Brain-Computer Interfaces

A brain-computer interface is usually described by what it is made of — the electrode, the channel count, the surgical route. That is the device layer, and its sibling brief Neural Interfaces carries it. This brief asks the question that decides whether any of it matters: what are these things used for, how many people use each one, and what is the honest clinical endpoint in each case. Answered that way the field reorganises around a single unflattering fact. The only neural prosthesis with a six-figure deployed base is the cochlear implant — roughly 600,000 recipients worldwide — and a cochlear implant does not record from the brain at all. It stimulates the auditory nerve. Every device the press calls a brain-computer interface has a deployed base in the tens.

Two structural facts follow, and this brief is organised around them. The first is that the record results in the application people care about most — restoring speech — are each a single participant, in a different person, at a different institution, scored on a different metric, with no shared test set and no replication. “The record” in speech decoding is not a coherent object; it is a small set of incomparable demonstrations that the field and its readers have agreed to line up in a row. The second is that the most instructive failure in the whole cluster was not a failure of signal. Argus II was a retinal implant with a CE mark, an FDA humanitarian device exemption, a published thirty-patient pivotal trial and a price near $150,000, and in 2020 its manufacturer stopped supporting it while the devices were still inside people. A BCI is not only an interface. It is a product with a company attached, and the company is on a shorter clock than the implant.

Established Start with the denominator, because almost every published projection of BCI adoption is missing one. The cochlear implant reached about 600,000 recipients worldwide as of 2016 and 217,000 in the United States through the end of 2019. Established Device failure requiring reimplantation is estimated at 2.5–6%, revision surgery at 3.8% to 8%, and children are implanted from as young as nine months. Frontier Those figures reach this brief through a tertiary index rather than the primary epidemiology and should be traced before being treated as exact. Established What survives the caveat is the shape: an implanted neural device can be manufactured, reimbursed, revised and lived with at population scale. Established It also took roughly half a century from first demonstration to that scale, and the device does not record. Frontier Every recording interface in human use today is investigational, and the total number of people who have ever carried a chronic intracortical recording array is in the low hundreds at most.

Established Communication restoration is the application with the strongest results and the smallest number of participants, and the two facts are the same fact. The laboratory ceiling is carried in Human-Machine Symbiosis and this brief does not re-derive it: Frontier 62 words per minute at 9.1% and 23.8% word error rate on 50-word and 125,000-word vocabularies (Willett et al. 2023), Frontier 78 words per minute at about 25% word error rate with synthesised voice and an avatar (Metzger et al. 2023), Frontier and 97.5% accuracy sustained over 8.4 months at roughly 32 words per minute after a thirty-minute calibration (Card et al. 2024). Established Each of those is n=1 — three different people, at three institutions, on three vocabularies, scored by three protocols, with no common test set and no independent replication of any of them. Frontier Lining them up as a rising curve is a rhetorical act, not a measurement: 62 and 78 and 32 are not points on the same axis, and the one number that is directly comparable across them — how long the system kept working — only Card et al. report. Established Natural conversational speech runs near 160 words per minute, so even the most generous reading has the best laboratory result at about half of speech, once, in one person.

Established The floor of the same application is a factor of about 150 below the ceiling, and it is the version that actually went home. Vansteensel and colleagues implanted subdural electrodes over motor cortex with a subcutaneous thoracic transmitter in a patient with locked-in ALS (New England Journal of Medicine 375:2060, 2016). Established By attempting to move the contralateral hand, the patient controlled a typing program accurately and independently 28 weeks after electrode placement, at an equivalent of two letters per minute, and the system offered autonomous communication that at times supplanted the patient’s eye-tracking device. Established Two letters per minute is about 0.4 words per minute. Frontier It is also fully implanted with no percutaneous connector, it ran at home, and it is one of very few results here that names the assistive technology it displaced. Established The gap to the laboratory record is roughly 150-fold in throughput, and the deployed system is the one that survived contact with a house.

Established Communication in the completely locked-in state has been demonstrated once, and the group carries a research-integrity history the brief has to state. Chaudhary et al. (Nature Communications 13:1236, 2022) implanted two 64-microelectrode arrays in supplementary and primary motor cortex of a patient in a completely locked-in state with ALS; the patient modulated firing rates using auditory feedback and selected letters one at a time to form words and phrases. Established Separately, the German Research Foundation investigated a 2017 study from the same senior author claiming communication in four completely locked-in patients and found the research data incomplete and the results flawed, with reported consequences including funding revocation and dismissal from the University of Tübingen; the parties settled in April 2022, and the authors claim to have won litigation supporting the earlier report’s integrity. Frontier The retraction status of that 2017 paper is not established here — an attempt to fetch it returned a different article entirely — and this brief makes no claim about it. Established The episode belongs in the record precisely because completely-locked-in communication is the claim a reader most wants to be true.

