Some devices classify choices in controlled laboratory tasks; others estimate continuous motion commands; speech interfaces may decode features of attempted articulation. The quality of a device depends on the reliability of an entire user–sensor–decoder–output system. A successful demonstration of a particular task does not establish universal interpretation of mental states or practical suitability for every person.
How do brain-computer interfaces work?
BCI operation can be broken into acquisition, preprocessing, feature representation, decoding, output, feedback, and adaptation. Each stage affects real-world reliability. Engineers may combine stages inside one computational model, but the model still requires a meaningful measured signal and a validated mapping to a specific output. Understanding what EEG, MRI, fMRI, and PET actually measure prevents many misleading claims.
Stage 1: A person produces task-related neural activity
A person attempting to move a hand, imagine a movement, attend to a visual target, or attempt to articulate a sentence recruits interacting neural populations. Some aspects of their activity vary systematically with the task. An interface can exploit these differences even when the muscles cannot carry out the movement. There is no universal electrode that simply reads intention; scientists must define the behavior or representation the system is designed to estimate.
Motor interfaces commonly target movement-related activity, while spelling systems can use stimulus-related responses involving distributed visual and attentional processes. Speech interfaces often record from sensorimotor regions involved in articulation. Signal design therefore follows the intended control problem. Neural activity that predicts an attempted movement under training conditions should not be assumed to predict arbitrary beliefs or private memories.
Stage 2: The sensor records a physical signal
Scalp electroencephalography records voltage differences produced largely by synchronized electrical currents in neural populations, attenuated and mixed by head tissues. ECoG electrodes placed on or near the cortical surface record local field activity closer to the source. Intracortical microelectrodes can record local action-potential events and field potentials near their contacts. Functional near-infrared spectroscopy measures optical changes linked to oxygenated and deoxygenated hemoglobin, an indirect hemodynamic correlate. These signals are different in biological origin and information content.
EEG provides fine timing without surgery, but its scalp recordings mix sources and are vulnerable to muscle, eye, movement, and environmental signals. Intracortical recordings may provide detailed local information but involve surgery, a restricted sampling region, and long-term device concerns. Neither method samples the entire brain. A careful treatment of scalp methods is available in EEG: How Electroencephalography Measures Brain Electrical Activity.
Stage 3: Preprocessing addresses contamination and changing signals
Recorded signals can contain electrode drift, poor contacts, electrical interference, eye blinks, facial muscle activity, body movement, and other unwanted components. Preprocessing may involve quality control, filtering, rereferencing, artifact detection, signal normalization, or spike extraction. The chosen procedure must preserve meaningful neural information rather than manufacture apparent task discrimination. The systematic review of motion artifact reduction shows why artifacts are especially important for portable EEG use.
The time window matters. More data can improve a prediction but introduce delays. Less data increases responsiveness while often reducing signal certainty. Researchers should report the full closed-loop delay from neural acquisition to an audible word or executed movement, not only the time taken by one software component.
Stage 4: A decoder estimates a task-specific outcome
A decoder can be a threshold rule, linear model, probabilistic classifier, regression algorithm, recurrent neural network, or another trained estimator. Features may include EEG rhythms, event-related potentials, spike counts, or spectral power measured over specific intervals. A classifier might choose a symbol; a regression model might estimate cursor velocity; a speech model might map neural sequences onto candidate phonemes or text.
Training generally requires a supervised relationship between measurements and a defined reference, such as an instructed target, attempted movement, or prompted sentence. A high prediction score only establishes performance under the tested conditions. Language-model output adds statistical knowledge about plausible words; such text is not automatically a verbatim record of an individual's internal speech.
Stage 5: Commands produce an effect and feedback supports learning
A decoded output might move a cursor, select an item, operate a computer, control a robotic arm, generate audio, or engage a stimulator. The device can display, speak, or otherwise signal the result to the user. Feedback enables correction and learning; adaptive algorithms may update their parameters as signal statistics or user behavior change. This interaction is often a collaboration between two adaptive systems: human neural control and engineered decoding.
