Brain-Computer Interfaces and Neural Immersion
Linas JuozenasShare
Intelligence Unleashed · Neural interfaces
Brain–computer interfaces from neural signals to restored communication
A brain–computer interface does not extract a hidden stream of finished thoughts. It measures selected neural activity, learns a task-specific mapping and turns that estimate into an action—sometimes with feedback returning to the user. Human studies now include rapid speech decoding, long-term home communication and movement with artificial touch. Most high-performance implantable communication and general motor-control systems remain investigational.
The field in one sentence
The clinical promise is real; the generalisation is not yet earned
Research participants with paralysis have used neural activity to type, speak, control computers, move robotic and biological limbs, and receive artificial tactile information. These are genuine achievements. Most leading results still come from single participants or small feasibility cohorts using bespoke systems and intensive support.
The responsible reading is neither “science fiction” nor “mind-reading has arrived.” It is that specific functions can sometimes be restored under specific conditions, while surgery, signal stability, daily usability, evidence at scale and long-term stewardship remain open engineering and clinical problems.
Educational and research context
This guide is not medical advice, a product recommendation or an invitation to join a trial. Eligibility and risk must be assessed by qualified clinical and research teams under a current protocol. People seeking communication access should also receive an individual assessment of established augmentative and alternative communication options.
01 · Draw the boundary
What a brain–computer interface is—and is not
The label is useful only if it identifies the biological source, the intended function and the feedback loop.
In the classic sense, a BCI measures activity produced by the brain, extracts features from it and translates them into an output that can be used to communicate, control a device or alter an intervention.1 The system may only read neural activity, or it may also write information back through stimulation.
Brain signal becomes useful output
Examples include EEG selection of a highlighted symbol, intracortical decoding of attempted handwriting, or motor-cortex activity controlling a cursor. The relevant control signal originates in brain activity.
Interfaces may connect elsewhere
Spinal-cord, peripheral-nerve and muscle interfaces can be essential parts of a neuroprosthesis. They are not automatically BCIs when no brain signal drives the system. The broader word neurotechnology includes much more than BCIs.
Muscle and behavioural signals are different
Eye tracking, voice control and surface electromyography (sEMG) can be excellent interfaces, but they do not record brain activity. Meta’s wrist research, for example, decodes neuromuscular signals with sEMG; calling it a BCI obscures the actual mechanism.5
| System | What it measures | A BCI? | Why the distinction matters |
|---|---|---|---|
| EEG speller | Electrical activity at the scalp associated with attention to a stimulus or an attempted task. | Yes, when the neural signal selects or controls output. | Performance must be separated from eye movements and other artefacts. |
| Wrist sEMG band | Electrical activity generated by muscles and peripheral motor-unit activation. | No, under the usual brain-based definition. | It may feel like “intention control,” but residual neuromuscular output is the signal. |
| Eye-gaze keyboard | Eye position or movement. | No. | It is often faster, simpler and less risky than a BCI when reliable eye control remains. |
| Closed-loop brain stimulation | Neural state used to trigger or shape stimulation. | Often described as a BCI, depending on whether useful communication or control is part of the definition. | Terminology varies; authors should state exactly what is sensed, inferred and stimulated.2 |
Decoding is not unrestricted thought-reading
A decoder is trained on relationships between recorded activity and known tasks—such as attempting to say prompted sentences or imagining a hand movement. Its output is a probabilistic estimate shaped by electrodes, training data, task instructions and often a language model. That is far narrower than direct access to a person’s private inner monologue. The privacy stakes are still serious because models and uses can expand.
02 · Follow the information
A BCI is a chain, not a chip
The weakest link may be signal quality, latency, calibration, feedback, setup burden—or a user interface that does not fit the person’s goal.
Every impressive output sits at the end of a sequence. A higher channel count can help in some designs, but it does not replace appropriate placement, stable recording, robust decoding, safe stimulation or usable feedback.
-
Define the intended task
Speak a sentence, point to a target, select a symbol, open a hand or detect contact. “Read the brain” is not a testable function.
-
Acquire a neural signal
Electrodes or sensors capture voltage changes, magnetic fields or blood-oxygen responses at a particular spatial and temporal scale.
-
Clean and represent it
Processing suppresses artefacts and extracts features such as frequency-band power, field potentials or threshold-crossing activity.
-
Decode an estimate
A statistical or machine-learning model maps features to a command, phoneme, character, trajectory or state. Confidence and errors matter as much as the top guess.
-
Return feedback and adapt
The person sees, hears or feels the result and changes strategy; the software may recalibrate. In bidirectional systems, stimulation can provide tactile or other information.
The clinically meaningful unit is not the implant. It is the complete loop that a person can use safely, reliably and on their own terms.
03 · Compare access routes
Closer signals usually require greater intervention
“Non-invasive,” “surface,” “endovascular” and “intracortical” describe anatomical access—not a simple ladder from bad to good.
Interfaces trade signal scale and specificity against surgery, biological response, coverage, portability and maintenance. The best route depends on the function and the person—not on a generic bandwidth ranking.
| Route | Signal and timescale | Potential advantage | Constraint to keep visible |
|---|---|---|---|
| Scalp EEG | Summed electrical activity from large neural populations; millisecond timing. | No surgery, portable hardware and a long research history. | Signals are spatially blurred and easily contaminated by eye, facial and neck activity; setup and training can still be substantial.4 |
| fNIRS | Changes in cortical blood oxygenation; delayed haemodynamic response. | Non-ionising and can complement electrical measures. | Slow response, limited depth and sensitivity to scalp blood flow and motion. |
| MEG / fMRI | Magnetic fields or blood-oxygen signals with powerful whole-brain research measurements. | Useful for mapping and scientific decoding. | Large, expensive infrastructure; fMRI is slow and neither is an ordinary wearable control route. |
| ECoG / cortical surface | Local field activity recorded on or near the cortical surface. | Stronger, more spatially specific signals than scalp EEG; high-gamma activity can support speech decoding. | Requires cranial surgery; coverage is limited to electrode placement. “Surface” does not mean non-invasive. |
| Endovascular | Field activity recorded from electrodes deployed inside a cerebral blood vessel. | Avoids an open craniotomy and can use established neurointerventional access. | Current chronic endovascular BCI systems are permanent implants with vascular, antithrombotic, lead and signal-resolution considerations.20 |
| Intracortical arrays | Local field and action-potential-related activity close to neurons. | Fine spatial and temporal information has enabled the highest-performing motor and speech demonstrations. | Penetrates cortex; tissue response, connector design, channel stability, surgery and lifetime support are central. |
| Stimulation channels | Electrical patterns delivered to cortical or other neural tissue. | Can create artificial tactile percepts or complete a bidirectional sensorimotor loop. | The location, quality and stability of percepts must be mapped; safe charge and long-term exposure matter. |
“Minimally invasive” is not a complete risk statement
The phrase is used inconsistently. An endovascular or subdural device may avoid one surgical route while introducing another. Compare the actual procedure, duration, implanted components, medication, infection and haemorrhage risks, revision plan, removability and expected support—not the adjective.