Established Motor restoration has three routes, and the interesting result is that the one with a write channel is the one that got faster. Route one drives the person’s own muscles: Bouton et al. (Nature 533:247, 2016) decoded motor cortex in real time to control a neuromuscular stimulation sleeve in a 24-year-old man with cervical spinal cord injury, producing six distinct wrist and hand movements including isolated finger movements, with clinical assessment showing improvement from a C5–C6 to a C7–T1 neurological level. Established Route two drives an external robot: Flesher et al. (Science 372:831, 2021) added intracortical microstimulation of somatosensory cortex to a motor decode and halved median trial times, 20.9 to 10.2 seconds, the gain coming almost entirely from less time spent attempting to grasp, in one participant. Established Route three drives reach and grasp by brain-controlled functional electrical stimulation (Ajiboye et al., The Lancet 389:1821, 2017); this brief prints no numbers from it, because no abstract was deposited and the publisher is unreachable. Frontier The strongest application-layer result in the cluster is a fourth thing: the brain-spine interface of Lorach et al. 2023, carried by Neuroplasticity Engineering, whose participant regained walking with crutches with the device switched off, WISCI II improving from 6 to 16 after a three-year plateau — the only result here shaped like disease modification rather than prosthesis, and the only endpoint shape a payer routinely reimburses.

Established Sensory restoration splits into one mature case — the cochlear implant above — and one cautionary one, and neither is a cortical BCI. Established The cortical visual work is genuine and small: Fernández et al. (Journal of Clinical Investigation 131(23), 2021) placed a 96-channel intracortical array in the occipital cortex of one 57-year-old participant with complete blindness for six months, measured a mean single-electrode threshold of 66.8 ± 36.5 µA, found that simultaneous multi-electrode stimulation significantly reduced thresholds (P < 0.001) rather than interfering, and obtained discriminable phosphenes permitting identification of letters and recognition of object boundaries. Frontier Recordings from visually deprived neurons were of high quality — a non-trivial negative result about cortical disuse. Established The cautionary case is Argus II: a retinal prosthesis with 60 electrodes, 55 enabled, in a 6×10 grid, CE marked March 2011 and granted an FDA humanitarian device exemption in February 2013 for up to 4,000 patients per year. Established Its pivotal single-arm trial enrolled 30 patients with severe retinitis pigmentosa, longest follow-up 38.3 months; with the device on, 23% showed visual-acuity improvement and 60% walked accurately to a door 20 feet away against 5% with the device off. Established There were nine serious adverse events among the 30 subjects, and the price was about $150,000 excluding surgery and training. Established In 2020 Second Sight ceased technical support with the devices still implanted; an IEEE Spectrum investigation reported that users risk, and in some cases have already experienced, a return to blindness. Frontier Second Sight merged with Nano Precision Medical in August 2023 with a commitment to restore support for existing patients. Established The failure mode was not the electrode. It was the company.

Established Treating disease is where the deployed base actually is, and it is stimulation, not decoding. Deep brain stimulation passed 244,000 implants worldwide by 2019, with approvals running from tremor in 1997 to epilepsy in 2018; those numbers, the effect sizes and the complication rates are carried in full by Human-Machine Symbiosis and are not re-derived here. Frontier What this brief adds is the psychiatric frontier case: Scangos et al. (Nature Medicine 27:1696, 2021) used multi-day intracranial electrophysiology and focal stimulation to identify a personalised, symptom-specific biomarker — gamma power across amygdala, orbitofrontal cortex, subgenual cingulate and hippocampus — and a treatment site in right ventral capsule/ventral striatum, before implanting a chronic sensing-and-stimulating device, reporting rapid and sustained improvement in depression. Established The authors state in the abstract that whether this n-of-1 approach generalises is unknown. Frontier The methodological point outruns the clinical one: the biomarker was discovered in the individual rather than imported from the literature, which is a different research design from every other implant programme in this cluster and is expensive per patient in a way that does not obviously scale.

Established Now the arithmetic that governs how any of this should be read. Across every result above and in the sibling briefs, the participant counts are: n=1 for Willett 2023, Metzger 2023, Card 2024, Bouton 2016, Flesher 2021, Fernández 2021, Scangos 2021, Chaudhary 2022, Vansteensel 2016 and Lorach 2023; n=4 for the SWITCH endovascular study and for the randomised adaptive-DBS feasibility trial; n=25 for the closed-loop memory-stimulation study; n=30 for the Argus II pivotal; Established and n=14 for every BrainGate participant who has ever been implanted, across twenty years. Established That is the largest interventional intracortical BCI dataset in existence. Established Set it beside 600,000 cochlear implants and 244,000 deep-brain stimulators and the field’s evidentiary position is legible: a strong mechanistic case, a handful of spectacular individual outcomes, and no population data of any kind. Established The regulatory shape follows directly. Every approved neural implant in every major jurisdiction — deep brain stimulation, vagus nerve stimulation, responsive neurostimulation, cochlear implants, Argus II — is a therapy for a named disease, while the Stentrode, the Neuralink N1 and the BrainGate arrays all operate under investigational device exemptions. Established There is no approval pathway anywhere for a neural implant whose indication is the augmentation of a healthy person.

