Human-Machine Symbiosis

Established The founding argument of this subject is that the bottleneck between a person and a machine is not channel capacity but the division of labour. J.C.R. Licklider put it in print in 1960: the human should set goals, formulate hypotheses and choose criteria; the machine should do the routinisable work that stands between a decision and its execution. On that reading a speech interface that returns 62 words per minute to a person with ALS is a total success, and a hypothetical gigabit cortical link that does not change what the pair can decide together is a failure. Speculative Sixty-six years later the field spends most of its capital on the second thing.

Established The measured record is real, clinically meaningful and uniformly smaller than its reputation. Intracortical implants restore communication to people who have lost speech, at rates between roughly 32 and 78 words per minute against 160 for natural conversation. Endovascular electrodes reach cortex through a vein with no serious adverse events in four patients at one year. Adaptive deep brain stimulation crossed from research device to approved product in 2025. Established And every headline result in the speech literature is a single participant, in a single laboratory, unreplicated.

Frontier The brief's organising claim is that the field's public metric — electrode count — is measuring the wrong thing. The reading problem has a well-posed information-theoretic framing and steady engineering progress. The writing problem does not: nobody can say what “full bandwidth into cortex” would even mean without a forward model of what cortex does with arbitrary multi-site stimulation, and no such model exists. Speculative Meanwhile the ethical problem that was supposed to be decades away has already been demonstrated in four people: inner speech is partially decodable from motor cortex, and the group that showed it published the countermeasure in the same paper.

Established The intracortical speech record for rate is 62 words per minute, in one person. Willett and colleagues implanted four microelectrode arrays of 64 electrodes each — two in area 6v, two in area 44 — in a single participant with ALS, and decoded attempted speech at 9.1% word error rate on a 50-word vocabulary and 23.8% on a 125,000-word vocabulary, the first large-vocabulary demonstration, at 62 words per minute. That is 3.4 times the previous record and roughly 39% of natural conversational speech at 160 words per minute. Established Inside the engineering result sits a neuroscientific one that matters more for the long run: a detailed articulatory representation of phonemes persists years after paralysis, and the tuning to speech articulators is spatially intermixed closely enough that accurate decoding is possible from a small region of cortex. Frontier That is what makes late intervention plausible — the signal does not decay away while the patient waits.

Established The expressivity record belongs to a different technology, a different laboratory and a different participant. Metzger and colleagues used a high-density ECoG array — surface electrodes, not penetrating shanks — and reported a median 78 words per minute at a median 25% word error rate over a 1,024-word vocabulary, with three simultaneous outputs: text, synthesised audio in a voice reconstructed from the participant's pre-injury recordings, and a facial avatar. The decoders reached that performance after less than two weeks of training. Established One participant, one centre, unreplicated at that level. Frontier The three-modality output is the underrated part: it is the first result in this literature that treats communication as something other than a text channel, and expression is not a bit rate.

Established The accuracy record is a third paper, and reading it carefully dismantles the idea that there is a single speech-BCI record at all. Card and colleagues implanted four arrays — 256 intracortical electrodes — in the left ventral precentral gyrus of one 45-year-old man with ALS, five years post-onset, under BrainGate2. On the first day of use, 25 days after surgery and after 30 minutes of cortical recordings for calibration, the system reached 99.6% accuracy on a 50-word vocabulary. On the second day, after 1.4 additional hours of training, it reached 90.2% accuracy on a 125,000-word vocabulary. With further training data it sustained 97.5% accuracy over 8.4 months, used for self-paced conversation at approximately 32 words per minute for more than 248 cumulative hours. Established The accuracy record therefore runs at roughly half Willett's word rate and 41% of Metzger's. Frontier The variable the paper is titled for — calibration time — is the one that decides whether a device is usable at home, and it is what the field is quietly converging on while coverage still counts electrodes.

Established The intermediate benchmarks show the slope, and one of them is misquoted almost everywhere. Before speech there was handwriting: in 2021 the same Stanford-Brown lineage decoded imagined handwriting from a participant with tetraplegia at about 86 characters per minute and 18 words per minute. The figure in wide circulation is “about 90 characters per minute,” which is wrong, and the DOI most often printed alongside it returns HTTP 404. Frontier A vocal-tract decoding result of roughly 15 words per minute is also commonly cited from 2021; this brief does not print that number, because it could not be confirmed against a reachable source. Established The trajectory that survives verification is nevertheless steep: 18 words per minute in 2021, 62 and 78 in 2023, and large-vocabulary error rates falling from about one word in four to about one in ten within a year. In four different people.