04 · Read claims in context
A spectacular number may answer a narrow question
Speed, accuracy and “bandwidth” are not interchangeable, and results from different tasks cannot be ranked without their full task context.
The essential qualifiers are online or offline, fixed or open vocabulary, prompted or self-generated, laboratory or home, peak or average, assisted or independent, and one session or sustained use. Language-model correction can be useful, but it must not be mistaken for information supplied by the neural signal alone.
Use task-native outcomes
Word or character error rate for text; time to target and successful selections for a cursor; task completion, object handling or validated clinical scales for movement; psychophysics for sensation.
Measure the day around the demo
Setup time, calibration, care-partner work, donning and doffing, uptime, fatigue, repair, portability, environmental robustness and whether the person can choose when to use it.
Ask who benefits, and how
Number of participants and diagnoses, adverse events, duration, comparison condition, user-valued goals, quality of life and performance against safer existing access methods.
| Reported number | What to ask | Common source of inflation or confusion |
|---|---|---|
| Words per minute | Attempted speech or selected text? Average, median or peak? Were corrections included? | Comparing decoding speed with natural speech rate, or omitting errors and correction time. |
| Word / character error | Online or offline? Vocabulary size? Prompts known? Language model? Out-of-vocabulary words possible? | Quoting an offline autocorrect result as live raw accuracy. |
| Accuracy | Accuracy of what, across how many choices and trials, with what chance level? | A binary classifier at 80% and a 40-choice speller at 80% do not convey the same information. |
| Information-transfer rate | Which formula, assumptions, trial duration, errors and idle time were included? | Using a theoretical selection metric as though it were useful everyday communication bandwidth.67 |
| Electrode channels | How many were recorded, usable and stable, and where were they placed? | Treating hardware channel count as a clinical outcome. |
Use an evidence ladder, not a hype leaderboard
A company announcement, trial registration and conference demo show activity but do not, by themselves, establish efficacy. A peer-reviewed single case can support feasibility. Replication across participants, prespecified multicentre studies, comparative outcomes, regulatory marketing authorisation, coverage and supported real-world use answer progressively different questions.
05 · Locate the field
For most leading platforms, the frontier is human feasibility
Multiple implant routes are being studied. The existence of an IDE, registry entry or Breakthrough Device designation does not authorise commercial use.
As of this review, high-performance implanted communication and motor-control BCIs are chiefly research systems. Human evidence ranges from temporary intraoperative recordings to years of use by individual participants. There is no responsible basis for calling the field a mature consumer implant market.
Academic intracortical systems
BrainGate and collaborating groups have published peer-reviewed cursor and text, speech, robotic-arm, own-limb and long-term home-use studies.811162124 The registered BrainGate2 study remains an investigational feasibility study.31
Endovascular recording
Data from the four implanted participants analysed in the SWITCH study supported 12-month safety and computer-control feasibility. The U.S. COMMAND study is registered as an early feasibility study—not a pivotal approval trial.2032
New fully implanted platforms
Commercial developers are enrolling small studies of distinct electrode and telemetry designs. Neuralink’s PRIME study, for example, is registered as an early-feasibility study evaluating the N1 implant, its R1 implantation robot and initial BCI control.33 Public participant updates can indicate study activity, but they are not peer-reviewed trial results.
Breakthrough is not approval
The FDA programme is intended to speed development and review for qualifying devices. It does not itself demonstrate safety or effectiveness and is not permission to market.34
In March 2026, China granted a narrow market authorisation. The National Medical Products Administration approved Neuracle’s NEO implantable hand-movement compensation system for selected people with tetraplegia after cervical spinal cord injury. The epidural recording system decodes motor intent to operate an external pneumatic glove.52 That jurisdiction- and indication-specific authorisation is an important milestone; it is not approval for speech, general computer control or consumer enhancement.
Trial language, translated
IDE permits investigation under specified conditions. Early feasibility asks whether a design can be used safely enough to keep studying and whether the proposed function is plausible. Pivotal evidence is intended to support a marketing submission. Cleared and approved are regulatory outcomes with different pathways. None of these words should be substituted for another.
06 · Restore digital access
Cursor control, selection and imagined typing
Communication BCIs do not all decode language. Many first decode movement intent and use it to operate a conventional interface.
A neural cursor can point and click on an on-screen keyboard; a decoder can instead classify imagined handwriting or key presses. These routes differ in speed, training, error correction and the brain areas they use.
| Study and route | Participants | Online result | What it does not establish |
|---|---|---|---|
| Point-and-click typing, intracortical (2017) | Three people with paralysis. | Up to 39 correct characters per minute using a cursor and on-screen keyboard.8 | That every user will reach the peak rate, or that a research rig is ready for unsupervised daily use. |
| Attempted handwriting, intracortical (2021) | One person with spinal cord injury. | 90 characters per minute with 94.1% raw online accuracy. Greater than 99% with general autocorrect was an offline analysis.9 | That offline autocorrect accuracy was achieved as live raw output, or that the result generalises beyond this participant. |
| Attempted QWERTY typing, intracortical (2026) | Two people: one with ALS and one with spinal cord injury. | Real-time bimanual key-intent decoding; peak rates differed markedly between participants—47 and 110 characters per minute—with language-model assistance.10 | A population average, ordinary touch typing, or performance outside structured laboratory tasks. |
| Endovascular click plus eye gaze (2021) | Two people with ALS. | Neural click selection was combined with eye tracking for cursor position; reported text entry was roughly 14 and 20 correct characters per minute.19 | Pure neural two-dimensional cursor control. It demonstrates a hybrid system, which may be the more useful design. |
A hybrid interface is not a lesser BCI. If gaze, a head switch or residual movement performs one part more reliably while neural control supplies a missing click or mode change, the combination may reduce fatigue and increase autonomy. The scientific obligation is to disclose each signal’s contribution.
