Brain-Computer Interfaces
Linas JuozenasShare
Brain–Computer Interfaces
A brain–computer interface is a signal path—not a shortcut to a mind. It can translate selected patterns of neural activity into text, cursor movement, speech or physical action. For selected study participants, experimental systems are already adding or re-establishing routes through which people can express intention, humour, knowledge and skill. These achievements are extraordinary. So are the responsibilities.
agency
or task
activity
trained pattern
or action
the result
system adapt
A decoder between neural activity and an outcome
A BCI measures activity associated with a chosen task, extracts features and converts them into an output. Interactive BCIs close a loop by returning feedback to the user; some bidirectional systems additionally stimulate neural tissue to return artificial sensory information. Biomarker-triggered therapeutic stimulation is also called “closed loop,” but it is a distinct use from communication or device control.
The term covers very different systems. An electroencephalography (EEG) cap that selects letters, a sensor placed on the brain’s surface, an array of microelectrodes within cortex and an electrode deployed in a brain blood vessel can all support a BCI. Their resolution, risks, portability and possible uses are not interchangeable. The US Food and Drug Administration treats implanted BCIs for people with paralysis or amputation as medical devices that interface with the nervous system to restore motor or sensory capabilities; its 2021 guidance addresses non-clinical testing and the design of early-feasibility and pivotal studies.1
The phrase direct brain control can also mislead. The person usually performs a defined action—attempting speech, imagining or attempting movement, attending to a flashing target, or changing a learned brain rhythm. Sensors capture a noisy biological correlate. Software then estimates which trained class or trajectory best fits the data. The interface is powerful precisely because useful intent can remain represented in neural activity even when injury or disease interrupts the usual route to muscles.
What the system decodes
It decodes a statistical relationship learned under particular conditions. A movement decoder does not suddenly know a person’s private memories. A speech decoder trained on attempted speech is not an all-purpose window into every passing thought. Capability depends on where and how signals are measured, what task was performed, what training data exist and how the output model constrains its guesses.
- TaskThe user attempts speech, movement or another agreed action.
- RecordElectrodes or optical sensors capture a relevant signal.
- CleanNoise, drift and artifacts are reduced; useful features are extracted.
- DecodeA model estimates a letter, word, direction, click or continuous movement.
- ReturnThe result appears; feedback helps the user and model adapt.
A read-out BCI measures neural signals and drives an external result such as text, a cursor or a robotic limb.
A read–write system also stimulates neural tissue to return information, such as an artificial touch sensation.
The person supplies the intention, knowledge and meaning. The interface supplies a route.
Core distinction
Every route trades signal access for burden
There is no single ladder from “worst” to “best.” The relevant question is whether an interface offers enough reliable information for a person’s goal at an acceptable medical, practical and financial cost.
| Approach | Where it measures | Main strengths | Main constraints | Current position |
|---|---|---|---|---|
| Intracortical arrays | Microelectrodes penetrate cortex and record local field potentials plus, on some channels, action potentials from nearby neurons. | Provides high spatial and temporal resolution and has supported high-bandwidth, multidimensional decoding in individual human studies. | Neurosurgery, infection or bleeding risk, tissue response, connector or hardware burden, signal change over time and demanding support. | Clinical proof-of-conceptPowerful results, still investigational for these uses. |
| Electrocorticography (ECoG) | Electrodes are placed over cortex either above the dura (epidural ECoG) or beneath it (subdural ECoG); their contacts do not penetrate cortex. | Broad cortical coverage and useful high-frequency activity without penetrating the tissue at every contact. | Requires cranial surgery and has lower spatial resolution than penetrating arrays. Five- and seven-year fully implanted home-use case reports exist, but multi-participant durability and scalable support evidence remain sparse. | Clinical proof-of-conceptNotable speech and motor demonstrations. |
| Endovascular electrodes | An electrode-bearing stent is delivered through a vein and positioned beside relevant cortex. | Avoids open-skull placement and uses familiar catheter techniques. | Requires general anaesthesia, jugular venous access, an intracranial vascular implant connected to subcutaneous chest hardware, antiplatelet therapy and imaging follow-up. Venous anatomy constrains placement, and present devices record lower-spatial-resolution field potentials rather than single-unit activity. | Early feasibilityIn SWITCH, five men enrolled; four had suitable venous anatomy, were implanted and completed 12 months. No serious adverse event, target-vessel occlusion or migration was observed, but eight mild adverse device effects affected all four. All initially used eye tracking for cursor movement and the BCI for selection; BCI-only automatic scanning was tested in one participant.2 |
| EEG | Electrodes on the scalp record summed electrical activity through skin, skull and other tissue. | Non-surgical, comparatively accessible, portable and repeatable. | Lower spatial resolution and signal-to-noise ratio; eye, facial-muscle and movement artifacts; setup and calibration; variable real-world performance. | Established research platformUseful in selected communication, control and rehabilitation paradigms; limitations remain substantial.3 |
| fNIRS | Light at the scalp estimates blood-oxygen changes near the cortical surface. | Non-surgical, silent and electrically tolerant; can complement EEG. | The blood-flow response is indirect and slow, limiting rapid control; hair, motion and systemic physiology can interfere. | Active researchPromising mainly for slower state or task classification and hybrid systems. |
| MEG, fMRI and functional ultrasound | MEG measures brain-generated magnetic fields; fMRI and functional ultrasound infer activity indirectly from changes in blood flow. Conventional MEG and fMRI use large external scanners. Functional ultrasound may be more portable, but adult human task mapping and decoding have generally required a skull defect or surgically installed acoustic window.3 | Valuable scientific mapping and, in some settings, rich signals. | MEG and fMRI are costly and immobile. For functional ultrasound, the adult skull attenuates and distorts the signal; no clinical assistive fUS BCI is established. Transcranial focused ultrasound is a stimulation method, not a neural recording method. | Research toolDo not confuse a laboratory decoder with a wearable product. |
Adjacent is not identical
Eye tracking, electromyography (EMG), peripheral-nerve cuffs and muscle sensors can be excellent assistive interfaces, but they do not necessarily measure the brain. Hybrid systems may combine them with a BCI. Precise labels matter because the benefits and privacy implications differ.