Established The hardware numbers are smaller and more specific than the marketing. BrainGate's clinical implant is a 96-channel Utah-type array — the silicon grid has 100 electrode sites, the clinical NeuroPort system wires 96 of them, and the technical reader who checks will find the second number. The participant history is short and traceable: four patients with tetraplegia in the Cyberkinetics trial of 2004–2006, BrainGate2 opening in July 2009, two brainstem-stroke participants controlling a robotic arm in May 2012, and two participants using a fully wireless transmitter in 2021 with typing accuracy and speed reported as identical to wired. Frontier The cumulative participant count over two decades could not be verified for this brief and is not stated here. Established Synchron's Stentrode takes the opposite trade: 16 electrodes on a self-expanding nitinol stent, about 5 cm long and at most 8 mm across, delivered endovascularly into the superior sagittal sinus with no craniotomy. The SWITCH study reported no serious adverse events in the first year in all four implanted patients, who used the device at home for over a year. Established Neuralink's human N1 carries 1,024 active electrodes distributed over 64 flexible leads; the widely quoted 3,072 belongs to the 2019 animal research platform, which is 96 threads of 32 electrodes. Frontier So the live design question is now quantified: a 64-fold channel deficit against a one-year home-use safety record that the high-channel route does not have.

Established The most instructive fact in the invasive record is a failure. Neuralink's first human participant, implanted 28 January 2024, suffered retraction of a reported 85% of implant threads — brain movement was approximately three times what the company had expected — and software changes restored much of the lost performance. Frontier That is the field's first public case of a high-channel implant degrading in vivo and being rescued by decoder adaptation, and it is the strongest available evidence that channel count is not the binding variable it is sold as. Frontier Company-reported figures put 12 participants implanted and more than 15,000 hours of cumulative use by September 2025; a 2026 clinical review separately reports 21 patients enrolled in the PRIME study, which is a different quantity and should not be merged with the implant count. Established Both are vendor-traceable numbers and are labelled as such here.

Established The non-invasive ceiling is not a matter of better electrodes. The best reported information transfer rate for a non-invasive interface in the clinical literature is 360 bits per minute, and pooled P300 speller accuracy sits around 74%. Non-invasive EEG has been controlling physical robots since 1988, and Vidal coined the term brain-computer interface in 1973; five decades have not moved the bandwidth by orders of magnitude, because the limit is the head. Established Marblestone and eighteen co-authors derived the constraints in 2013 and nobody has overturned them: single-photon optical imaging is scattering-limited to 1–2 mm of depth; magnetic resonance has a temporal floor near 100 ms and a spatial floor near 40 micrometres set by water diffusion; and any recorder is thermally capped. Frontier Those are properties of a skull and of water, not of an instrument.

Established The motor and rehabilitation half of the field is where the mature comparators live, and they are chastening. Deep brain stimulation is the reference point: more than 244,000 patients implanted worldwide through 2019, with FDA approvals running from essential and Parkinsonian tremor in 1997 through Parkinson's disease in 2002, dystonia in 2003, OCD in 2009 and epilepsy in 2018. Established Its closed-loop successor is now a product: Medtronic's BrainSense adaptive DBS was FDA-approved in 2025 under PMA P960009 supplement S478, alongside a CE mark, and Newronika's AlphaDBS was CE-marked in 2021. Frontier Its randomised blinded evidence is four male participants in a crossover trial whose own authors describe it as establishing feasibility. Approval and proof are different things and here the gap is wide. Frontier A clinical review scores it at a 16.3% decrease in bothersome symptoms on Level 3 evidence; ADAPT-PD reports approximately 98% of participants opting to continue with adaptive rather than conventional stimulation, which is a real signal and is preference, not efficacy; and a sympathetic real-world series reports 8 patients, 6 still on adaptive stimulation at four months, with 3 of 8 improving individually.