Home use changes the question
Laboratory throughput asks whether a decoder works. Home use asks whether a person can choose the time, topic and software; recover from errors; protect privacy; and keep the system running. In 2026, the same intracortical participant first reported by Card and colleagues in 2024 was shown to have had a BCI active at home for 3,801 logged hours over 19 months, including idle time between interactions—first using speech and, after it was added later, cursor control—following care-partner setup but without researchers present during operation.16 That is unusually strong longitudinal evidence of practicality—not independent replication—and still a one-person result with wired percutaneous connectors and a cart of research computers.
07 · Decode attempted speech
From text to a voice with timing and expression
Progress is best understood as a sequence of credible, carefully bounded results.
Speech BCIs typically learn activity associated with attempts to articulate—not complete, unconstrained thoughts. A model may first predict phonemes or acoustic features, then use language or voice-synthesis components to produce words and sound.
| Study | Interface and participant | Core result | Boundary |
|---|---|---|---|
| Willett et al. (2023) | Four 64-electrode intracortical arrays; one participant with ALS. | Online attempted-speech text at 62 words per minute; 9.1% word error with a 50-word vocabulary and 23.8% with a 125,000-word vocabulary.11 | One participant; the small- and large-vocabulary results are different conditions. |
| Metzger et al. (2023) | 253-contact high-density ECoG; one participant with anarthria after brainstem stroke. | Median 78 words per minute; about 25% word error in the 1,024-word online condition, with text, personalised sound and avatar animation.12 | The expanded 39,378-word analysis was offline; 150 words per minute was not achieved output. |
| Card et al. (2024) | 256 intracortical electrodes; one participant with ALS and severe dysarthria. | After brief calibration, 0.44% word error for 50 words on day one and 9.8% for 125,000 words on day two; later large-vocabulary performance improved further.13 | Accuracy and speed came from defined benchmarks; the vocabulary did not accept every possible out-of-vocabulary term. |
| Littlejohn et al. (2025) | High-density ECoG; the same participant as the 2023 avatar study. | Continuous streaming synthesis in 80-millisecond increments, with online text and audio rather than waiting for a complete sentence.14 | In the 1,024-word incremental evaluation, online error remained substantial; “streaming” describes latency, not perfect recognition. |
| Wairagkar et al. (2025) | 256 intracortical electrodes; the same participant as Card 2024. | Voice generated from neural features in roughly 10-millisecond increments, including attempts to alter intonation and sing short melodies.15 | A one-person feasibility study; voice quality and phoneme error are not interchangeable with text word accuracy. |
| Card et al. (2026) | Same intracortical participant, now evaluated in extended home use. | The system logged 3,801 active-at-home hours over 19 months, including idle periods, and decoded 1.96 million words. Output during decoded utterances averaged 56 words per minute. Formal prompted tests reached 99.2% word accuracy; the participant rated 92% of everyday sentences at least mostly correct.16 | Home-hour duration, utterance speed, self-rated natural conversation and prompted-copy accuracy are different measures; daily setup took about 20 minutes with care-partner help. |
Easier to inspect and correct
Words on a screen support review, deletion and conventional text-to-speech. A language model can improve likely sequences but may also insert a plausible word the user did not intend.
Preserves conversational timing
Streaming acoustic synthesis reduces turn-taking delay and may convey rhythm, pitch or emphasis. Evaluation needs intelligibility, latency, naturalness and the user’s control—not a single “accuracy” score.
A voice can be personally meaningful
Voice banking or reconstruction may help a synthesizer resemble a pre-injury voice. The user should govern when that model speaks, how it is secured and whether anyone else may clone or reuse it.
No peer-reviewed implanted speech BCI had demonstrated output above 150 words per minute by the review date
Rates around 150–160 words per minute commonly appear as a comparator for natural conversation. They must not be rewritten as achieved BCI throughput. The verified 2023 online study rates were 62 and 78 words per minute; a particularly important 2026 home study reported a 56-word-per-minute average during decoded utterances alongside 3,801 logged system-active hours.
08 · Rebuild a sensorimotor loop
Returning information can improve movement control
Motor intention can drive a cursor, robotic arm, orthosis or electrical stimulation of the person’s own muscles. In some studies, artificial touch has reduced reliance on vision.
Decoding a reach is only half of ordinary movement. Natural control continually uses touch and proprioception to correct force and position. Bidirectional research therefore links motor-cortex recording with sensors and stimulation of somatosensory cortex.
Robotic reach and grasp
People with tetraplegia have controlled multi-joint robotic arms through intracortical activity, including reaching, grasping and bringing a drink toward the mouth.2122 These landmark studies established possibility; they did not yield a plug-and-play household product.
Artificial tactile feedback
In a 2021 study with one participant, sensors on a robotic hand triggered intracortical stimulation in somatosensory cortex. Median time on an arm-function task fell from 20.9 to 10.2 seconds with tactile feedback.23 That is a compelling controlled comparison, not a population effect estimate.
Brain-controlled muscle stimulation
Intracortical signals have been routed to functional electrical stimulation to produce grasping and reaching with a participant’s paralysed arm.2425 Movement required external hardware, calibration and careful positioning; it did not show biological restoration of the disrupted pathway.
Shape, motion and location
Small studies now show that patterned stimulation can convey more than a simple “contact” signal, including localised percepts and constrained information about edge orientation or motion. The vocabulary of artificial touch remains sparse and highly individual.4849
Bidirectional does not mean “full feeling”
Electrical stimulation usually evokes a limited percept at a mapped body location. Researchers test whether it is detectable, localizable, stable and useful for a task. Natural texture, temperature, pain, proprioception and embodiment are not automatically restored.
09 · Connect brain, body and training
A bypass can assist now; paired activity may also train surviving pathways
Assistive function while a system is on and therapeutic change that remains when it is off are separate outcomes.
Some systems route decoded intention past an injury to spinal stimulation or muscle activation. When intention, stimulation and practice are paired repeatedly, researchers also investigate whether neuroplasticity can produce persistent recovery.