Feedback can be more than a screen image. In an important study, a participant with tetraplegia used a brain-controlled robotic arm with tactile information evoked by intracortical stimulation of somatosensory cortex. Task performance improved when artificial touch accompanied visual feedback. The original experiment illustrates a bidirectional interface, while its participant-specific evidence still requires careful generalization.
Main types of BCI technology
Noninvasive EEG and other wearable approaches
Most accessible research BCIs record from outside the skull. EEG systems are common because they can monitor brain-related electrical signals with high temporal resolution and without intracranial surgery. Tasks include spelling, visual target selection, motor imagery, and control of robotic devices. Their effectiveness depends on signal quality, environment, calibration, stimulus requirements, user state, and the actual device task. He and colleagues' review surveys these advantages and limitations.
Some EEG interfaces use the P300 response associated with an attended target in an event sequence. Others use steady-state visually evoked potentials from visual stimuli modulated at identifiable frequencies. Sensorimotor rhythm systems measure activity changes associated with movement imagery. These paradigms are not interchangeable: a user who cannot rely on visual attention may need a different design from one who can, and an imagery-based method can impose different learning demands.
Cortical-surface recordings
Electrocorticography uses electrodes at the surface of the cortex or in closely related placements. Such recordings may have different spatial and spectral properties from scalp EEG and have supported high-density speech neuroprostheses. Surgical placement brings medical risk and long-term system considerations. The decision to use surface recording must be justified by the expected functional purpose and clinical evidence rather than by the assumption that more invasive always means better.
Intracortical microelectrode systems
Intracortical arrays enter brain tissue and can measure local signals associated with neuronal events. They have supported cursor, robotic, and speech applications in specialized studies. Recording is local and depends on implant location, tissue response, hardware design, and evolving signal quality. Medical teams must address surgery, infection risk, reliability, maintenance, revision, and ongoing device access.
Implant technologies also include emerging endovascular approaches, with electrodes introduced through blood vessels. Placement changes potential risk and information bandwidth. The 2025 systematic review and individual-patient meta-analysis documents a varied landscape and explicitly notes continuing challenges in portability and chronic use.
Active, reactive, and passive interfaces
An active interface uses voluntary modulation such as motor imagery. A reactive interface recognizes brain responses to an external event, such as a flashing letter. Passive systems attempt to infer task-related states, such as changes in workload, without an explicit command. Passive estimates raise pronounced concerns about contextual validity and privacy. Calling them passive does not make their measurements exact assessments of mood, attention, or character.
Which brain regions and networks do BCIs use?
Motor BCIs commonly exploit neural populations in motor and premotor regions and interactions with sensory and association networks. Speech BCIs have drawn on ventral sensorimotor cortex, which participates in attempted articulation. Sensory feedback has been delivered through stimulation in somatosensory areas. Visual spelling BCIs exploit population-level responses across perceptual and attention-related systems. The anatomy is determined by the control signal and device architecture.
Decoding information from a cortical region does not prove that this area uniquely contains a psychological function. Correlated activity can reflect task demands, feedback, context, or interconnected circuits. Neural systems are distributed and dynamic; for their broader organization, see Neural Circuits: How Brain Networks Generate Behavior.
How does a neural decoder learn, and why does it fail?
Decoder training often matches recorded signal sequences to known attempted actions or target labels. Algorithms optimize mappings and evaluate them on data that were not used for fitting. For a real-time interface, independent evaluation should also involve new trials, new sessions, and realistic tasks. Testing a model on data that leak training information can overstate its usefulness. Demonstrating neural rather than facial-muscle or eye-motion control may require dedicated artifact controls.
Signals change across days because of electrode characteristics, neural adaptation, behavior, fatigue, and environmental effects. A decoder can therefore deteriorate even when it performed well during initial calibration. Adaptive models and periodic retraining may improve control but require monitoring and safeguards. Reliability is a system property that includes the human participant, hardware, algorithms, and support services.
BCI accuracy is not a universal percentage. A classification score for a small set of choices, a word error rate for unrestricted speech, and an outcome such as independently completing a home activity represent different achievements. Appropriate evaluation includes error correction, delay, unintended activation, task completion, sustained usage, fatigue, and user-defined benefit.