Established And sixty years after the first powered exoskeleton, the entire randomised evidence base for robot-assisted gait training after spinal cord injury is eight trials and 241 people. GE's Hardiman in the 1960s weighed 680 kg, promised a strength amplification factor of 25, managed 0.76 m/s, and was never made to work — its control loops produced violent and uncontrollable motion. ReWalk became the first FDA-approved medical exoskeleton in 2014. Established A 2026 systematic review searched to June 2025 and pooled everything randomised: 6-minute walk distance SMD 0.57 (95% CI 0.12–1.03), Walking Index for Spinal Cord Injury II SMD 0.49 (0.13–0.84), functional independence SMD 0.39 (0.05–0.72), and lower-extremity motor score SMD 0.03 (−0.27 to 0.34) — no effect on strength at all. GRADE certainty ranged from low to moderate. Frontier The pool mixes treadmill systems with wearable overground devices, so it is not a clean exoskeleton result, and the authors flag both the small number of trials and the scarcity of long-term follow-up. Established The scale is the correction. A technology with sixty years of history, several cleared devices and an enormous popular literature has a total randomised evidence base smaller than one mid-sized drug trial, and the one thing it demonstrably does not do is make legs stronger.

Frontier On sensory substitution, closed-loop prosthetics and proprioceptive feedback this brief declares a gap rather than filling it. The write side of the loop — intracortical microstimulation producing tactile or proprioceptive percepts, sensorised prosthetic hands closing a control loop through the user's own somatosensory cortex — is an active and serious literature, and no primary source in it was reachable for this brief. Established What can be said from verified sources is the shape of the historical record: Dobelle implanted visual-cortex arrays producing phosphenes from 1978; Warwick, Gasson and Kyberd implanted an array into peripheral nerves for bidirectional signalling in 2002; and Neuralink's Blindsight cortical visual prosthesis received FDA breakthrough device designation in September 2024, which is a regulatory scheduling decision and not evidence of efficacy. Speculative The honest position is that the read side of this field has numbers and the write side mostly has demonstrations, and that asymmetry is the subject of the next section.

Frontier The question the field is actually converging on is calibration, not accuracy. Card's headline is not 97.5%; it is that 30 minutes of cortical recording produced a working 50-word system on day one, and 1.4 further hours produced a 125,000-word system on day two. A device that needs a technician and a morning every week is a research instrument. A device that calibrates in half an hour and holds for 8.4 months is a product. Frontier Nobody has published how often recalibration is needed across a cohort, because there is no cohort.

Frontier Is high-bandwidth non-invasive interfacing physically bounded? The strongest available answer is yes, and it is arithmetic rather than opinion. Every non-invasive modality is limited by the skull (electrical smearing, acoustic impedance), by scattering (optical), or by diffusion (magnetic resonance). Speculative For the bound to break, one of two things would have to be true: either a modality that is not electrical, optical, acoustic or magnetic — nobody has one — or an in-body relay, molecular or ultrasonic “neural dust” reporting out through a transcranial link, which converts the problem from non-invasive to minimally invasive and therefore concedes the point. Handwave Marblestone's own estimate of the electrical channel requirement is instructive about how far off the arithmetic is: with perfect spike sorting, roughly 13,000 to 210,000 electrodes; with current spike-sorting algorithms, roughly 7.5 million.

Frontier Is the bandwidth race a category error? Licklider's position says yes: the pair's performance is set by the division of labour, not the pipe. Speculative The steelman for racing anyway is a phase-transition argument — if the interface becomes fast enough that the user stops noticing it, qualitatively new joint behaviour becomes possible, in the way interactive computing was not merely faster batch computing. Nobody can name the threshold in advance, and betting on unnamed thresholds is how the GUI and the web happened. Speculative The counter is that the human side may have no matching sink: conscious information throughput is often estimated at tens of bits per second, which would make a gigabit cortical link a firehose aimed at a teaspoon. Speculative The counter-counter is that the sink need not be conscious — motor learning absorbs enormous bandwidth without any of it reaching report.

Frontier Minority position worth taking seriously: channel count is not binding; chronic stability is. One patient lost 85% of threads and recovered most performance in software. Utah arrays degrade over years through gliosis and insulation failure. If a decoder can absorb an 85% channel loss, then adding channels buys headroom against degradation rather than capability. Frontier What would have to be true for channel count to matter directly is a task whose decoding accuracy is still climbing as a function of channels at today's counts — and the observation that accurate speech decoding was possible from only a small region of cortex cuts against it. Speculative No such scaling curve has been published by anyone, in either direction.