A digital bridge after spinal cord injury
In a 2023 proof-of-concept study, one participant with chronic incomplete tetraplegia used cortical signals to modulate epidural spinal cord stimulation and stand and walk with crutches in structured settings. The participant had already completed extensive stimulation-assisted rehabilitation, and the study had no comparator. Improvements were reported, but this was not a trial of unaided walking.26
A “double neural bypass”
A 2026 study combined intracortical motor decoding, muscle activation, an orthosis, cervical stimulation, rehabilitation and patterned cortical stimulation. With the complete bypass active, one participant performed structured tasks including self-feeding; after months of stimulation and training, selected elbow and wrist measures also improved, while fingers and thumb remained insensate to natural touch even though cortical stimulation could evoke artificial percepts.27 The specialised, multi-component programme cannot show which component caused persistent change or what a broader group would obtain.
| Outcome | Question | Example measure | Interpretation |
|---|---|---|---|
| Immediate assistance | What can the person do while the full device is active? | Successful grasp, walking distance, task time or independence level. | The system bypasses or supplements a disrupted pathway. |
| Training effect | Does performance improve across sessions with the device? | Learning curve, reduced calibration, force control or clinical motor score. | May reflect user learning, decoder adaptation, conditioning or recovery. |
| Persistent recovery | What remains when stimulation and decoding are off? | Prespecified motor or sensory measures after a washout period. | Suggests biological change, but spontaneous recovery, co-interventions and measurement bias need controls. |
Vision is a related but distinct frontier
Visual prostheses can stimulate retina or visual cortex to create spots or patterned percepts. These are not camera-quality sight. In one 2021 cortical study, a 96-electrode array implanted for six months in one participant who was blind evoked discriminable phosphenes and supported some constrained letter and boundary tasks; implantation and explantation were uncomplicated in that case.28 Current cortical programmes remain very small early-feasibility studies. Retinal devices have a different anatomy, evidence base and regulatory history, so they should not be presented as proof that a high-resolution cortical “pixel grid” is available.
One participant’s result is evidence, not an average
Single-participant studies are indispensable in rare, invasive research: they can establish mechanism, feasibility and personally meaningful benefit. They cannot estimate how often the result will occur, identify uncommon harms or settle which component caused a persistent change.
10 · Keep the accessible route realistic
Non-invasive BCIs trade surgery for signal and setup constraints
Avoiding an incision does not remove the burden—and lower neural resolution does not mean no useful application.
EEG BCIs can support selection, communication, neurofeedback, gaming research and rehabilitation protocols. Their greatest strengths are reversibility, accessibility and repeatability. Their limitations are not solved by attaching an AI label.
Detect a task-relevant response
A rare or attended stimulus evokes an event-related pattern. Repeated flashes can help select a row, column or symbol, but attention, vision and timing demands matter.
Attend to a flicker frequency
Visual cortex responds at the frequency of an attended target. Laboratory spelling by participants without the target disability can be fast, yet flicker comfort, gaze demands and performance in intended users require separate testing.
Modulate sensorimotor rhythms
Imagined or attempted movement can change EEG rhythms. Training success varies, and eye or muscle activity must be controlled rather than allowed to masquerade as neural decoding.
Use a slower vascular signal
Changes in oxygenated and deoxygenated haemoglobin can classify some tasks and complement EEG. The haemodynamic delay makes rapid control difficult.
Independent home use is possible for some non-invasive designs: a 2018 study documented home use of a P300-based system by people with ALS.41 The wider literature nevertheless shows heterogeneous protocols, small samples and a gap between healthy-volunteer classification and dependable assistive use.4
Before calling a headset a BCI
Ask whether the device uses the neural signal to produce a functionally useful output; whether the evaluation excluded eye, jaw and scalp-muscle artefacts; whether performance was online; whether the target population was tested; and whether its advertised wellness or medical claim matches its regulatory status.
11 · Separate restoration from speculation
The case for cognitive enhancement is much weaker
A technology that restores a missing route for communication does not thereby upload knowledge, expand intelligence or reveal memory on demand.
The strongest evidence concerns specific unmet functions in people with severe impairment. Non-medical experiments can adapt a game, estimate workload or let a trained user select among targets. Those demonstrations are far from a general-purpose “upgrade.”
| Claim | What exists | What remains unshown |
|---|---|---|
| Silent communication | Implanted systems can decode trained attempted speech; wrist sEMG can infer subtle motor output without using brain activity. | Unrestricted, accurate decoding of unspoken inner thought from a convenient consumer device. |
| Memory support | Adjacent research explores stimulation and closed-loop neuromodulation in clinical contexts. | A validated consumer “memory assistant” that records, uploads or retrieves autobiographical memories. |
| Adaptive media | EEG and other biosignals can sometimes classify coarse states under controlled conditions. | Reliable reading of a person’s exact emotion, preference or truthfulness across real-world contexts. |
| Cognitive enhancement | Training can improve control of the trained BCI task; assistive access can expand what a person is able to do. | A safe implant that raises general intelligence or transfers knowledge to a person without the relevant impairment. |
| Brain-to-brain exchange | Laboratory demonstrations can pass a very small, experiment-defined signal through two separate interfaces. | Shared experience, telepathy, consciousness transfer or direct high-bandwidth exchange of meaning. |
Restoration and enhancement are not a perfect moral binary
A speech BCI may restore communication and also let its user work faster than with an earlier eye-gaze system. A sensory prosthesis may create a novel kind of percept rather than reproduce biology. The ethical questions turn on purpose, proportionality, evidence, consent, alternatives, distribution and power—not on a single label.
12 · Think in failure modes
Risk belongs to the procedure, hardware, software and care pathway
There is no defensible universal “BCI complication rate” or annual percentage of signal loss.
An external EEG cap and a penetrating cortical array have different risk profiles. Even two implants differ by surgical route, location, connector, medication, stimulation, duration and the health of the intended users. Safety evidence must remain device- and protocol-specific; estimates borrowed from deep-brain stimulation or another implant can inform planning without becoming a BCI-wide complication rate.30
Procedure and tissue
Possible harms include bleeding, infection, seizure, stroke or neurological deficit, anaesthesia complications, wound and skin problems, pain, vascular occlusion or migration, and risks of revision or explantation. Which apply depends on the route.
Implant and external stack
Lead or connector damage, loss of channels, telemetry failure, battery or power problems, heating, moisture, mechanical stress, electromagnetic compatibility and unavailable replacement components can interrupt access.
Drift and wrong output
Neural activity, electrode coupling, fatigue, posture, disease and context change over time. False activation, missed intent, confident language-model substitution and updates that alter performance need monitoring and recovery paths.
| Evidence | Observed | Responsible interpretation |
|---|---|---|
| BrainGate interim profile (2023) | Fourteen adults contributed 12,203 participant-days. Investigators reported 68 device-related adverse events and six related serious adverse events; none led to explantation for safety, intracranial infection, device-related death or permanently increased disability.29 | Encouraging Class IV evidence for this percutaneous intracortical research system in carefully selected participants—not a universal rate for implanted BCIs. |
| SWITCH endovascular case series (2023) | Four implanted participants completed 12 months; no serious adverse event causing death or permanent disability, target-vessel occlusion or device migration was reported, alongside mild device-related events.20 | Early safety and feasibility in four people over one year; too small and short to exclude uncommon or late harms. |
| Fully implanted ECoG life cycle (2016–2024) | One person with ALS used a communication BCI for about seven years. Control eventually became unreliable as neural-signal amplitude declined; disease-related cortical atrophy was a leading explanation, though alternatives remained possible.1718 | Durability is a property of person, disease, target, hardware and algorithm together—not a fixed percentage lost per year. |
Loss of a useful device is itself a harm
A participant may build communication, work and relationships around an interface. Trial closure, company failure, an incompatible update or loss of technical staff can then remove more than hardware. Consent and funding should address maintenance, data access, transition to alternatives and the circumstances and costs of continued implantation or explantation.