Frontier The most consequential new datum in this subject is a scaling curve, and it points the wrong way for the industry that markets channel count. Hahn et al. (medRxiv 2025.07.02.25330310) pooled twenty years of BrainGate data — 2,319 recording sessions, 20 arrays, the first 14 participants, up to 7.6 years — and reported that decoding signal-to-noise ratio increases logarithmically with the number of electrodes. Frontier Neural Interfaces owns that finding; the application-layer consequence belongs here. Established Logarithmic growth is simultaneously the authors’ stated “pathway for scaling performance” and the mathematical form of diminishing returns: ten times the channels buys a fixed increment, not ten times the capability. Frontier If that curve holds, capability gains cannot come from channel count. They have to come from the decoder, and from choosing what to decode. Frontier The same paper reports that 11 of 14 arrays provided meaningful movement decoding throughout enrollment — the first time anyone could say what fraction of implanted arrays stay useful and for how long. Frontier It is a preprint, unreviewed when this brief was written.

Frontier The decoder has become a research object in its own right, and it does much of the work the hardware is credited with. Card et al. 2024 reduced calibration to about thirty minutes and held 97.5% accuracy across 8.4 months — a claim about model stability and re-training schedules rather than about electrodes. Frontier Neuralink’s first participant lost roughly 85% of threads and recovered most performance through software changes — the strongest available demonstration that a decoder can absorb catastrophic hardware loss. Frontier Modern speech decoders lean on language models as priors, so a share of the reported word-error-rate improvement is progress in language modelling rather than in neuroscience, and no published result separates the two. Speculative If channel count scales logarithmically and decoders do not, the field’s optimisation target is mis-set: the marginal research dollar buys more in the decoder and in the choice of decoded variable than in the array. Frontier That choice is already doing real work — handwriting beat cursor control by changing what was decoded, not what was implanted.

Frontier Non-invasive BCI has a measured ceiling, and it sits closer to the invasive route than the field admits. The highest reported information transfer rate for a non-invasive BCI in a recent clinical review is 360 bits per minute, with pooled P300-speller accuracy of 74%; carried by Human-Machine Symbiosis, it corresponds to roughly 60–70 English characters a minute at typical entropy — approximately the intracortical handwriting rate, without surgery. Frontier Non-invasive stimulation now also reaches deep structures: Violante et al. (Nature Neuroscience 26:1994, 2023) demonstrated temporal-interference stimulation of the human hippocampus, validated by field modelling and in a human cadaver, then tested in n=20 for imaging and n=21 for behaviour, with a 12% increase in the odds of correct recall (95% CI 0.02–0.21) — the same size as the implanted result, as section 14 sets out. Speculative Either the invasive route is being wasted on that application, or both are measuring the same small, non-specific arousal-like effect, and no experiment currently distinguishes those readings.

Frontier The write side is behaving better than the field’s own theory predicted. Flesher 2021 halved task times by adding a percept; Fernández 2021 found simultaneous multi-electrode stimulation lowers thresholds rather than summing into interference; and chronic stimulation over 1,500 days in one participant saw detection thresholds fall. Frontier Against that, Human-Machine Symbiosis argues that writing is under-specified and may be harder than reading, because a stimulation pattern has no ground truth to be scored against. Speculative The untested reconciliation is that low-dimensional percepts — contact, edge, phosphene — are easy to write and high-dimensional ones are not, in which case both positions are right about different targets.

Frontier Two capabilities arrived that nobody requested. Inner speech is robustly represented in motor cortex and imagined sentences are decodable in real time in four participants — carried by Human-Machine Symbiosis, along with the “mental password” gating the same group published alongside it, the best available evidence about how neurotechnology governance behaves: well, voluntarily, without a regulator. Speculative The second is passive BCI, reading state rather than command, which needs few channels, tolerates high error, and is therefore the application with the largest plausible market and the worst ethics. Frontier Meanwhile human single-neuron recording at Neuropixels density is routine intraoperatively — over 200 well-isolated cortical single units at once (Paulk et al., Nature Neuroscience 25:252, 2022) — but acutely only: no chronic human Neuropixels implant has been reported. Frontier And the application nobody has solved is the unglamorous one: there is still no shared benchmark task, no common test set and no agreed clinical endpoint across groups, and Hahn’s decoding-SNR metric is the first candidate for a cross-participant common currency. Speculative Whether the field adopts it is the thing to watch, because a field that cannot compare its own results cannot be wrong in public.