Frontier Endovascular and peripheral routes may be the ones that scale to users. The Stentrode reaches cortex through a vein with a hundredfold reduction in surgical risk and a one-year home-use record; the trade is 16 electrodes. Synchron's US COMMAND trial is confirmed to be underway in patients with quadriplegia, and no enrolment figures, endpoints or results were reachable for this brief. Speculative The hypothesis worth stating plainly is that the first brain interface to reach a million users will not be cortical. Speculative What would have to be true is that the useful applications live below about 100 channels — and cursor control, communication and switch access all do.

Frontier Inner speech is decodable, and that finding is one year old. Using multi-unit recordings from four participants — unusually many for this literature — Kunz and colleagues found that inner speech is robustly represented in motor cortex and that imagined sentences can be decoded in real time. The representation correlates highly with attempted speech, but a separable neural “motor-intent” dimension distinguishes the two, and some aspects of free-form inner speech could be decoded during sequence-recall and counting tasks. Frontier The same paper demonstrates high-fidelity strategies that prevent speech interfaces from unintentionally decoding private inner speech. Speculative A mental password that gates decoding is not an ethics slogan; it is an engineering primitive, and the group that found the leak shipped the gate in the same publication. That is a governance pattern the Institute should name and want repeated.

Speculative And the forbidden question: is the read/write bandwidth asymmetry fundamental? Reading from cortex is a measurement problem. Writing to cortex is a control problem in a recurrent system whose dynamics are not modelled, where every additional stimulation channel enlarges the space of unintended states. Handwave There is no established theory of how many independent percepts can be written into cortex without interference, and there is therefore no defensible unit in which to state a target. This is the strongest argument that “full-bandwidth bidirectional neural interface” is not merely hard but under-specified.

Frontier The workback plan to a full-bandwidth bidirectional interface has nine links, and the binding two are scientific rather than engineering. In order: (1) chronic channel survival — under 10% channel loss at five years across more than twenty human implants, against a public record containing an 85% single-patient loss inside a year; (2) decoder invariance to channel loss, flat performance under ablation to 50%, which is partly demonstrated already; (3) a published scaling law relating task accuracy to channel count over two orders of magnitude on a fixed task; (4) a thermal and data budget that meets the physical constraints at scale; (5) write-side psychophysics — how many independent, discriminable percepts multi-site intracortical microstimulation can evoke without interference; (6) a forward model that predicts cortical response to a novel stimulation pattern; (7) closed-loop co-adaptation stable over months without oscillation; (8) surgical throughput and a cost structure compatible with reimbursement; (9) a regulatory pathway for a non-therapeutic indication.

Speculative Links (5) and (6) jointly are the binding constraint, and they are the least funded. Everything on the read side has a well-posed information-theoretic framing and a decade of steady progress. The write side has no framing at all: without a forward model there is no way to specify what a full-bandwidth input is, which means the programme's headline objective cannot currently be written as a measurable target. Handwave A field that cannot state its goal in units is not close to achieving it, whatever the electrode count does.

Frontier Link (3) is the cheapest and most consequential unpublished result in the subject. If task accuracy saturates below 1,000 channels, the entire high-bandwidth programme is answering a question nobody needed answered, and the correct investment moves to stability, calibration and surgical risk. Nobody has published the curve. Speculative The uncomfortable possibility is that some laboratories have it internally and it saturates.

Established Link (1) is genuinely tractable and is not binding. Hermetic packaging, flexible substrates and gliosis mitigation are materials and manufacturing problems with an industrial path; the wireless intracortical result of 2021, which removed the percutaneous pedestal and with it the main chronic infection route, is the shape of progress here. Frontier Link (9) is the secondary binding constraint and it is institutional, not scientific: every approved neural implant in every major jurisdiction is a therapy for a diagnosed disease, and an augmentation device has no door.

Frontier This brief waits on a specific result that another field produces. Closed-loop co-adaptation — link (7) of the workback plan — is not a decoder problem. It is a plasticity problem: the user's cortex is the second learner in the loop, and whether joint learning converges or oscillates depends on how the nervous system's own adaptation rate can be measured and steered. That belongs to Neuroplasticity Engineering, and the single most striking observation in that subject — a participant whose walking improved and stayed improved after the stimulator was switched off — is the clearest sign that an interface can leave a nervous system better than it found it, which is a different object from a prosthesis.