13 · Protect the whole stack
Cybersecurity is a safety and communication problem
The attack surface extends from an implant or headset through receivers, laptops, apps, cloud services, models, support accounts and updates.
“Neural data” is not one file. A system may hold raw signals, processed features, calibration tasks, inferred commands, decoded conversations, personalised voice data, stimulation parameters, logs and ordinary account credentials. Each has different consequences if exposed or altered.
Limit observation
Collect only what the function needs; process locally where practical; encrypt data in transit and at rest; isolate research from personal accounts; restrict and log support access; and set explicit retention and secondary-use rules.
Prevent silent alteration
Authenticate devices and users, sign software and model updates, validate inputs, monitor unexpected behaviour, preserve rollback and test changes against safety and performance specifications.
Keep essential access working
Plan for power, connectivity and component failure; retain a backup communication route; publish support lifetimes; patch vulnerabilities without abruptly removing a critical function.
Make stopping easy
The user needs an accessible pause, undo and safe state. Stimulation should have hardware and software bounds, while remote operation should use least privilege and fail safely.
The FDA’s February 2026 guidance addresses cybersecurity design, labelling and premarket documentation for medical devices with cybersecurity risk, including statutory expectations for certain connected “cyber devices.” It does not create a BCI-specific device classification.35
Protect inferences, not only raw voltage
A raw signal may be difficult to interpret outside its model and task. The decoded sentence, predicted intention, personalised voice and behavioural log may be immediately revealing. Future models may also extract information that was not part of the original purpose. Governance should follow data through transformation and reuse.
14 · Keep the person in authorship
Agency is designed through correction, consent and control
A decoder can help complete an action without becoming its author. The interface must make uncertainty and automation governable.
Every BCI output is co-produced by neural activity, model assumptions, training data, user strategy and feedback. Shared autonomy can make a robotic reach smoother and a language model can reduce typing effort. Both can also create a plausible action the person did not choose.
Consent must be accessible and ongoing
Explain uncertainty, alternatives, foreseeable burdens, data uses, commercial relationships, withdrawal, post-trial access and explant choices in a format the person can use. Supported decision-making is not the same as substituting someone else’s preference.
Make intent revisable
Provide confidence cues where helpful, confirmation for high-consequence actions, rapid correction, an accessible emergency stop and a way to distinguish user-selected text from automatic completion.
Let experience change the design
Track user-defined benefit, fatigue, embodiment, frustration, privacy and care-partner effects—not only decoder accuracy. Consent to research or data reuse can be revisited as function and circumstances change.
Voice and avatar identity
A reconstructed voice or animated face can restore more than information; it can affect recognition, intimacy and self-presentation. The user should choose the voice, expressive range, audience and persistence. A system should not silently make someone sound more certain, polite or emotionally uniform than intended.
Consent when communication is severely limited
Severe motor impairment does not imply impaired understanding. Researchers must provide time, reliable yes/no methods, independent advocacy where appropriate and repeated checks of willingness. At the same time, uncertain communication demands careful validation: neither excluding a capable person nor manufacturing consent is acceptable. The current Declaration of Helsinki and disability-rights principles remain relevant to research design.4044
A person should be able to use a decoder without surrendering the right to hesitate, revise, remain silent or stop.
15 · Count the full pathway
Access is not the price of the electrode
Real access includes assessment, alternatives, surgery, equipment, training, support, connectivity, repair, upgrades and an exit plan.
A device can be technically successful and still be inaccessible because a person lives far from a study site, lacks care support, uses an underrepresented language, cannot tolerate the setup, or loses coverage for the services that make the implant functional. Public technology assessments likewise identify access, privacy, security and long-term support as policy questions—not side issues.45
| Stage | Resources and decisions | Equity question |
|---|---|---|
| Choose | Communication and functional assessment; comparison with eye gaze, switches, voice, EMG and other assistive technology; independent counselling. | Can the person make an informed choice without having to accept surgery to obtain good support? |
| Implant or fit | Specialist team, travel, hospital care, medication, accessible accommodation and management of complications. | Who is excluded by geography, language, comorbidity, trial criteria or unpaid care requirements? |
| Learn and use | Calibration, rehabilitation, external computers, mounting, power, software, connectivity and care-partner training. | Does the measured benefit subtract setup time, fatigue and the labour of other people? |
| Maintain | Monitoring, patches, replacement parts, retraining, compatible operating systems, data portability and clinical follow-up. | Who pays after a study ends, and what happens if a developer is acquired or fails? |
| Change or stop | Transition to another interface, deletion and retention choices, continued safe implantation, revision or explantation. | Is withdrawal practically possible without losing all communication or bearing unaffordable surgery? |
One model does not speak for everyone
Speech BCIs learn participant-specific neural patterns and use language resources that may encode vocabulary, dialect and domain. A one-participant bilingual ECoG study shows possibility, not solved multilingual access.46
No unsupported implant
Long-term obligations should be planned before recruitment: named responsibility, funding or insurance, minimum support periods, security maintenance, interoperable data where safe, component supply, clinical records, transition assistance and explant criteria. “The trial ended” is not a care plan.
16 · Match the rule to the use
No single “neuro-rights law” governs every BCI
Medical-device law, research ethics, privacy, cybersecurity, disability rights, consumer protection and employment rules overlap.