“Minimally invasive” is relative
A catheter route or surface electrode may avoid penetrating cortex, yet it can still involve anaesthesia, an implanted lead, medication, imaging and long-term clinical follow-up. It should never be heard as “risk-free.”
More channels are not the endpoint
Channel count matters only when it produces stable, usable information. Placement, signal quality, decoder design, task fit, comfort, reliability and support can matter more than a large headline number.
What people and research teams have actually demonstrated
The field has crossed important thresholds. It has not crossed all of them. These examples deserve celebration for what they accomplished—and respect for the participants whose sustained effort made the work possible.
Fast brain-to-text communication
In 2021, one participant with paralysis used intracortical recordings of attempted handwriting to produce text online at 90 characters per minute with 94.1% raw accuracy; accuracy exceeded 99% only in a separate offline analysis using general-purpose autocorrection.4 The result was a major improvement in that research setting. It did not show that everyone will reach the same speed, that the implant can be placed without risk, or that performance will remain unchanged for life.
Why it matters: attempted handwriting produced distinguishable neural patterns for individual characters.Large-vocabulary attempted-speech decoding
A 2023 intracortical study with one participant who had ALS decoded attempted speech at 62 words per minute. It reported a 9.1% word-error rate for a 50-word vocabulary and 23.8% for a 125,000-word vocabulary.5 In a separate one-participant ECoG study, attempted-speech text decoding reached a median 78 words per minute with a 25% median word-error rate and also drove personalised synthetic speech and an animated facial avatar in laboratory tasks.6
Why it matters: the target is conversation, not merely a laboratory classification score.Rapid calibration and personalised voice
In 2024, a speech neuroprosthesis for one man with ALS achieved 90.2% word accuracy with a 125,000-word vocabulary on day two, after 1.4 additional hours of training data. With further longitudinal training, structured prompted tests reached a 2.5% word-error rate by session 15 and approximately 97.5% accuracy was maintained through 8.4 months. A separate personalised text-to-speech component rendered decoded text using pre-ALS voice recordings.7 This is an impressive individual result, not a population estimate.
Why it matters: calibration burden and personal identity are part of usefulness.Streaming voice during attempted speech
A 2025 one-participant ECoG study synthesised speech from 80-millisecond neural chunks, but 80 milliseconds was the update interval—not end-to-end latency; median onset latency was about 1.1 seconds. Its 1,024-word task reached 47.5 words per minute with a 58.8% word-error rate, whereas a restricted 50-phrase AAC set reached 90.9 words per minute with a 12.3% word-error rate. A separate one-participant intracortical study generated 10-millisecond audio frames with approximately 25-millisecond algorithmic feedback latency, but open listener transcription still produced a median 43.75% word-error rate.8 These were real-time engineering advances, not yet reliable open-vocabulary conversation.
Why it matters: communication quality includes social timing, not just correct words.A rare but important step toward everyday use
A 2026 report followed one man with ALS using an intracortical speech-and-cursor BCI at home for nearly two years without researchers present. Across more than 3,800 hours, he used it for communication, computer control and full-time work. The paper reports 183,060 sentences at an average 56 words per minute; he marked 92% of sentences as at least mostly correct, while structured tests exceeded 99% word accuracy. A care partner donned and removed the hardware and initiated the software, so the report demonstrates participant-directed use with setup assistance rather than fully independent setup.9
This is unusually valuable evidence because it concerns long-duration, participant-directed use outside the laboratory. It remains a report about one participant. Both truths belong in the same sentence.
Robotic movement with artificial touch
In a 2021 one-participant study, motor-cortex signals controlled a robotic arm while arm sensors triggered somatosensory-cortex microstimulation that evoked artificial tactile percepts. With stimulation enabled, median object-transfer time fell from 20.9 to 10.2 seconds versus vision-only control.10 The deeper lesson is that feedback can reduce uncertainty and improve measured performance.
A brain–spine digital bridge
In 2023, bilateral epidural ECoG was linked to implanted lumbosacral spinal stimulation in one man with chronic spinal-cord injury. After prior spinal-stimulation rehabilitation, he stood and walked using aids such as a walker or crutches, including on stairs and complex terrain.11 The uncontrolled one-participant study did not demonstrate unaided walking or isolate the contribution of the brain interface.