Established Beyond that: Bioelectric Medicine supplies the closed-loop controllers, the biomarker-triggered stimulation logic and the regulatory precedents that adaptive DBS established and that any augmentation device would inherit. Materials science owns chronic stability — hermetic feedthroughs, flexible substrates that survive brain micromotion of the magnitude that detached 85% of one patient's threads, and coatings that slow gliosis. Frontier Machine learning owns decoder invariance and rapid calibration, and it is the one input that has been reliably delivering. Speculative Systems neuroscience owns the missing forward model of stimulation, and nobody is commissioning it.

Frontier There is also a dependency the field does not name: a shared task. Biological Computing has the same defect on its wetware side, and both fields would be transformed by a benchmark that let two laboratories disagree about the same measurement rather than about their own priors. Speculative And one dependency runs the other way, which is unusual enough to state: the rehabilitation half of this subject depends on a health-economics result rather than a scientific one, because an exoskeleton competes with a wheelchair that is faster, cheaper and more reliable, and only a redefinition of the clinical endpoint changes that comparison.

Frontier Experiment one, and the cheapest: publish the channel-count scaling curve. Take one fixed task on one participant with a high-channel implant, subsample channels from the full array down two orders of magnitude, and report task accuracy as a function of channels retained. This requires no new surgery, no new hardware and no new consent beyond re-analysis. Frontier It is the experiment that decides whether the field's headline metric means anything, and its absence after a decade of channel-count marketing is itself a finding.

Speculative Experiment two: write-side psychophysics. Measure discriminability — d-prime — as a function of the number of simultaneously stimulated intracortical sites, in people already implanted for a therapeutic indication. The output is a curve saying how many independent percepts can be written before interference dominates, which is the first number in the whole bidirectional programme that would have units. Handwave Nobody is funding it, and until somebody does, every roadmap past the read side is a sketch.

Speculative Experiment three: a head-to-head, task-matched, pre-registered comparison of invasive against non-invasive interfaces on one fixed clinical endpoint. Every paper in this field compares against its own prior; there is no common benchmark, no shared test set and no agreed endpoint. The obstacle is institutional rather than technical — the field would have to accept that some of its demonstrations would lose — and that is exactly why an institute rather than a laboratory should run it.

Frontier Experiment four: replication. Run one of the three speech results in a second laboratory, in a second participant, with the original decoder and no retuning. Established Every headline speech figure quoted anywhere in this brief is n=1, and a second participant is worth more to this field than a thousand more electrodes. Speculative The reason it has not happened is not cost but incentive: a successful replication publishes as a footnote and a failed one publishes as an attack.

Frontier Experiment five: publish per-channel signal-to-noise against time, per subject, for every chronic implant in service. That single reporting convention would convert chronic stability from an anecdote into a curve, and it costs nothing that is not already being recorded. Speculative Experiment six: measure the exoskeleton endpoint nobody has powered a trial for. Bone mineral density, bowel transit and cardiovascular load under upright weight-bearing, against a wheelchair control, over two years. If upright loading has no wheelchair substitute, that trial is the one that decides whether wearable exoskeletons have a clinical future, and the ambulation literature has been measuring the wrong outcome for a decade.

Established The physical budget is quantified and it is tight. Steady-state heat dissipation must stay under roughly 40 milliwatts in a 500 mg mouse brain to hold a 2-degree rise, and that constraint scales unfavourably: the power budget grows with brain mass while the channel count anyone wants grows with neuron count. The recorder should displace no more than about 1% of brain volume. And the data has to leave the skull — recording 75 million neurons at 1 kHz with minimal compression is on the order of 100 gigabits per second, which is a transcutaneous link problem before it is a neuroscience problem. Frontier The useful engineering figures of merit are therefore watts per thousand channels and bits per second per milliwatt off-head, and almost nobody reports either.

Established On the insertion side the published animal figures are specific. Neuralink's platform paper reports threads 4–6 micrometres thick and 5–50 micrometres wide, about 20 mm long, at 50 or 75 micrometre electrode spacing; a robot inserting up to 6 threads (192 electrodes) per minute in automatic mode with an achieved average of 29.6 electrodes per minute; an insertion success rate of about 87% across 19 surgeries; digitisation at 19.3 kHz and 10 bits; and a spiking yield of roughly 46% of channels at a 0.35 Hz detection threshold. Frontier The last number is the one to hold onto: fewer than half the electrodes in the flagship animal platform were returning spikes.