Governance depends on jurisdiction, intended use and setting. An investigational speech implant, a wellness EEG headband and an employer’s attention-monitoring system can involve different regulators and legal duties even when all are marketed with neural language.
| Layer | What it addresses | Boundary or misconception |
|---|---|---|
| Medical-device regulation | Safety, performance, manufacturing, clinical evidence, labelling, post-market duties and cybersecurity for a defined intended use. The FDA also issued specific 2021 guidance for implanted BCIs for paralysis or amputation.3 | An IDE or Breakthrough designation is not marketing authorisation. For example, FDA clearance of a cortical electrode for temporary use does not clear a permanent complete BCI system.53 |
| Human-participant research | Scientific validity, independent ethics review, accessible consent, risk–benefit assessment, monitoring, withdrawal and fair selection. | Consent alone does not make a poorly designed or unsupported study ethical. |
| Data and cybersecurity | Personal, health or biometric data rules where applicable; security of connected medical devices; contracts and consumer-health protections. | “De-identified” or raw neural data are not automatically harmless, and no single privacy label covers every inference. |
| Disability and human rights | Autonomy, non-discrimination, accessibility, participation, research inclusion and access to assistive technologies.40 | Protection should not become paternalistic exclusion from research or care. |
| Work, school and consumer settings | Power imbalance, voluntariness, surveillance, accommodation, discrimination, product claims and unfair practices. | A click-through “consent” does not make coercive monitoring acceptable. |
Responsible innovation
The OECD Recommendation on Responsible Innovation in Neurotechnology is an international policy instrument covering stewardship, safety, inclusivity, societal deliberation and protection of personal brain data.37 It is not a claim that every decoded output has one universal legal classification.
Ethics across the life cycle
UNESCO adopted its Recommendation on the Ethics of Neurotechnology on 11 November 2025. It addresses dignity, autonomy, privacy, identity, freedom of thought, equality, misuse and access.36 It is a global normative recommendation, not directly binding worldwide legislation.
Regional and national rules may apply differently to medical devices, AI-enabled features, personal data, employment and consumer products. Application requires a case-specific legal analysis. There is no substitute for reading the enacted instrument and the device’s intended purpose.
17 · Use milestones, not prophecy
What would make a BCI clinically ready?
The next decade should be judged by supported autonomy across people and years—not by the most cinematic demonstration.
The likely direction is more complete systems: implanted or wearable sensing, on-device processing, adaptive models, ordinary computer compatibility, multiple fallback inputs and user-controlled feedback. The decisive advances will be those that reduce the surrounding burden.
Wireless, home-centred systems
Fully implanted telemetry and simpler external hubs may reduce infection and setup burdens. Progress must be shown through uptime, independent use, battery and heat performance, security and multi-year support.
Hybrid and multimodal control
BCI, gaze, residual movement, switches, automated target selection and shared autonomy can each do the part they handle best. A useful system need not maximize the amount controlled directly from brain activity.
Richer communication
Streaming voice, intonation, bilingual models, personalised vocabulary and combined cursor control may support more natural participation—if errors remain visible and the user retains authorship.
Bidirectional function
More stable, informative patterns of cortical stimulation may improve grasp, embodiment and rehabilitation. Multi-participant, longer-term evidence is still required.
Prescription for selected users
Routine clinical access would require favourable comparative benefit, device-specific safety, reliable manufacturing, supported home use, reimbursement and a credible lifetime plan—not merely a successful early feasibility study.
General neural enhancement
Memory upload, knowledge transfer, shared consciousness and mass-market cognitive implants do not follow from present motor and speech decoding evidence.
| Milestone | A convincing demonstration would include |
|---|---|
| Generalisation | Prespecified results across a meaningful range of users, diagnoses, languages and sites—with results for participants who do not benefit also reported. |
| Durability | Years of signal, function and safety data using the final intended hardware, not a temporary or percutaneous substitute. |
| Independence | Chosen daily activities at home, setup burden and care-partner input measured, faults recoverable without a laboratory team. |
| Comparative value | User-valued outcomes against the best available alternative interface and against the risks and opportunity costs of implantation. |
| Stewardship | Patch and component commitments, portability, post-trial access, funded clinical follow-up and a company-failure plan. |
18 · Evaluate a claim
Twelve questions that cut through a BCI headline
If the answer to several is missing, the claim is not ready to carry the conclusion.
- What biological signal is used? Brain, spinal cord, peripheral nerve, muscle, eye movement or a hybrid?
- What exact function is decoded? A binary selection, cursor trajectory, prompted sentence, free text, acoustic feature, grasp or clinical state?
- How many people contributed the result? Report how many people were enrolled, implanted and analysed, and how many withdrew—not only the number labelled “patients.”
- Who were they? Diagnosis, residual movement, time since injury, sensory and cognitive requirements determine transferability.
- Was performance online? A held-out offline reconstruction is scientifically useful but is not live closed-loop control.
- What assistance was present? Researcher setup, eye tracking, manual channel selection, care-partner labour, a language model or shared autonomy?
- Which denominator and comparator were used? All trials or successful blocks; raw or corrected output; device on or off; best alternative access method?
- How long and where did it work? A peak block, repeated sessions, independent home use or multi-year follow-up?
- What went wrong? Errors, no-control false activations, fatigue, adverse events, signal loss, downtime and non-use days belong in the result.
- What is the regulatory status? Research permission, Breakthrough designation, component clearance and system marketing authorisation are different.
- What happens to data and software? Storage, model training, voice rights, remote access, updates, deletion limits and third parties should be explicit.
- Who supports the user for life? Name the responsible organisation, funding, repair, security, transition and explant plan.
A strong claim survives added context
“One participant produced 62 words per minute online with 23.8% word error in a 125,000-word condition” is more informative—not less impressive—than “AI reads thoughts at conversational speed.” Precision lets readers see both the breakthrough and the work still ahead.
19 · Quick answers, careful boundaries
Frequently asked questions
The distinctions most likely to change how a BCI claim should be understood.
Can a BCI read thoughts?
Not in the broad science-fiction sense. A trained system can infer a constrained target—such as attempted movement, an attempted phoneme, a selected symbol or a response to a flashing item—from recorded activity under particular conditions. It does not reveal every belief, memory or unspoken idea. Calling decoded attempted speech “thought reading” erases the task, training data, model errors and the user’s intentional effort.
Are all BCIs brain implants?
No. Scalp EEG and fNIRS are non-invasive; ECoG records on or near the cortical surface, above or below the dura; intracortical arrays enter cortex; and some experimental systems record from vessels near the brain. Each route trades signal detail, coverage, portability, durability and medical risk differently. A device should be described by its actual signal and placement, not simply as “neural.”
Is a wrist EMG band a brain–computer interface?
Not under a brain-signal definition. Surface electromyography records muscle activity, including very small residual motor signals at the wrist. It can make an excellent hands-free or low-effort interface, but it is a neuromotor or muscle–computer interface rather than direct brain recording.5 Naming the signal accurately helps users compare options.
What is the fastest communication BCI?