Movement, sensation and a person’s own hand
A 2026 first-in-human case report combined motor-cortex decoding, somatosensory-cortex microstimulation, forearm neuromuscular stimulation, an active hand orthosis, reinforcement-learning force control, transcutaneous cervical spinal stimulation and extensive training in one man with complete tetraplegia. He performed self-feeding and a delicate-eggshell task while assistive components were active. Off-system gains were principally elbow-flexion strength and wrist tactile detection; independent finger movement and finger sensation were not restored.12 Because this was an uncontrolled, multi-component one-participant programme, the contribution of each element cannot be separated.
Do not erase the participant from the achievement
These systems do not simply arrive, switch on and perform. Participants undergo procedures, repeat calibration tasks, practise control, tolerate failures, explain what feels useful and help teams redesign the technology. Their time, skill, risk and lived expertise are part of the scientific contribution. Public storytelling should name that contribution, avoid depicting people as passive demonstrations, and never treat disability as a dramatic prop.
Do not count papers as people
Several landmark publications are later stages of work with the same participant. The 2023 ECoG speech-and-avatar study and 2025 streaming-voice study followed one woman; the 2024 rapid-calibration, 2025 instantaneous-voice and 2026 long-term home-use reports followed one man. Continued work with the same person provides valuable depth and durability evidence, but it does not enlarge the number of independent participants.
How to read a “breakthrough” headline
A dramatic video can show that something happened. It cannot, by itself, show how often it worked, what setup or assistance was involved, how long it lasted, whether harms occurred or how another person would fare.
| Evidence level | What it can support | What it cannot yet support |
|---|---|---|
| EstablishedMeasurement principle or adjacent clinical technology | EEG records scalp electrical activity; implanted stimulation and sensory prostheses show that neural interfaces can be clinically useful in defined indications. | That every new BCI use is safe, effective or reimbursed. |
| Clinical proof-of-conceptHuman participant uses the system meaningfully | Feasibility for a specific task, person, implant, decoder and support environment. | Typical benefit, rare harms, durability for decades or suitability for the wider population. |
| Active researchEarly human, laboratory, animal or engineering evidence | A mechanism worth testing and technical possibilities that may mature. | Clinical benefit or a delivery timeline. |
| SpeculativeForecast, concept or company ambition | A question that can guide research or public debate. | That the capability exists. A roadmap is not a result. |
Metrics that belong together
- Accuracy and error type: character accuracy, command accuracy and word-error rate are not interchangeable. WER counts substitutions, deletions and insertions per reference word and can exceed 100%.
- Speed and latency: how much useful communication or control occurs, and how quickly feedback arrives.
- Calibration: minutes or hours, frequency of retraining, and whether researchers must intervene.
- Robustness: performance across days, posture, fatigue, environments and disease progression.
- Workload: concentration, discomfort, visual demand and the cost of correcting errors.
- Life impact: tasks chosen by the user, participation, privacy and time returned to the person.
Labels that do not prove approval
An FDA Breakthrough Device designation can increase communication with the agency and support efficient review. It is not marketing authorisation and not a finding that the device is safe or effective. Likewise, an early-feasibility study asks whether a concept can be studied safely and whether it works well enough to justify further research.
Company videos can add a valuable participant perspective, but performance claims should be checked against registered protocols, peer-reviewed results and complete adverse-event reporting.
Numbers without context create false certainty
Electrode count, hardware sampling or telemetry rate, and user information-transfer rate are different quantities. Report task throughput—such as words, characters or targets per minute—together with errors, vocabulary, language-model assistance, calibration and independent-use time. The often-repeated figure of approximately 50 bits per second is a natural-speech comparator, not demonstrated clinical BCI throughput. No single metric tells readers whether a person can set up the system, depend on it at home, correct it, afford its support, use it without exhaustion or keep it functioning as their body and software change.
Safety is more than surviving implantation
A BCI can become part of how someone speaks, works, moves and relates to other people. Safety therefore includes surgery and biology, but also reliability, maintenance, cybersecurity, continued support and an honourable exit.
Medical risk
Depending on the approach: anaesthesia, infection, bleeding, stroke, seizure, vessel injury or clotting, inflammation, scar response, medication effects, pain and explantation risk. The exact profile belongs to the exact device and procedure.
Biological change
Signals can drift as electrodes, tissue, medication, fatigue or disease change. A seven-year implanted-BCI report in ALS described years of useful home communication followed by unreliable control. No technical malfunction was identified; declining neural-signal amplitude and CT-observed atrophy suggested ALS-related neurodegeneration, although the authors stated that alternative explanations remained plausible.13
System dependence
Batteries, cables, wireless links, operating systems, cloud services, calibration tools and trained personnel can all fail or become obsolete. A clinically useful signal is not enough unless the whole chain works.
Exit and aftercare
Before implantation, participants need a funded plan for trial completion, company failure, unsupported software, repair, continued use, safe deactivation and—when medically appropriate—retention or removal.
Device abandonment is a patient-safety issue
A 2024 systematic review and expert consensus defined neurological-device abandonment around failures of consent and reasonable medical, technical or financial support before the end of an implant’s intended life, including failures that create safety concerns or make the device ineffective.14 This should change research design: long-term stewardship cannot be an optional promise made after funding disappears.
The largest published chronic intracortical series is informative—not conclusive
A multicentre report covering 14 participants and 12,203 participant-days recorded 68 device-related adverse events, including six serious events. It reported no intracranial infection, unanticipated adverse device effect, device-related death, permanent disability or adverse-event-driven explantation.15 This was open-label, nonrandomised Class IV evidence from research systems, not proof of commercial-scale safety.