Established The workback target is a BCI a healthy adult would rationally choose, that is reimbursed, and that has a deployed base above 100,000. The chain runs: (1) a clinical indication with a population above 100,000 that a low-channel device serves; (2) safety at one year established in at least twenty people; (3) a published cross-group benchmark with a shared test set; (4) performance stability without recalibration for ninety days; (5) chronic yield at five years across at least twenty implants; (6) a revision and explantation protocol; (7) a device-support obligation that survives the manufacturer; (8) a reimbursement code; (9) a regulatory pathway for a non-therapeutic indication; (10) a demonstrated advantage over the best non-implanted alternative on a task a healthy adult cares about.

Established Links 7, 8 and 9 are institutional and they bind today. Even a perfect device has no door: there is no reimbursement decision for any BCI indication in any major payer system, no cleared predicate whose indication is augmentation, and no instrument anywhere requiring that an implanted device’s servicing outlive the company that sold it. Established Argus II is the proof of need for the last of those and it is eight years old. Frontier Link 10 binds ultimately and is scientific: no published BCI result would beat a keyboard for a healthy adult, no head-to-head against a keyboard, a touchscreen or an eye-tracker has ever been attempted, and the field’s own physical-limits analysis says the non-invasive route cannot be pushed far enough to change that.

Frontier Link 5 is the interesting one, because it just moved and almost nobody has noticed. Chronic yield was the field’s stated bottleneck for twenty years; Hahn et al. 2025 now supply it for 20 arrays in 14 people, and the answer is better than the animal literature predicted. Speculative A field whose declared bottleneck quietly dissolves should become suspicious of the rest of its bottleneck list, and this one deserves that suspicion: several items on it are engineering folklore rather than measurements. Frontier Link 6, revision, is the sleeper. Established Nobody has published a successful replacement of a failed intracortical array in a human with recovery of prior performance. Established Cochlear implants have a 2.5–6% reimplantation rate and a mature revision practice built around it; intracortical BCI has neither, Speculative and revision is unglamorous, entirely tractable, and unfunded — the highest-leverage cheap item in the cluster. Frontier Link 1 is the quiet structural failure: the field selects participants, not indications. Trials recruit an individual who can be studied, not a population that could be served, which is why no BCI programme can state its addressable population and why link 8 has nothing to price.

Established This brief waits on the device layer for exactly one thing: the joint distribution of channel count, chronic yield and years. Neural Interfaces supplies it, and every application claim here is downstream of it — the logarithmic scaling result, the 35.6% mean human electrode yield with only a 7% decline across enrollment, the meningeal-fibrosis and connector failure modes, and the absence of any published human chronic yield for any high-channel flexible implant. Frontier That last absence is a gap in the world rather than in the literature search, and it is the single fact that would most change this brief.

Established The second dependency is on decoding as a discipline: non-stationarity across days, transfer across participants, and the contribution of language-model priors to reported word error rates. Frontier None of these has a shared evaluation, so progress in them is currently unmeasurable in principle rather than merely unmeasured. Established The third is on Neuroplasticity Engineering, which owns the off-state recovery result that gives this field its only disease-modification-shaped endpoint, Speculative and whose warning that movement-contingent stimulation can drive maladaptive as well as adaptive change applies directly to any BCI judged on what the patient can do with it switched off.

Established The fourth is institutional and sits outside science entirely: health-technology assessment methods for a one-time implanted device with a decades-long support tail, which no jurisdiction has built. Frontier The fifth is the one the field rarely admits — a dependency on not being made obsolete. Speculative If spinal repair, gene therapy or regenerative medicine restores the biological pathway, the prosthesis is unnecessary, and a decade-stable implant is a much weaker investment if the disease it addresses is curable in fifteen years. Frontier That argument is live inside funding bodies and almost never appears in public writing about the field.

Established Six experiments would settle more than the next decade of channel-count increases. Frontier One: the head-to-head nobody runs. Take the best current BCI in its best participant and score it against eye-tracking, a switch scanner and a touchscreen on the same communication task, same scorer, same day, reporting words per minute, error rate, setup time and fatigue. Established The measurement is trivial, the equipment costs nothing, and it has never been published; the closest anything comes is Vansteensel’s note that a two-letters-per-minute implant supplanted the participant’s eye-tracker at times — which is a statement about eye-tracking’s failure modes in advanced ALS, not about throughput.

Frontier Two: a shared benchmark with a held-out test set. Two independent groups reporting on the same task with the same scoring, with the test set published before the systems are built. Speculative Hahn’s decoding-SNR metric is the first candidate currency and adopting it costs nothing but status. Frontier Three: safety at one year in at least twenty people, reported as a serious-adverse-event rate with a confidence interval. Established The best that currently exists is n=4 from the endovascular SWITCH study, which reported no serious adverse events at one year with home use beyond a year, at 16 electrodes against the N1’s 1,024. Established This link binds everything downstream, because no payer prices a device whose harm rate has an interval spanning an order of magnitude.