There is no fair one-number league table. Studies differ in participant, signal, vocabulary, prompts, correction, language-model assistance, averaging and whether results were online. Verified implanted speech studies reached 62 and 78 words per minute in defined 2023 tasks, while one participant averaged 56 words per minute during extensive home use reported in 2026.111216 Error rate and independence matter as much as peak speed.
Are high-performance implanted communication and motor BCIs approved treatments?
Most remain investigational. As of 4 September 2026, implanted systems for open-ended speech and general computer control were not generally marketed treatments. China had authorised one epidural BCI for a narrow hand-grasp compensation indication after cervical spinal cord injury.52 Trial authorisation, an FDA Breakthrough Device designation and FDA clearance of a cortical electrode for temporary use are not approval of a permanent complete system.53 Status must be checked for the exact device, intended use and jurisdiction.
Can a BCI cure paralysis?
No current evidence supports that general claim. A BCI may bypass an interrupted pathway while powered—for example by controlling a cursor, robot, muscle stimulation or spinal stimulation. Repeated paired training may also contribute to persistent functional gains in selected participants, but assistive performance and biological recovery must be measured separately.
Does a higher electrode count guarantee a better BCI?
No. Useful performance depends on placement, signal quality, stability, spatial coverage, decoder design, task, calibration and the user’s needs. Advertised channels may exceed the number recorded or usable. Clinical value is better judged by safe, reliable function, error recovery, independence and durability than by a hardware count alone.
Can a BCI work when a person has no reliable voluntary movement?
Possibly, but the evidence is especially sparse and assessment is difficult. A 2022 single-case study reported spelling through intracortical signals and auditory neurofeedback in a person described as completely locked-in.47 One case cannot establish how often the approach works, and diagnosis, sensory access, arousal, training time and false-positive control all matter. Established access methods should be assessed wherever any dependable movement remains.
Why are neural data treated as especially sensitive?
Recordings can support inferences beyond the immediate control command, may be difficult to replace if compromised and can be linked with health, behaviour and identity data. The raw signal is not a transparent transcript of the mind, but that does not make it harmless. Collection should be purpose-limited, secured, auditable and governed by explicit rules for retention, model training, sharing and deletion.
What should someone ask before considering a BCI trial?
Ask what function is realistically expected; what alternatives exist; which procedures, medications and travel are required; what adverse events and unknowns are plausible; who owns and can reuse data; what daily support is needed; how withdrawal works; and who pays for maintenance, follow-up or explant after the study. An independent clinician, accessible consent materials and a trusted support person can help separate research goals from personal benefit.
Primary research, registries, standards and official guidance
Sources and verification record
These references support the definitions, participant-level results, trial-status checks and governance distinctions used above. A registry establishes that a study is listed; it does not establish that its target enrolment was reached or that the device was effective.
Link behaviour: External source links open in a new tab.
- Slutzky, M. W., Vansteensel, M. J., Herff, C. & Gaunt, R. A. (2025), “A brain–computer interface working definition,” Nature Biomedical Engineering, 9, 792. DOI record.
- Chen, Y. et al. (2024), “Considerations and discussions on the clear definition and definite scope of brain-computer interfaces,” Frontiers in Neuroscience, 18, 1449208. DOI record.
- U.S. Food and Drug Administration (2021), “Implanted Brain-Computer Interface (BCI) Devices for Patients with Paralysis or Amputation—Non-clinical Testing and Clinical Considerations.” FDA guidance.
- Edelman, B. J. et al. (2025), “Non-Invasive Brain-Computer Interfaces: State of the Art and Trends,” IEEE Reviews in Biomedical Engineering, 18, 26–49. DOI record.
- Kaifosh, P., Reardon, T. R. & CTRL-labs at Reality Labs (2025), “A generic non-invasive neuromotor interface for human–computer interaction,” Nature, 645, 702–711. DOI record.
- Thompson, D. E. et al. (2014), “Performance measurement for brain–computer or brain–machine interfaces: a tutorial,” Journal of Neural Engineering, 11, 035001. DOI record.
- Speier, W., Arnold, C. & Pouratian, N. (2013), “Evaluating True BCI Communication Rate through Mutual Information and Language Models,” PLOS ONE, 8, e78432. DOI record.
- Pandarinath, C. et al. (2017), “High performance communication by people with paralysis using an intracortical brain-computer interface,” eLife, 6, e18554. DOI record.
- Willett, F. R. et al. (2021), “High-performance brain-to-text communication via handwriting,” Nature, 593, 249–254. DOI record.
- Jude, J. J. et al. (2026), “Restoring rapid natural bimanual typing with a neuroprosthesis after paralysis,” Nature Neuroscience. DOI record.
- Willett, F. R. et al. (2023), “A high-performance speech neuroprosthesis,” Nature, 620, 1031–1036. DOI record.
- Metzger, S. L. et al. (2023), “A high-performance neuroprosthesis for speech decoding and avatar control,” Nature, 620, 1037–1046. DOI record.
- Card, N. S. et al. (2024), “An Accurate and Rapidly Calibrating Speech Neuroprosthesis,” The New England Journal of Medicine, 391, 609–618. DOI record.
- Littlejohn, K. T. et al. (2025), “A streaming brain-to-voice neuroprosthesis to restore naturalistic communication,” Nature Neuroscience, 28, 902–912. DOI record.
- Wairagkar, M. et al. (2025), “An instantaneous voice-synthesis neuroprosthesis,” Nature, 644, 145–152. DOI record.
- Card, N. S. et al. (2026), “Long-term independent use of an intracortical brain–computer interface for speech and cursor control,” Nature Medicine, 32, 2504–2510. DOI record.
- Vansteensel, M. J. et al. (2016), “Fully Implanted Brain–Computer Interface in a Locked-In Patient with ALS,” The New England Journal of Medicine, 375, 2060–2066. DOI record.
- Vansteensel, M. J. et al. (2024), “Longevity of a Brain–Computer Interface for Amyotrophic Lateral Sclerosis,” The New England Journal of Medicine, 391, 619–626. DOI record.
- Oxley, T. J. et al. (2021), “Motor neuroprosthesis implanted with neurointerventional surgery improves capacity for activities of daily living tasks in severe paralysis: first in-human experience,” Journal of NeuroInterventional Surgery, 13, 102–108. DOI record.
- Mitchell, P. et al. (2023), “Assessment of Safety of a Fully Implanted Endovascular Brain-Computer Interface for Severe Paralysis in 4 Patients,” JAMA Neurology, 80, 270–278. DOI record.
- Hochberg, L. R. et al. (2012), “Reach and grasp by people with tetraplegia using a neurally controlled robotic arm,” Nature, 485, 372–375. DOI record.