The clinical consent conversation
- State the evidence level plainly. “Investigational” should be understood in practical terms, including uncertainty about personal benefit.
- Describe alternatives. Eye gaze, switch access, AAC, environmental control and other supports may be safer or better for a person’s priorities.
- Separate research from care. Participation can help science without guaranteeing therapy.
- Revisit consent. Preferences can change after implantation, disease progression, new risks, software updates or a change in the research team.
- Presume decision-making capacity. Speech, language or motor disability is not evidence of incapacity. Provide the person’s preferred AAC, interpreters, accessible materials, additional time, teach-back and supported decision-making before considering proxy consent.
- Explain withdrawal precisely. Stopping research procedures, refusing future data reuse, continuing clinical or technical support, deactivating the system and surgical removal are separate decisions. Removal has its own risks and may not be immediate.
- Preserve communication choice. A person’s existing speech, eye-gaze, AAC, switch access or partner-assisted method must remain available unless the person asks to stop using it; reliability alone does not justify displacement.
Security is part of clinical quality
- Minimise collection. Record only the signals and metadata needed for the agreed function.
- Process locally where feasible. Avoid sending raw neural streams to remote systems without a clear, necessary reason.
- Encrypt data and updates. Authenticate devices, log access and maintain a vulnerability-response process.
- Fail safely. Provide a simple user-controlled stop, prevent unsafe actuation and make degraded performance visible.
- Plan updates conservatively. A decoder change can alter behaviour; validation, rollback and user approval matter.
Mental privacy is not a science-fiction issue
Today’s systems do not reveal an entire mind. They can nevertheless produce intimate, reusable data about neural activity, tasks, performance, health and behaviour. Protection should begin before decoding becomes more powerful.
What neural data can—and cannot—say
A neural recording is not a pure transcript of thought. Meaning comes from context: sensor location, task instructions, calibration labels, timing, other physiological signals and the model used to interpret them. In a trained setting, a decoder may distinguish attempted movements, infer intended speech units or estimate attention to a stimulus. That does not justify claims that a headset can discover a person’s hidden beliefs, political identity or every private feeling.
Yet modest inferences can still matter. A dataset may reveal a diagnosis, fatigue, response to stimuli, error patterns or behavioural preferences. Repeated recordings can become more informative when combined with voice, gaze, location, health records or platform activity. Privacy risk is therefore not only “Can someone read a sentence from my brain today?” It is also “What could this dataset reveal after it is retained, combined, sold or re-analysed tomorrow?”
Privacy should follow the person, not the product label
A hospital implant, a research headset and a consumer “wellness” device may fall under different rules even when all collect sensitive signals. Ethical protection should not vanish because a company calls a device entertainment, productivity or meditation technology.
A meaningful control set
- Purpose choice: the user knows exactly which function is being trained and can reject secondary uses.
- Collection choice: raw signals, inferred features, errors and contextual data are distinguished rather than bundled together.
- Access choice: the user can see who accessed data, for what reason and under which authority.
- Output choice: the user can preview, correct, cancel or privately discard decoded communication.
- Model choice: consequential updates require explanation, validation and a safe way back.
- Lifecycle choice: retention deadlines, deletion limits, export and portability are explained separately for raw signals, derived features, decoded output, telemetry and user-specific models.
- Future-use choice: operating the device is not bundled with consent for secondary research, commercial reuse, advertising or AI-model training.
- Exit choice: pausing, disconnecting or leaving a study does not remove access to ordinary care or existing assistive communication.
The following are ethical safeguards this article argues should be protected. They are not presented as a claim that every jurisdiction already recognises each one as a freestanding legal right.
Mental quiet
People need times when no employer, school, platform, clinician or device scores their attention or emotion. Protected time to disconnect supports recovery, inward thought and the freedom to change one’s mind without generating a permanent record.
Protection from covert steering
Stimulation, adaptive prompts and persuasive interfaces must not be used to coerce or manipulate. Disclosure alone is insufficient: the user needs understandable controls, meaningful consent, the ability to refuse or stop without penalty, and protection against hidden or exploitative optimisation.
An ordinary alternative
No one should have to surrender neural data to receive education, work, insurance or a basic service. Refusing a BCI should not be treated as refusing to participate in society.
The 2025 UNESCO Recommendation on the Ethics of Neurotechnology places human dignity, autonomy, mental privacy, freedom of thought, informed consent, fairness and protection against coercive uses at the centre of governance.16 It is a global normative standard, not a binding world law. Its importance is that it sets expectations before workplace, educational and consumer uses become ordinary.
The interface carries a person’s ideas; it does not become their author
When familiar routes for speech or movement become less accessible, a BCI can add or re-establish a route for communication and action. Its value should be defined first by the person: agency, relationships, self-expression, learning, creation and—if they choose—work. It is not conditional on making them productive for someone else.
Speech or movement limitations do not establish the absence of thought, preference or personhood. A fluent decoder can add a route; it does not create the person or the meaning.