What clinical applications have the strongest evidence?
BCI research is especially promising for communication and assistive control after severe motor impairment. A 2025 systematic review identified 93 human invasive-BCI studies involving 214 patients, illustrating both the extent of research and the relatively small clinical population studied. The review summarized heterogeneous task outcomes and stressed the need for standardized functional assessments, portability, chronic reliability, and long-term home use.
A dramatic laboratory performance may be an important feasibility demonstration while leaving questions about independent daily operation. Research participants may require sophisticated engineering teams, repeat calibration, externally prompted tasks, and device maintenance. Clinical benefit should be defined by meaningful function for the particular person, compared with relevant alternative assistive technologies.
Text and speech for people with paralysis
In 2023, Willett and colleagues described an intracortical speech-to-text neuroprosthesis that decoded attempted speech in a participant with paralysis. The work advanced large-vocabulary, comparatively rapid communication, while remaining a specialized clinical research demonstration with measurable decoding errors.
A separate 2023 study by Metzger and colleagues used high-density cortical surface recording to produce text, speech-related output, and facial-avatar animation in a participant with severe paralysis. Its reported median text performance included 78 words per minute and a 25% median word error rate for a defined evaluation. Both measurements matter: language output can be fast while still requiring substantial error management.
In 2025, Littlejohn and colleagues described a streaming brain-to-voice approach using dense cortical surface recordings in one clinical-trial participant. The system decoded speech-related information in 80-millisecond increments. This describes the computational update interval and should not be mistaken for the ability to infer arbitrary thoughts with 80-millisecond certainty.
Another 2025 Nature paper reported a brain-to-voice system using 256 implanted microelectrodes in a participant with amyotrophic lateral sclerosis and severe dysarthria. Closed-loop auditory feedback supported expressive synthesized voice, including modulation of intonation. This illustrates the importance of natural conversational interaction and individual voice identity; it does not establish uniform performance across people or settings.
Cursor, computer access, and robotic control
An interface can map attempted hand movements to pointer position, clicking, tablet navigation, or the trajectory of a robotic limb. Such systems may restore opportunities to communicate and interact with everyday environments, but success depends on the usability of the complete setup. A stable interface also needs safeguards against unwanted commands and graceful recovery from errors.
Robotic manipulation depends on feedback: vision alone may provide incomplete information about contact and grasping. The bidirectional tactile-feedback experiment demonstrates one technical response to that challenge. A high-performing robotic demonstration under supervision is distinct from long-term independently supported access at home.
Brain–spine systems and movement
A brain–spine interface can decode neural activity associated with intended movement and translate it into electrical stimulation of spinal circuits below an injury. In Lorach and colleagues' 2023 Nature investigation, a personalized system enabled walking in one participant with spinal cord injury under studied conditions. The observation is important clinical proof of concept and does not amount to a general cure for paralysis.
These systems combine recording, decoding, stimulation, rehabilitation, and clinical supervision. Their indications and risks differ from noninvasive EEG spelling or a consumer device. Mechanisms and efficacy must be evaluated for the complete protocol, not transferred from unrelated neurotechnology.
Can BCIs read thoughts or memories?
A BCI can classify or reconstruct aspects of specifically trained, task-related neural signals. In speech research, successful decoding depends on recordings made during attempted articulatory behavior and on mappings learned from participant-specific data. The model can use linguistic probabilities to predict likely words. None of these findings demonstrates unrestricted recovery of hidden memories, beliefs, or spontaneous inner monologue.
Decoding accuracy depends on known targets, stimulus structure, vocabulary, participant identity, training sessions, and the difference between prompted and unprompted behavior. Statements about mind reading commonly ignore those constraints. For a wider account of reconstruction, reverse inference, and media claims, see Can Brain Scans Read Minds? Neural Decoding and Its Limits.
Neural data remain sensitive even when unrestricted thought reading is impossible. Combining recordings with behavioral information may permit increasingly rich inferences. Privacy safeguards should follow realistic present capacities and foreseeable developments rather than relying on sensationalism or dismissing meaningful risks.