- Collinger, J. L. et al. (2013), “High-performance neuroprosthetic control by an individual with tetraplegia,” The Lancet, 381, 557–564. DOI record.
- Flesher, S. N. et al. (2021), “A brain-computer interface that evokes tactile sensations improves robotic arm control,” Science, 372, 831–836. DOI record.
- Bouton, C. E. et al. (2016), “Restoring cortical control of functional movement in a human with quadriplegia,” Nature, 533, 247–250. DOI record.
- Ajiboye, A. B. et al. (2017), “Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia: a proof-of-concept demonstration,” The Lancet, 389, 1821–1830. DOI record.
- Lorach, H. et al. (2023), “Walking naturally after spinal cord injury using a brain–spine interface,” Nature, 618, 126–133. DOI record.
- Chandrasekaran, S. et al. (2026), “A neuroprosthesis for restoring hand movement and sensation in a person with complete tetraplegia,” Nature Medicine, 32, 2591–2601. DOI record.
- Fernández, E. et al. (2021), “Visual percepts evoked with an intracortical 96-channel microelectrode array inserted in human occipital cortex,” Journal of Clinical Investigation, 131, e151331. DOI record.
- Rubin, D. B. et al. (2023), “Interim Safety Profile From the Feasibility Study of the BrainGate Neural Interface System,” Neurology, 100, e1177–e1192. DOI record.
- Bullard, A. J. et al. (2020), “Estimating Risk for Future Intracranial, Fully Implanted, Modular Neuroprosthetic Systems: A Systematic Review of Hardware Complications in Clinical Deep Brain Stimulation and Experimental Human Intracortical Arrays,” Neuromodulation, 23, 411–426. DOI record.
- ClinicalTrials.gov (2009–present), “BrainGate2: Feasibility Study of an Intracortical Neural Interface System for Persons With Tetraplegia” (NCT00912041). Trial record.
- ClinicalTrials.gov (first posted 2021; study started 2022), “COMMAND Early Feasibility Study: Implantable BCI to Control a Digital Device for People With Paralysis” (NCT05035823). Trial record.
- ClinicalTrials.gov (2024–present), “Precise Robotically Implanted Brain-Computer Interface” (PRIME; NCT06429735). Trial record.
- U.S. Food and Drug Administration, “Breakthrough Devices Program” (accessed 4 September 2026). FDA program page.
- U.S. Food and Drug Administration (2026), “Cybersecurity in Medical Devices: Quality Management System Considerations and Content of Premarket Submissions.” FDA guidance.
- UNESCO (2025), “Recommendation on the Ethics of Neurotechnology.” Official text.
- Organisation for Economic Co-operation and Development (2019), “Recommendation of the Council on Responsible Innovation in Neurotechnology” (OECD/LEGAL/0457). OECD legal instrument.
- United Nations Human Rights Council (2025), “Neurotechnology and human rights” (A/HRC/RES/58/6). Official record.
- Council of Europe, “Neurotechnologies: Human Rights and Biomedicine” (accessed 4 September 2026). Council of Europe project page.
- United Nations (2006), “Convention on the Rights of Persons with Disabilities.” Official text.
- Wolpaw, J. R. et al. (2018), “Independent home use of a brain-computer interface by people with amyotrophic lateral sclerosis,” Neurology, 91, e258–e267. DOI record.
- Huggins, J. E., Wren, P. A. & Gruis, K. L. (2011), “What would brain-computer interface users want? Opinions and priorities of potential users with amyotrophic lateral sclerosis,” Amyotrophic Lateral Sclerosis, 12, 318–324. DOI record.
- Branco, M. P. et al. (2021), “Brain-Computer Interfaces for Communication: Preferences of Individuals With Locked-in Syndrome,” Neurorehabilitation and Neural Repair, 35, 267–279. DOI record.
- World Medical Association (2024), “WMA Declaration of Helsinki—Ethical Principles for Medical Research Involving Human Participants.” Official declaration.
- U.S. Government Accountability Office (2024), “Brain-Computer Interfaces: Applications, Challenges, and Policy Options” (GAO-25-106952). Official report.
- Silva, A. B. et al. (2024), “A bilingual speech neuroprosthesis driven by cortical articulatory representations shared between languages,” Nature Biomedical Engineering, 8, 977–991. DOI record.
- Chaudhary, U. et al. (2022), “Spelling interface using intracortical signals in a completely locked-in patient enabled via auditory neurofeedback training,” Nature Communications, 13, 1236. DOI record.
- Greenspon, C. M. et al. (2025), “Evoking stable and precise tactile sensations via multi-electrode intracortical microstimulation of the somatosensory cortex,” Nature Biomedical Engineering, 9, 935–951. DOI record.
- Valle, G. et al. (2025), “Tactile edges and motion via patterned microstimulation of the human somatosensory cortex,” Science, 387, 315–322. DOI record.
- International Organization for Standardization & International Electrotechnical Commission (2025), “ISO/IEC 8663:2025—Information technology—Brain-computer interfaces—Vocabulary.” Standard record.
- IEEE Standards Association, “P2731—Standard for a Unified Terminology for Brain-Computer Interfaces” (active project; accessed 4 September 2026). Project page.
- National Medical Products Administration of China (2026), “First invasive brain–computer-interface medical device approved for market” [translation of Chinese page title]. Official announcement.
- U.S. Food and Drug Administration (2025), “Layer 7 Cortical Interface—Premarket Notification K242618.” FDA 510(k) record.
Evidence reviewed through 4 September 2026. Sources do not carry equal evidential weight. Peer-reviewed studies establish specific results under their reported conditions; single cases establish feasibility rather than frequency; registries describe planned or current studies rather than outcomes; regulatory records apply to a defined product, indication and jurisdiction; and recommendations or standards have the force described by their issuing bodies. Because trials and authorisations change, verify current status before making a clinical, legal or investment decision.
Continue the Alternative Realities series
Technology, simulation and the futures they make possible
- 1 · Technological Innovations and the Future of Reality
- 2 · Virtual Reality: Technology and Applications
- 3 · Augmented and Mixed Reality
- 4 · The Metaverse
- 5 · Artificial Intelligence and Simulated Worlds
- 6 · Brain–Computer Interfaces and Neural Immersion
- 7 · Video Games as Immersive Alternative Realities
- 8 · Holography and 3D Projection
- 9 · Transhumanism and Post-Human Realities
- 10 · Ethics of Virtual and Simulated Realities
- 11 · Future Prospects Beyond Current Technologies
Written as an evidence-aware educational guide. Last evidence review: 4 September 2026.