Frontier Four: decoder stability without retraining. Plot task accuracy against days since the last decoder update, for ninety days, in more than one person. Established Card et al. 2024’s 97.5% over 8.4 months is the closest anything has come and it is one participant with an unreported retraining schedule. Frontier Five: a revision protocol. Explant a failed array and implant a replacement in a human, and report whether prior performance returns. Established No such report exists. Frontier Six: the indication-first study. Pair an incidence or prevalence estimate for a candidate indication — severe dysarthria of any cause rather than ALS alone, refractory epilepsy warning, severe chronic pain — with a demonstrated function at 32 channels or fewer. Speculative That experiment is epidemiology plus a modest device, it would tell the field which market it is actually in, and nobody is running it because it is not a publication in a high-impact journal.

Established The first engineering requirement is the one the failure data actually name: get rid of the connector. In the founding non-human-primate failure analysis, acute mechanical failures were the most common single category and 83% of those were connector issues — carried in detail by Neural Interfaces. Established The fully implanted, wirelessly transmitting systems in this cluster are the ones that went home and stayed: Vansteensel’s ECoG typist, the Stentrode, the WIMAGINE epidural implants behind the brain-spine interface. Frontier Percutaneous pedestals are a laboratory affordance that the application layer cannot ship.

Frontier The second is calibration burden, which is a product requirement disguised as a research parameter. A system needing a technician and an hour before use is not an assistive device; a thirty-minute self-service calibration, as reported by Card et al., is the first figure in this literature that reads like a specification rather than a result. Speculative The target that would matter is zero — a decoder that is stable across days and does not need the user to produce labelled data at all — and no published system reaches it.

Established The third is that the applications with the largest addressable populations live below roughly 100 channels. Switch access, environmental control, cursor control, binary communication and state monitoring are all reachable at 16 to 32 electrodes, which is the regime where an endovascular or epidural route with a one-year safety record is available. Frontier A 60-electrode retinal implant and a 96-electrode cortical implant both delivered letter-level vision; the spread from a working clinical device to a research probe is two orders of magnitude in channels and nothing like two orders of magnitude in clinical outcome. Established Fourth, power and thermal budgets bound the whole enterprise: the field’s own physical-limits analysis puts roughly 40 mW of tolerable thermal load in a mouse brain and about 100 Gb/s for 75 million neurons at 1 kHz, figures carried by Human-Machine Symbiosis. Speculative Fifth, and least discussed: an implanted device needs a firmware update path that works for twenty years, and no programme has published one.

Established The cochlear implant is the adjacent technology that matters most and is cited least usefully. It supplies the manufacturing, audiology, revision-surgery and reimbursement template that a recording interface would need, Frontier and it supplies the honest caution that a stimulating interface with a well-understood target organ still took about fifty years to scale. Established Deep brain stimulation, vagus nerve stimulation and responsive neurostimulation form the second adjacency: they are the proof that implanted, chronically powered neural hardware is a normal part of medicine, and their approval histories are the only real template for a BCI regulatory submission. Frontier Adaptive DBS is the live case, having crossed from research device to approved product between 2021 and 2025, with a real-world open-label series finding 3 of 8 patients individually improved — carried by Human-Machine Symbiosis.

Established The comparator set that BCIs are supposed to beat is assistive technology, and it is where the field’s claims are weakest. Eye-tracking, sip-and-puff, single-switch scanning, head-mouse and partner-assisted scanning are mature, cheap, non-surgical and in daily use by far more people than every BCI participant in history combined. Frontier They fail in specific, known circumstances — oculomotor involvement in late ALS, fatigue, positioning, caregiver availability — and those circumstances, not throughput, define the honest BCI indication. Speculative A field that measured itself against this comparator would be organised around a much smaller and much more defensible claim than the one it currently makes.

Established Functional electrical stimulation and exoskeletons are the effector-side adjacency, with robot-assisted gait training meta-analysed at standardised mean differences of 0.39–0.57 on function and 0.03 on strength across 8 randomised trials and 241 participants. Frontier Language modelling is an adjacency the field does not name: modern speech decoders are neural front ends on statistical language models, and their improvement curve is partly borrowed. Frontier Non-invasive deep stimulation by temporal interference is the newest adjacency and the most disruptive if it holds. Speculative And the adjacency that could end the subject is biological repair — spinal-cord regeneration, gene therapy for inherited retinal disease, hair-cell regeneration — each of which removes an indication rather than serving it.