Human-first communication and authorship
Protect intelligence without turning it into a hierarchy of human worth
Cognitive abilities vary across people and across a lifetime. When a well-designed assessment validly measures reasoning, memory, learning or another ability, real differences in performance matter; so do creativity, expertise and forms of understanding that one score does not capture. These capacities grow from biology and development together with education, sustained curiosity, deliberate practice, health and years of accumulated knowledge. They are not toys that can simply be bought and inserted.
People who devote a lifetime to understanding a field may carry models, distinctions and hard-won judgement that no tool, database or instruction manual can reproduce on demand. Their continued presence can guide others, correct errors and open paths no group had yet imagined. Protecting their sleep, health, concentration, independence and ability to communicate is a genuine social priority. So is recognising their achievement and giving them credit.
Human dignity, health care, communication and support never depend on IQ, productivity or exceptional achievement. That universal principle does not require society to pretend that intelligence, learning or expertise are unimportant. We can protect equal personhood while also cultivating cognitive growth and celebrating people whose disciplined minds contribute something rare.
A BCI can serve both principles magnificently: not by manufacturing wisdom or human worth, but by preventing a motor or speech barrier from concealing a person’s intentions, knowledge and style. Fidelity means preserving timing, vocabulary, humour, uncertainty and the right to revise.
An authorship contract for neural communication
- The person sets the goal. The interface serves an intention chosen by the user, not a metric selected for publicity.
- The person controls release. By default, decoded drafts remain private until the user confirms release through an accessible control. Any safety exception must be narrowly defined, disclosed and agreed in advance.
- Correction must be easy. A decoder error is not the person’s belief, consent or promise.
- Automation must be visible. Prediction, paraphrase, translation and generative completion should be distinguishable from directly decoded selections.
- Credit follows documented contribution. Using a decoder does not by itself transfer credit away from the user. Direct decoding, predictive correction, human editing and generative expansion should be distinguishable; ethical credit and legal authorship are related but not identical.
- Records preserve uncertainty. Systems should not turn an ambiguous signal into a falsely confident statement.
The system estimates a letter, phoneme, word, click or movement from a trained neural pattern. Accuracy and uncertainty can be measured.
A language model ranks likely continuations or corrects errors. It can increase speed, but may also substitute common phrasing for unusual intent.
An AI system turns a short selection into a longer sentence, voice or image. That expansion should be identified and approved, not silently presented as direct thought.
The person intentionally selects, edits or confirms a message using their preferred communication supports. Decoder fluency alone is not proof of consent, capacity, legal responsibility or authorship.
Protect originality from well-meaning interference
Original ideas often need periods of undirected, private thought. Constant prompts, autocomplete and social attention can pull a person toward what is already expected. A responsible interface therefore includes room to wander, compose silently and keep unfinished ideas private. Then, when the thinker chooses to return, other people become essential: they can test, support, develop, communicate and celebrate the idea. Solitude and community are partners when both remain voluntary.
Enhancement should be described honestly
BCIs may eventually expand how quickly people interact with tools or provide new sensory channels. That would be a meaningful capability. It is not the same as demonstrating growth in general intelligence, wisdom or expertise. If an enhancement claim is made, researchers should specify which ability improved, for whom, on what task, for how long and at what cost. Respect for intelligence requires better evidence, not smaller ambitions.
A brilliant implant is not useful without an accessible support system
Access includes surgery, but it also includes fitting, training, transport, home setup, repair, software compatibility, language, supporter time, electricity, broadband, clinical follow-up and a funded plan for the anticipated service life and post-trial transition.
Begin with the person’s priority
Researchers may optimise a cursor while a prospective user most values private conversation, pain reduction, environmental control or a way to call family. Studies of potential BCI users and people with locked-in syndrome show that priorities and preferred control strategies cannot be safely assumed by designers.17
Design around real days
Electrode and headgear design—including compatibility with coily hair and protective hairstyles—perspiration, spasticity, fatigue, medication, posture, breathing support and changing speech can determine whether a system works. Studies should measure setup time, user-directed use with the person’s preferred level of assistance and recovery from failure, and should not make an unpaid caregiver an unexamined eligibility requirement.
Keep existing routes available
BCI should complement—not prematurely displace—eye tracking, switches, AAC, personal assistance and the communication methods a person already trusts. Redundancy is dignity when technology fails.
| Access dimension | A narrow approach asks | A responsible approach asks |
|---|---|---|
| Affordability | What does the hardware cost? | Who pays for evaluation, procedure, rehabilitation, consumables, support, upgrades, repair and clinically appropriate continued use, deactivation, retention or removal? |
| Usability | Can the participant complete the test? | Can the person choose, set up, pause, troubleshoot and recover from errors at home with acceptable fatigue? |
| Language | Does the decoder work in English? | Does it preserve the user’s language, accent, names, technical vocabulary, code-switching and communication style? |
| Representation | Is the recruitment target met? | Do trial populations, investigators and advisers reflect relevant variation in disability and communication method, age, sex and gender, race and ethnicity, language, socioeconomic and geographic context, and support availability—and are exclusion, attrition and subgroup performance reported? |
| Continuity | Did the study endpoint succeed? | What happens after the grant, the company, the operating system or the participant’s condition changes? |
Disability-led does not mean technology-opposed
People can welcome a tool that restores communication or movement while rejecting the claim that their life was empty before it. They can value independence and interdependence at the same time. Co-design means paying people for expertise, including them before requirements are fixed, and letting different users choose different definitions of benefit.