Is AI necessary for a brain-computer interface?
Statistical models and machine learning are common, but BCI does not require a giant neural network. Systems may use basic threshold rules, linear classifiers, adaptive regressors, or deep sequence models according to the task. Sophisticated models can help interpret complex recordings and synthesize speech, while raising questions of transparency, latency, generalization, and the source of generated content.
A decoder that produces a plausible sentence should identify how much came from measured neural evidence and how much from the language model's learned expectations. High-quality communication requires error correction and clear presentation of uncertainty. Bigger AI models cannot independently eliminate surgical risks, electrode limitations, lack of clinical access, or long-term support obligations.
How precise and dependable are current devices?
Reliability includes false-positive commands, successful tasks per hour, accuracy, latency, setup time, fatigue, calibration burden, device stability, and the ability to recover from failure. Comparing percentages across studies is hazardous when users, outcome definitions, equipment, and test conditions differ. Reporting should distinguish offline predictions from online closed-loop behavior and externally prompted trials from spontaneous tasks.
Usability also varies across individuals. The 2025 systematic review of EEG-BCI performance factors reported associations between task outcomes and characteristics such as attention, cognition, fatigue, and training, with variation by paradigm. Its findings support individualized assessment, not a deterministic claim about any person's capacity to use a BCI.
For real-world adoption, independent home operation and multi-month or multi-year follow-up are major questions. Studies need transparent reporting of technical failures, abandoned sessions, adverse effects, assistance from researchers, user satisfaction, maintenance, and continuing access.
Safety, surgery, and the regulatory status of BCIs
Implanted interfaces involve surgical and chronic-device risks that may include bleeding, infection, tissue responses, hardware failure, electrode changes, and revision procedures. Their risk profile depends on device placement, the individual's medical condition, and the research or clinical protocol. Noninvasive systems avoid intracranial surgery but can still have limitations connected with visual stimulation, fatigue, unreliable control, and misuse of inferred information.
The FDA's 2021 guidance on implanted BCI devices sets out considerations for nonclinical testing and clinical studies involving paralysis or amputation. The status of a research device is not equivalent to FDA approval or clearance for general use; regulatory decisions are device- and indication-specific.
Medical eligibility must be evaluated by qualified professionals, with comparison to existing communication and assistive tools and with explicit plans for device upkeep, adverse events, and possible study termination. No consumer EEG score should be used as an individual psychiatric diagnosis solely because a decoder produces a brain-state label.
Do BCIs treat mental health conditions or enhance intelligence?
Research into EEG feedback, stimulation, and state estimation is sometimes grouped under neurotechnology, but these methods require distinct evidence. A brain-driven interface that controls a computer is not automatically a validated treatment for depression, ADHD, PTSD, or anxiety. An algorithm can classify a recording without establishing that its output reflects a clinically meaningful biomarker or that training changes health outcomes.
Therapeutic claims need condition-specific trials, adequate comparators, clinical endpoints, adverse-event monitoring, and sustained benefits. Consumer promises of improved intelligence, mood, or concentration should be evaluated against these standards. The clinical evidence for an assistive communication interface cannot be borrowed to prove another product's mental-health efficacy.
Privacy, consent, cybersecurity, and autonomy
Neural records should be treated as potentially sensitive personal data. Risks include excessive collection, covert inference, reidentification, insecure transmission, data brokerage, secondary use, dependence on proprietary algorithms, and uncertainty over who can terminate processing. The practical safeguards include granular and revocable consent, data minimization, access controls, transparency, security audits, retention limits, and meaningful user control.
A device can also influence action because a decoder's mistakes change outputs. Interfaces should expose uncertainty, support correction or emergency stop mechanisms, and avoid irreversible actions where decoding is unreliable. Long-term implanted devices raise obligations involving software updates, service availability, manufacturer failure, and maintenance when trials end. The U.S. GAO policy assessment addresses many of these implementation problems.
The UNESCO 2025 Recommendation on the Ethics of Neurotechnology establishes an international normative framework addressing privacy, agency, human rights, and responsible research and deployment. A systematic scoping review of ethical considerations synthesizes dilemmas associated with neurotechnology and enhancement. Ethical guidance frames responsible practice; actual compliance requires independent evaluation.