Frontier The requirement nobody has specified is the stimulation waveform library. A read-only interface needs amplifiers and a decoder. A bidirectional one needs a catalogue of stimulation patterns with known, discriminable, reproducible percepts attached — and the catalogue does not exist because the psychophysics has not been done. Handwave Until it does, “write bandwidth” specifications should be read as marketing arithmetic on the number of wires.

Established The mature neighbour is deep brain stimulation, which supplies both the surgical infrastructure and the honest risk baseline for anything implanted: standard leads carry four platinum-iridium contacts, segmented directional leads arrived in 2015, and the accumulated implant base past 244,000 people is the only chronic-neural-implant dataset of real size that exists. Frontier Responsive neurostimulation for epilepsy is the closed-loop precedent, with reported seizure reductions in the range of 38–67%, and it demonstrates that a device can detect a neural state and act on it without a person in the loop.

Established Bioelectric Medicine is the adjacent field that has just done the thing this one has not: a double-blind, sham-controlled, 242-patient randomised trial of an implanted vagus-nerve device for rheumatoid arthritis, published with a controlled primary endpoint of 35.2% against 24.2% on sham. Frontier An eleven-point difference is a modest result and a real one, and it is the only trial in this cluster of subjects large enough to be believed on its own. It is the standard the interface field should be measured against, not the one it is measured against.

Established The genuinely successful neural interface, and the one this literature rarely cites, is the cochlear implant — a low-channel, restoration-oriented, reimbursed device with hundreds of thousands of users, which won by choosing an application that lives comfortably below the bandwidth ceiling. Frontier Also adjacent: rehabilitation robotics and Robotics in Infrastructure for the actuation and control problems; Biological Computing for cultured-neuron interfacing, which is this subject with the human removed; Human Cognitive Augmentation and Distributed Cognition for the Licklider question of what a human-machine pair can do that neither can; and Cognitive Liberty, which owns the legal half of the inner-speech problem. Speculative The adjacency that has produced least and promises most is materials science: brain micromotion three times larger than a well-funded company expected is a mechanical mismatch problem, and it is the sort of problem that yields to substrates rather than to neuroscience.

Established There is no regulatory door for a non-therapeutic neural implant. Every approved neural device in every major jurisdiction is a therapy for a diagnosed condition; the approval routes are built on risk-benefit judgements that assume a disease on the benefit side. An augmentation device in a healthy adult has no predicate, no cleared comparator and no evidentiary standard. Speculative That is not a gap that closes by itself, and it is a stronger brake on the exotic end of this subject than any physical constraint in this brief.

Established The pathway that does work is worth studying because it is the field's one clean example. Adaptive DBS reached market as a supplement to an existing premarket approval — PMA P960009 supplement S478 — riding an approved therapy's regulatory history rather than opening a new category. Frontier The lesson generalises uncomfortably: in neural interfaces the route to market runs through an already-approved indication, which biases the whole field towards restoration and away from anything genuinely new.

Frontier The field has no shared benchmark and no obligation to report negatives. Trial registration is not a substitute: Synchron's US COMMAND trial is confirmed to exist, and no enrolment, endpoint or result was reachable through any open route for this brief. Established Company-reported usage statistics — cumulative implant hours, participant counts — enter the literature through press and stay there uncontested.

Established And the citation chain itself is an institutional failure worth naming. A 2026 peer-reviewed clinical review of brain-computer interfaces — a good review, cited repeatedly in this brief — reports the 78 words-per-minute avatar result and attributes it to “Willett et al. (2021).” That result is Metzger and colleagues, 2023; Willett's 2021 paper is the handwriting study and his 2023 paper is the 62-wpm one. Frontier The misattribution passed peer review in a clinical journal. Any institution building a research programme on secondary literature in this field should treat “it was in a peer-reviewed review” as an unverified claim, and should budget for resolving DOIs.

Frontier Mental privacy stopped being a future problem in 2025. Inner speech is robustly represented in motor cortex, imagined sentences can be decoded in real time, and some aspects of free-form inner speech were decodable during structured tasks in four participants. An always-on device that decodes attempted speech is an always-on device sitting on top of a partially decodable inner monologue. Established The mitigation exists and was published alongside the risk: gating strategies that stop a speech interface from decoding unintended inner speech. Speculative The governance question is therefore narrow and answerable — whether such a gate is mandatory, auditable and user-controlled, or a vendor option.