There is no single “BCI law”
A medical implant, research prototype, consumer headset and workplace-monitoring system can engage different combinations of medical-device, research, data-protection, consumer, employment, disability and cybersecurity law.
| Framework | What it contributes | Important limit |
|---|---|---|
| US FDA medical-device oversight | The FDA’s implanted-BCI guidance addresses bench, animal, biocompatibility, electrical, software, cybersecurity and clinical-study considerations for devices intended for paralysis or amputation.1 | FDA has no single BCI classification. Intended use, risk, predicates and technological characteristics can determine whether 510(k), De Novo or premarket approval applies. Significant-risk investigations generally require FDA investigational-device-exemption authorisation and institutional-review-board approval; neither an IDE nor Breakthrough Device designation permits commercial marketing. A genuinely nonmedical, low-risk wellness product may instead fall outside device regulation or within FDA’s general-wellness policy. |
| OECD neurotechnology recommendation | The 2019 recommendation calls for responsible innovation, safety assessment, inclusive deliberation, privacy, stewardship, trust and anticipation of misuse across the innovation cycle.18 | This is a nonbinding Council recommendation, principally focused on responsible innovation in neurotechnology for health. It creates no directly enforceable duties and is not a product licence. |
| UNESCO ethics recommendation | Adopted in 2025, it supplies a global human-rights-oriented framework for medical and non-medical neurotechnology, including children, work, marketing, consent and neural data.16 | This is a nonbinding UNESCO recommendation, not a treaty or directly enforceable world law. Implementation requires action by member states and institutions. |
| European Union | The Medical Device Regulation governs qualifying medical devices. The GDPR applies when neural information relates to an identified or identifiable person, although “neural data” is not itself a named special category: Article 9 depends on whether the information qualifies, for example, as health, genetic or uniquely identifying biometric data.19 Under AI Act Article 6(1), medical-device AI is high-risk only when it is a covered product or safety component and third-party conformity assessment is required. A July 2026 amendment moved that Annex I regulated-product layer to 2 August 2028—not the entire AI Act.20 | A BCI is not automatically “high-risk AI” merely because it involves the brain. Classification requires the actual device, AI function, intended purpose and legal criteria. |
| Chile | Law 21.383 amended Chile’s Constitution in 2021, requiring scientific and technological development to respect physical and psychological integrity and directing legislation to safeguard brain activity and information derived from it.21 | The amendment did not itself enumerate the popular “five neurorights,” and a separate implementing bill remained pending at this article’s evidence date.21 “Neurorights” is not one settled worldwide code. |
| US state privacy laws | Colorado HB24-1058, effective 7 August 2024, added biological data—including neural data—to the state’s sensitive-data framework, although the umbrella definition includes an identification-purpose qualifier.22 California SB-1223, effective 1 January 2025, added measured central- or peripheral-nervous-system activity to its sensitive-personal-information framework while excluding information inferred solely from non-neural data.23 | Both laws retain their ordinary coverage thresholds and healthcare or research exclusions. Neither protects every neural datum or creates a comprehensive neurorights code. |
Existing rights still matter
Privacy, bodily integrity, freedom of thought and expression, non-discrimination, disability rights, workplace protections and research ethics already govern parts of neurotechnology. A new label should not weaken them.
Gaps are still real
Consumer neural data, model-derived inferences, cross-border cloud processing, post-trial support and coercive monitoring may fall between familiar categories. Targeted rules can clarify duties before harm becomes routine.
Govern capability, not mythology
Rules should respond to what a system measures, infers, stores, changes and controls. Inflated “mind-reading” stories can distract from present risks such as opaque consent, excessive retention and unsupported implants.
High-stakes settings need a presumption against coercion
Employers, schools, insurers, law enforcement and militaries hold power over individuals. Neural monitoring or stimulation should not proceed where refusal or withdrawal can trigger retaliation or loss of employment, education, insurance, care, benefits or liberty. Where use is considered, necessity, proportionality, independent oversight, accessible ongoing consent, viable refusal and enforceable protection against retaliation are minimum requirements—not optional refinements.
A responsible BCI is designed for a whole life
The strongest standard is not “the demo worked.” It is that the person remains safer, freer, better heard and properly supported across success, failure, change and exit.
Questions for a prospective participant and any supporter they choose
- What is the study’s primary goal: safety, feasibility or benefit?
- How many people have used this exact device, for how long and with which adverse events?
- Which tasks can I realistically perform, and what existing abilities or assistance does the system require?
- Which setup, cleaning, calibration, charging, troubleshooting and on-call tasks are expected from me or a chosen supporter, how much time do they take, and what paid technical support is available?
- What happens to raw neural data, inferred data and recordings of errors?
- Who controls the implant, external hardware, software licence, raw signals, inferred data and user-specific models—and can I access and export each?
- Can I review or stop a message or action before it is sent?
- Who pays for maintenance, travel, repair, software, emergency care and removal?
- What is the written plan if I withdraw, the study ends or the company closes?
- Which organisation is contractually responsible for support, security updates and transition if the sponsor is acquired, insolvent or closes?
- Will leaving affect my ordinary care or other assistive technology?
- Who can I contact independently of the sponsor?
Obligations for researchers and companies
- Publish protocols, endpoints, failures, adverse events and conflicts of interest.