Established Three institutional objects do not exist, and their absence is the binding constraint on this field today. The first is a reimbursement code. Established No payer in any major jurisdiction has made a coverage decision for a brain-computer interface indication, which means no BCI has a price, and a device without a price cannot have a market, a manufacturing plan or a service contract. Established The second is a regulatory pathway for a non-therapeutic indication: every approved neural implant anywhere is authorised against a named disease, and no regulator has published a framework under which an augmentation claim could even be filed. Frontier The third is the one Argus II proves is needed and nobody has built: a device-support obligation that survives the manufacturer.

Frontier An implanted device creates a maintenance obligation measured in decades against a company lifespan measured in years. Every current high-channel programme is venture-funded and pre-revenue. Speculative The fixes are unglamorous and entirely tractable: an escrowed device-support fund sized to the implanted base, a mandated open hardware and firmware specification held by a third party, or a condition of approval that servicing obligations transfer on insolvency. Established None exists in any jurisdiction. Speculative This is the highest-leverage low-cost intervention available to any government that wants a neurotechnology policy, and it is not on any published agenda.

Established The fourth institutional problem is the field’s own evidentiary culture. There is no shared benchmark, no common test set, no registry of implanted participants and no agreed clinical endpoint; every paper compares to its own prior. Frontier The consequence is that twenty years of results cannot be pooled, which is exactly why the Hahn 2025 dataset had to be assembled retrospectively by a consortium rather than read off a register. Speculative A mandatory implant registry, of the kind orthopaedics and cardiac devices have had for decades, would cost almost nothing and would answer more questions than the next several trials. Frontier Fifth, consumer and non-medical neurotechnology sits in a regulatory vacuum that data-protection law is beginning to fill ahead of device law — constitutional and statutory neural-data protections are widely reported, and international bodies have active work programmes. Frontier This brief verified no jurisdiction, date, statute or case name in that area and prints none; the substance is real and the specifics here are unconfirmed.

Established The central ethical fact in this field is not privacy. It is abandonment. Argus II patients gave informed consent to a surgical implant on the understanding that a company would stand behind it, and in 2020 that company simply stopped, leaving the devices in place and unsupported. Speculative No consent form in the field discloses the probability that the sponsor will not exist for the life of the implant, and no ethics committee requires it — remarkable, because it is the most likely adverse event for any device from a pre-revenue company. Frontier The honest disclosure would read: this device has an expected service life longer than the expected life of the company that made it, and there is no mechanism by which its support survives.

Frontier Mental privacy is the second issue and it is now a real capability rather than a thought experiment. Inner speech is decodable in a research setting in four participants, and the group that showed it published a gating mechanism in the same paper — carried by Human-Machine Symbiosis, alongside Cognitive Liberty. Speculative The harder case is passive BCI: a command interface is used by its subject, while a state interface is typically deployed on a subject by an employer or an insurer, and the inner-speech result establishes that the boundary between reading state and reading content is porous. Frontier The consent architecture governing medical implants does not transfer to a workplace fatigue monitor at all.

Established Third, the research participants themselves. Fourteen people carry the entire interventional record for high-performance intracortical BCI, each of them severely disabled, each undergoing craniotomy for a study whose direct benefit to them is uncertain and whose device is usually removed at the end. Frontier Their burden is enormous, their number is tiny, and the field’s publication incentives push toward spectacular single-participant demonstrations rather than the boring replication that would make their contribution cumulative. Speculative Fourth, access: a $150,000 device with nine serious adverse events in thirty patients and no reimbursement code is, whatever else it is, an intervention available to almost nobody, and every equity argument about neurotechnology has to start from that rather than from a hypothetical future of enhanced elites.

Speculative The most probable civilizational implication of brain-computer interfaces is a small one, and this brief is willing to say so. The likely end state is a mature clinical speciality resembling cochlear implantation: tens of thousands of recipients, a defined indication, a revision practice, an audiology-equivalent profession, and no cultural significance beyond the lives it changes completely. Frontier That outcome would be a major medical achievement and a total failure of the narrative that has funded the field.

Speculative The larger version requires a specific thing that has never happened: a healthy adult rationally choosing an implant. The steelman is real — the same was said of laser eye surgery and cosmetic dentistry, a sufficiently low-risk route changes the arithmetic, and there are genuine healthy-adult applications in high-stakes vigilance, communication in noise-isolated environments and control of complex machinery. Frontier Against it: the risk floor of any implant is not zero, the demonstrated throughput advantage over a keyboard is negative, there is no regulatory pathway, and the physical-limits literature says the non-invasive route cannot be pushed far enough to close the gap. Speculative The honest terminal position is a declared tie on the value question and a clear verdict on the timing one: nothing in the current evidence base supports a consumer market this century, and the argument that one arrives anyway rests on analogy rather than on measurement.

Speculative The reframing that would matter most is not about bandwidth at all. If the correct endpoint for most BCIs is plasticity rather than function — if the device should be judged by what the patient can do with it switched off — then the field is optimising a metric no payer reimburses, and the brain-spine interface result is not an outlier but the template. Handwave At the far horizon, coupled multi-user interfaces and shared representational formats between brains remain coherent to state and nowhere near evidence; nothing in this brief’s source base bears on them at all.