Scientific limitations and remaining uncertainties
The field has robust evidence that neural activity can support task-specific technological control. Its ability to provide consistent, accessible benefits at population scale remains under study. Common limitations are small clinical samples, selection effects, difficulty replicating personalized setups, heterogeneous performance reporting, decoder drift, laboratory dependence, uncertain long-term implant performance, and resource-intensive multidisciplinary support.
A decoding result does not explain all neural mechanisms involved in the task. Algorithms can exploit correlates or confounds without isolating a causal substrate. Reverse inference from recorded activity to a complex psychological state is especially problematic. Converging evidence from anatomy, controlled experiments, perturbations, and reproducible behavioral outcomes is needed for stronger mechanistic conclusions.
Media coverage can turn a participant-specific milestone into a promise of imminent universal access. Scientific communication should provide the number of participants, exact device and brain area, outcome measures, online test conditions, error rates, support burden, adverse events, and duration. A valuable clinical advance is most responsibly described with the boundaries that make it interpretable.
What should readers ask before trusting a BCI claim?
Start with the measurement: EEG, cortical surface potentials, intracortical activity, or a hemodynamic signal? Then identify the task, the source of training labels, the size of the choice set, and whether evaluation was independent and online. Ask if eye or muscle signals were ruled out where relevant, whether the user could correct errors, and whether performance transferred beyond the training session.
For a clinical device, ask how it compares with established assistive communication or access methods, whether it provides durable daily benefit, what surgery and ongoing maintenance require, who controls the data, and how the device can be repaired or replaced. For a consumer product, distinguish research-grade validation from a marketing demonstration. Specific answers are more informative than descriptions of direct thought control.
Frequently asked questions
Can brain-computer interfaces work without surgery?
Yes. EEG-based systems record from the scalp, while some systems use other noninvasive techniques. Implanted approaches sample signals closer to neural sources and introduce medical risks. The right design depends on the task and individual needs.
How does a BCI control a computer cursor?
A decoder associates measured task-specific brain activity with cursor direction or selections. The software updates movement, displays the result, and may adapt with the user. It requires calibration and error management.
Can a person with paralysis speak through a BCI?
Some research participants with severe motor or speech impairments have communicated through text or synthesized voice. The 2023 speech study and 2025 voice-synthesis study provide primary evidence. These are specialized clinical demonstrations with participant-dependent outcomes.
Can brain implants restore walking?
Specialized experimental brain–spine systems can couple neural decoding to spinal stimulation. A 2023 human study demonstrated walking for an individual participant. Clinical generalization and long-term effects require further evidence.
How much of the brain does a BCI record?
A BCI records signals through its actual sensor contacts and measurement modality. Scalp EEG captures broad mixtures; implanted arrays sample more local activity. Neither measures the full content of mental experience.
Are BCIs accurate enough to use every day?
Some systems have impressive laboratory or supervised clinical performance. Daily reliability varies by task, user, technical setup, and support. Long-term home use and standardized functional outcome evidence remain priorities.
Can a BCI read a thought someone wants to keep private?
Existing published decoders are trained for particular tasks and conditions, and do not establish unrestricted thought reading. Neural data still warrant privacy and security protections because task-specific inferences can be sensitive.
Is a BCI the same thing as artificial intelligence?
A BCI is a biological measurement and control system. It often incorporates AI or machine learning to decode signals, but its identity depends on the brain-to-technology channel, not any one model family.
Conclusion
Brain-computer interfaces connect measured neural activity with useful external action through sensing, computation, feedback, and adaptation. Their achievements in communication and assistive control are scientifically substantial, especially for people with severe paralysis. The strongest evidence concerns particular devices, recordings, and tasks, while broader clinical availability demands more standardized trials, long-term performance, safety, accessibility, and sustained support.
Understanding the source of the signal, the limits of decoding, and the outcomes that matter to actual users provides a rigorous basis for evaluating future breakthroughs. This framework also protects against confusing powerful specialized neurotechnology with unlimited access to thought or a universal treatment.