- Report distributions and individual variability, not only best-case sessions.
- Measure meaningful home outcomes chosen with prospective users.
- Build a user-controlled pause, confirmation and correction path.
- Distinguish decoded intent from algorithmic completion in logs and interfaces.
- Secure neural data across collection, transmission, storage, model training and deletion.
- Guarantee maintenance, vulnerability response, spare parts and clinical escalation for a defined period.
- Define hardware title, software rights, data access and export, support duration, end-of-life notice, transition assistance and obligations after acquisition or insolvency.
- Measure caregiver and supporter workload, obtain their separate consent for research tasks, compensate their work and design toward the user’s preferred level of independence or interdependence.
- Fund post-trial support and a user-centred, clinically appropriate plan for continued use, transition, deactivation, retention or removal before implantation begins.
- Compensate participants for expertise and include them in governance.
- Never use a participant’s life as proof of claims the study did not test.
Obligations for hospitals and review boards
- Assess the entire hardware–software–service chain, not only the implant.
- Provide accessible, independent consent support and the participant’s preferred communication method; do not infer decision-making incapacity from speech, language or motor disability.
- Review update, cybersecurity, emergency and explantation plans.
- Protect continued access to ordinary clinical care and communication.
- Monitor cumulative burden, fatigue, privacy and psychosocial effects.
- Reassess when the intended use, decoder or commercial partner changes.
Obligations for policymakers and payers
- Make coverage decisions from clinical value and user priorities, not hype.
- Cover training, support and maintenance when they are necessary to the benefit.
- Protect neural and inferred data across medical and consumer contexts.
- Prevent coerced use in employment, education, insurance and public services.
- Require credible continuity plans before long-lived devices enter trials or markets.
- Fund independent replication, accessible infrastructure and multilingual design.
A compact design test
Does the system increase the user’s effective agency? The person should be able to understand the system, decide when to use it, express and correct intended output, preserve privacy and receive long-term support. They must also be able to withdraw from research without abandonment or loss of ordinary care, with a funded plan for continued support, transition, deactivation, retention or removal.
Questions readers should ask
The most useful answer is often neither “impossible” nor “inevitable,” but a clear description of the signal, task, evidence and uncertainty.
Can a BCI read anyone’s thoughts?
No general-purpose system can open an arbitrary person’s mind and transcribe everything in it. A 2025 four-participant intracortical study decoded deliberately imagined sentences in structured, participant-specific tasks and detected some task-evoked private inner speech; large-vocabulary word-error rates remained approximately 26–54%, and a thought passphrase used to gate decoding achieved 98.75% detection.24 This establishes a real privacy and consent issue—not continuous recovery of arbitrary, unsolicited thoughts.
Can a consumer EEG headset reveal my exact inner speech?
Not with the reliability and open vocabulary implied by “reading inner speech.” Scalp EEG can detect broad patterns and task-linked responses, and experimental classifiers may distinguish limited trained categories. Muscle, eye and movement artifacts are also easy to misinterpret. A product should disclose independent validation, chance comparison, false positives and out-of-sample performance.
Will BCIs upload knowledge or make a person instantly more intelligent?
There is no demonstrated technology for uploading semantic knowledge, autobiographical memories, wisdom or general intelligence into a human brain. Small experimental stimulation studies have altered performance on specific recognition tasks, but they have not extracted, stored, replayed or transferred a person’s memory content.25 A BCI may improve access to information, communication speed or control of tools; training may also build a specific skill. Those are valuable outcomes. Claims of broader intelligence growth should be tested with valid measures and durable transfer beyond the trained task.
Are implanted BCIs already standard treatment?
Most implanted communication and control BCIs remain investigational. A narrow exception is China’s March 2026 marketing authorisation of an epidural implantable BCI hand-movement compensation system, used with a pneumatic glove, for selected adults with chronic cervical spinal-cord injury.26 No chronic assistive intracortical or ECoG BCI has US FDA marketing authorisation. Other neural interfaces—including cochlear implants and some stimulation devices—are established for defined indications, but their approval and evidence do not transfer to a new BCI. Check the exact device, indication, jurisdiction and regulatory status.
Does “non-invasive” mean harmless?
No surgery usually means a very different and often lower medical-risk profile, which is important. It does not erase risks from poor decisions, distraction, false output, skin irritation, data exploitation, misleading health claims or coercive monitoring. Risk follows the complete use, not only the sensor.
Does an AI language model improve or distort neural communication?
It can do either. Prediction can correct noise and make communication much faster. It can also favour common language, erase unusual names or ideas, and produce fluent words the user did not intend. Good design displays uncertainty, makes correction easy, preserves a literal option and identifies generated expansion.
If a BCI helps produce a book, design or discovery, who deserves credit?
Credit should track contribution. The person who originated, selected and approved the ideas is not displaced by the fact that an interface decoded their selections. Researchers and engineers deserve credit for the enabling system; a generative model’s material contribution should be disclosed. Records should make these roles visible rather than attributing all output to the machine or all engineering to the user.
Will only wealthy people benefit?
That outcome is possible but not predetermined. Public research, coverage policy, open standards, disability-led procurement and long-term service funding can widen access. Hardware price alone will not solve inequity; clinical infrastructure, language support, training and maintenance must also be distributed.
Should society ban enhancement uses?