These horizons track deployed populations and institutional objects rather than laboratory records, because the laboratory record has moved faster than anything downstream of it for twenty years.

• 10 yr: Frontier A cross-group benchmark with a shared test set exists and at least two consortia report on it; one-year safety is established in twenty or more people on at least one low-channel platform; and the first payer coverage decision for a BCI indication is made, most plausibly for communication in severe dysarthria rather than for motor control. Frontier Chronic-yield and stability data extend to a second cohort outside BrainGate. Speculative The surprise on this horizon would be a published head-to-head against eye-tracking that the BCI loses, which would be the most useful paper the field could produce and the one it has the least incentive to write.
• 25 yr: Speculative A deployed base in the low tens of thousands for one or two indications, on a low-channel fully implanted platform, with a revision practice and a reimbursement code. Speculative A device-support obligation that survives the manufacturer exists in at least one jurisdiction, prompted by a second Argus-shaped failure rather than by foresight. Frontier Decoders transfer across participants without individual calibration, which is the single capability change that would most alter the economics.
• 50 yr: Speculative Either the field looks like audiology — a settled speciality with a six-figure deployed base and no cultural charge — or one or more of its indications has been removed outright by biological repair, and the prosthesis is a bridge technology that was crossed. Speculative Both outcomes are consistent with everything currently measured, and this brief declines to choose between them. Handwave A regulatory pathway for a non-therapeutic indication may exist by this point; nothing in the current evidence suggests what would motivate a regulator to build one.
• 100 / 250+ yr: Handwave A general-purpose interface — one that reads intention across domains rather than decoding a trained variable, and writes percepts of arbitrary dimension — requires a forward model of cortical representation that does not exist and is not being built. Handwave The strongest argument that it is achievable is that the brain is a physical system with no known exemption; the strongest against is that the only published scaling curve is logarithmic and the only demonstrated write targets are one-dimensional. Handwave Multi-user coupled interfaces belong on this row and no further: everything shown so far is a BCI wired to a stimulator transmitting a bit.

• Depends on The device layer, carried by Neural Interfaces, and specifically its joint distribution of channel count, chronic electrode yield and years in a human. Every application claim in this brief is a claim about what can be decoded from a signal whose durability that brief measures: the 35.6% mean human electrode yield with a 7% decline across enrollment, the connector as the leading acute failure mode, the meningeal-fibrosis mechanism that replaced the glial-scar folk model, and above all the logarithmic relationship between decoding signal-to-noise ratio and electrode count. That last result is the hinge. If it holds, no amount of device progress in the currently contested 100-to-10,000 channel range produces a proportional application gain, and the application layer has to advance through decoders and through the choice of decoded variable instead. The dependency is asymmetric and worth stating plainly: better devices would help this brief modestly, whereas a published human chronic yield for a high-channel flexible implant — which does not exist anywhere — would change it substantially.
• Enables This brief supplies the demand signal for three others. Neuroplasticity Engineering inherits the question of whether a device should be judged on what the patient can do with it switched off, which is the only endpoint here shaped like disease modification. Memory Engineering inherits the finding that non-invasive temporal-interference stimulation and implanted closed-loop stimulation produce memory effects of indistinguishable size, which bears directly on whether the invasive route is worth its risk for that application. Cognitive Liberty inherits the inner-speech decoding result and the passive-BCI market argument, which together make mental privacy a present engineering question rather than a future one. Downstream of all three sits the institutional demand this brief cannot satisfy on its own: a reimbursement pathway, an implant registry, and a servicing obligation that outlives a company.
• Adjacent Cochlear implantation is the closest adjacent practice and the correct denominator, supplying the manufacturing, revision and reimbursement template at four orders of magnitude more scale. Deep brain stimulation and responsive neurostimulation are the regulatory template. Assistive technology — eye-tracking, switch scanning, sip-and-puff — is the comparator set every BCI claim should be scored against and almost never is. Functional electrical stimulation and exoskeletons are the effector side. Language modelling is an unacknowledged adjacency inside every modern speech decoder. And biological repair — spinal regeneration, retinal gene therapy, hair-cell regeneration — is the adjacency that removes indications rather than serving them, which makes it the strongest long-horizon argument against building a decade-stable implant at all.

Handwave “BCIs are approaching natural-speech bandwidth.” The laboratory record is 78 words per minute against natural conversation at about 160, in one participant, unreplicated. Established The deployed, fully implanted system that ran at home for years managed two letters per minute, about 0.4 words per minute. Established Both are true, they differ by a factor of roughly 150, and a reader who has met only the first has a picture of the field that is wrong by two orders of magnitude at the point of care.

Handwave “Channel count is the capability metric.” The only published curve relating decoding performance to electrode number is logarithmic. Established A 60-electrode retinal implant and a 96-electrode cortical implant both delivered letter-level vision. Frontier Neuralink’s first participant lost about 85% of threads and recovered most performance in software. Established The steelman is real: logarithmic in signal-to-noise need not be logarithmic in achievable applications, and there may be a threshold task — dexterous multi-finger control, high-fidelity somatosensation, naturalistic speech — that fails below some channel count and works above it. Speculative None has been demonstrated and none ruled out. Frontier What can be said is that the field spent a decade asserting a relationship it never measured, and the first measurement went against it.

Handwave “The results are validated in patients.” The complete interventional record for high-performance intracortical BCI is 14 people over twenty years, plus a handful of single-participant demonstrations elsewhere; the largest randomised evidence for adaptive deep brain stimulation is n=4. Frontier The phrase “n=1” does more work here than any other, and the pattern is more persuasive than any single correction: the 62-words-per-minute speech BCI, the 78-words-per-minute avatar BCI, the 97.5%-over-8.4-months speech BCI, the tactile-feedback robotic arm, the cortical visual prosthesis, the brain-spine interface, the closed-loop depression device, the completely-locked-in speller, and the fully implanted ECoG typist. Established All nine appear routinely in public writing without the qualifier.

Handwave “An approved device is a supported device.” Argus II had a CE mark, an FDA humanitarian device exemption, a published 30-patient pivotal trial and a $150,000 price, and its manufacturer stopped technical support in 2020 with the devices still implanted. Frontier This brief reaches that history through an encyclopaedic index citing an IEEE Spectrum investigation not read directly, which should be traced before the account is settled. Established The approval was never a promise about year twenty-five.

Frontier “BCIs read thoughts.” Deployed systems decode attempted movement or speech, and the standard correction has been that this differs from reading a thought. Frontier That correction is weaker than it was: inner speech is robustly represented in motor cortex in four participants and imagined sentences can be decoded in real time. Established The accurate statement is that they can, partially, in a research setting, and that the group which found it published the mitigation alongside it.

Frontier “BCIs are a rich person’s toy with no clinical value.” This is the sceptic’s error and it fails on the record: a person with ALS communicating at 62 words per minute, a tetraplegic participant whose neurological level improved from C5–C6 to C7–T1, a blind participant reading letters, and a paraplegic participant walking with the stimulator switched off are all real, peer-reviewed clinical outcomes. Frontier They are each one person, and one person is not nothing.

Frontier “Invasive beats non-invasive, obviously.” On memory, a scalp-applied temporal-interference protocol produced a 12% increase in the odds of correct recall in 21 healthy people, and implanted closed-loop stimulation a 15% relative increase in recall probability in 25 epilepsy patients. Established Those are the same size. Established On typing, a non-invasive P300 speller reaches 360 bits per minute with 74% pooled accuracy, the same range as the intracortical handwriting result. Speculative The invasive route’s advantage is real for continuous, high-degree-of-freedom control and for anyone whose residual motor function cannot drive a non-invasive system; it is not general, and stating it as general is a common error among the otherwise well informed.

Frontier “The cochlear implant proves neural interfaces scale.” It proves a stimulating interface scales: it does not record, its target is a peripheral nerve with a well-understood tonotopic code, and the problem it solved is not the one a recording BCI faces. Established “Closed-loop means the device understands you.” In every closed-loop system in this cluster the loop is closed on a scalar biomarker — gamma power, a spectral feature, an encoding-state classifier — not on content, and the authors of the depression result call it an n-of-1 of unknown generality.

Frontier “None of this has changed since 2006.” Word error rate on a 125,000-word vocabulary fell from unmeasurable to about 24% and then to about 10% in roughly a year; a fully implanted wireless system reached parity with a wired one; a bidirectional interface halved task times; and the first twenty-year yield dataset appeared in 2025. Established The rate of change is real; what has not changed is the number of participants.

Established Three source-level corrections this brief is obliged to carry. First, a sibling brief states that no scaling curve of task performance against channel count has been published; one has now been published, it postdates that brief, and it cuts the same way. Established Second, the DOI for the widely cited 2021 handwriting result returns HTTP 404 in two independent sessions; the result — about 86 characters a minute, one participant — is real, and this brief does not print that identifier. Frontier Third: bibliographic infrastructure in this cluster is actively unreliable. Established A major open metadata service returned the Flesher 2021 volume as 572 when it is 372, and attached one paper’s abstract to an unrelated record; a long query to a second service returned veterinary and cryobiology records; and a request for a specific DOI at a third publisher returned a different article entirely. Speculative That last failure is why this brief makes no claim about the retraction status of the 2017 completely-locked-in paper, and why a citation taken from a single automated lookup here is not a citation.