A single answer is too crude. The risk differs between voluntary adult exploration, an implanted medical device, a child’s school, an employer’s productivity programme and military use. Begin with evidence, power, accessible and ongoing consent, proportionality, reversibility, privacy and a real ability to refuse without penalty. Uses chosen by disabled people for communication, movement, sensation or symptom management deserve equitable access and rigorous evidence; enhancement claims deserve precise testing; coercive use should not proceed.
Primary studies and governing sources
Research reports establish what was shown in their participants and conditions. Official sources establish current frameworks. Neither should be stretched beyond its scope.
- US Food and Drug Administration. Implanted Brain-Computer Interface Devices for Patients with Paralysis or Amputation: Non-clinical Testing and Clinical Considerations (final guidance, 2021). Official guidance
- Mitchell P, et al. Assessment of Safety of a Fully Implanted Endovascular Brain-Computer Interface for Severe Paralysis in 4 Patients. JAMA Neurology (2023). doi:10.1001/jamaneurol.2022.4847. Human case series
- Edelman BJ, et al. Non-Invasive Brain-Computer Interfaces: State of the Art and Trends (2025). Review
- Willett FR, et al. High-performance brain-to-text communication via handwriting. Nature (2021). doi:10.1038/s41586-021-03506-2. One-participant study
- Willett FR, et al. A high-performance speech neuroprosthesis. Nature (2023). doi:10.1038/s41586-023-06377-x. One-participant study
- Metzger SL, et al. A high-performance neuroprosthesis for speech decoding and avatar control. Nature (2023). doi:10.1038/s41586-023-06443-4. One-participant study
- Card NS, et al. An Accurate and Rapidly Calibrating Speech Neuroprosthesis. New England Journal of Medicine (2024). doi:10.1056/NEJMoa2314132. One-participant study
- Littlejohn KT, et al. A streaming brain-to-voice neuroprosthesis to restore naturalistic communication. Nature Neuroscience (2025), doi:10.1038/s41593-025-01905-6; and Wairagkar M, et al. An instantaneous voice-synthesis neuroprosthesis. Nature (2025), doi:10.1038/s41586-025-09127-3. Two one-participant studies
- Card NS, et al. Long-term independent use of an intracortical brain-computer interface for speech and cursor control. Nature Medicine (2026). doi:10.1038/s41591-026-04414-6. One-participant study
- Flesher SN, et al. A brain-computer interface that evokes tactile sensations improves robotic arm control. Science (2021). doi:10.1126/science.abd0380. One-participant study
- Lorach H, et al. Walking naturally after spinal cord injury using a brain–spine interface. Nature (2023). doi:10.1038/s41586-023-06094-5. One-participant study
- Chandrasekaran S, et al. A neuroprosthesis for restoring hand movement and sensation in a person with complete tetraplegia. Nature Medicine (2026). doi:10.1038/s41591-026-04498-0. One-participant study
- Vansteensel MJ, et al. Longevity of a Brain–Computer Interface for Amyotrophic Lateral Sclerosis. New England Journal of Medicine (2024). doi:10.1056/NEJMoa2314598. Longitudinal case report
- Okun MS, et al. Definition of Implanted Neurological Device Abandonment: A Systematic Review and Consensus Statement. JAMA Network Open (2024). doi:10.1001/jamanetworkopen.2024.8654. Review & consensus
- Rubin DB, et al. Interim Safety Profile From the Feasibility Study of the BrainGate Neural Interface System. Neurology (2023). doi:10.1212/WNL.0000000000201707. Multicentre safety case series
- UNESCO. Recommendation on the Ethics of Neurotechnology (adopted 11 November 2025). Nonbinding UNESCO recommendation
- Branco MP, et al. Brain-Computer Interfaces for Communication: Preferences of Individuals With Locked-in Syndrome. Neurorehabilitation and Neural Repair (2021). doi:10.1177/1545968321989331. User-preference study
- OECD. Recommendation of the Council on Responsible Innovation in Neurotechnology (2019). Nonbinding OECD recommendation
- European Union. Regulation (EU) 2017/745 on medical devices; and Regulation (EU) 2016/679—the General Data Protection Regulation. Binding regulations
- European Union. Regulation (EU) 2024/1689—the Artificial Intelligence Act, as amended in timing by Regulation (EU) 2026/1744. Binding regulations
- Biblioteca del Congreso Nacional de Chile. Ley 21.383 (published 25 October 2021), constitutional amendment concerning scientific and technological development, integrity, brain activity and information; and Cámara de Diputadas y Diputados de Chile. Boletín 13.828-19 legislative file. National law & official bill record
- Colorado General Assembly. HB24-1058: Protect Privacy of Biological Data (2024). State law
- California Legislature. SB-1223: Consumer privacy—sensitive personal information—neural data (2024). State law
- Kunz EM, et al. Inner speech in motor cortex and implications for speech neuroprostheses. Cell (2025). doi:10.1016/j.cell.2025.06.015. Four-participant intracortical study
- Hampson RE, et al. Developing a hippocampal neural prosthetic to facilitate human memory encoding and recall. Journal of Neural Engineering (2018). doi:10.1088/1741-2552/aaaed7. Small experimental stimulation study
- National Medical Products Administration of China. First invasive brain–computer-interface medical device approved for marketing (official notice, 13 March 2026; Chinese). Official regulatory notice