Established The consent structure of this field is unusual and under-discussed. Every headline result is one person, identifiable within a small community, in a multi-year study where the participant is a collaborator rather than a subject, often with a terminal diagnosis. That produces excellent science and a consent relationship that standard trial ethics was not designed for. Frontier It also produces a specific harm that is now documented in adjacent implant fields: what happens to a person's device when the sponsor stops supporting it. Six elective explantations before twelve months in the one large sham-controlled implant trial in this cluster, three of them for perceived lack of benefit, is the only quantified glimpse the pack offers of the exit side.

Established The historical warning is Dobelle. Cortical visual prostheses were implanted from 1978 and run commercially into the 2000s outside the mainstream trial system, and the programme was rejected for inadequate ethical oversight, non-peer-reviewed outcome claims and patient harm. Speculative The technical premise — phosphene vision from cortical stimulation — was not wrong, and is exactly what has now received breakthrough designation under proper protocols. The steelman, which the Institute should sit with, is that the programme was rejected for its governance and the field then lost roughly twenty-five years of clinical learning by treating the governance verdict as a physics verdict.

Speculative If Licklider was right, the civilizational question is not how much bandwidth a person can accept but what a person-machine pair can decide that neither can decide alone. That reframing changes what counts as progress. On the bandwidth metric, this field has moved from 18 to 78 words per minute in four years and has a long way to go. On the division-of-labour metric, the most consequential human-machine symbiosis of the last decade happened entirely through a keyboard and a screen, at a few tens of bits per second, and nobody needed a craniotomy.

Frontier The near-term civilizational effect is therefore restorative and real: communication returned to people with ALS and brainstem stroke, upright loading returned to people with spinal cord injury with measurable gains in walking and independence and none in strength. That is a large good and it is bounded by cost and reimbursement rather than by physics.

Speculative The far-term effect divides on a single question: whether augmentation ever escapes the therapeutic frame. If it does not — if every neural implant remains a treatment for a diagnosis — then this subject stays a branch of rehabilitation medicine indefinitely, which is neither a failure nor the thing its advocates are selling. Handwave If it does, the distributional problem arrives immediately and in an unusually sharp form, because a surgical, individually-fitted, individually-calibrated device has no marginal-cost collapse of the kind that made mobile telephony universal. Speculative A cognitive advantage that requires neurosurgery and a maintenance contract is a different kind of inequality from one that requires a subscription. Handwave And it is an inequality with an unusual failure mode, because a device that stops being supported does not merely stop working; it has to be removed from a skull.

Speculative The Institute's own reason for holding this subject open is the Licklider one. If the pair is the unit of analysis, then the interesting long-run question is not how much data crosses the interface but whether joint human-machine reasoning is qualitatively different from either half, and that question is answerable today with keyboards, at no surgical risk, by anyone willing to design the experiment.

These horizons track what the measured record supports, not what is announced. Every dated claim below is anchored to a result in section 2, and the first horizon is deliberately unambitious because the field's binding problems are not on the read side.

• 10 yr: Frontier Speech decoding replicated across multiple participants and centres, with calibration under an hour and the first cohort-level data on how often recalibration is actually needed at home. Frontier Endovascular interfaces reporting multi-year safety in tens rather than fours of patients, which is the threshold at which the low-channel route stops being a curiosity. Frontier A published channel-count scaling curve settling whether the electrode race measures anything. Frontier Robot-assisted gait training either acquires a larger randomised base than 241 people or is formally redefined around bone, bowel and cardiovascular endpoints rather than ambulation.
• 25 yr: Speculative Chronic stability solved as an industrial problem — hermetic packaging, flexible substrates, under 10% channel loss at five years across a real cohort — and closed-loop co-adaptation demonstrated stable over years rather than months. Speculative Write-side psychophysics produces the first catalogue of discriminable evoked percepts, which would be the first honest specification of write bandwidth in the field's history. Speculative A regulatory category for a non-therapeutic neural device exists in at least one jurisdiction, or the augmentation programme is formally stalled and says so.
• 50 yr: Speculative A forward model of cortical response to arbitrary multi-site stimulation, accurate enough to design a percept rather than discover one by search. Speculative That is the result on which everything past sensory substitution depends, and it is nobody's programme today. Handwave If it does not arrive, bidirectional interfacing plateaus permanently at coarse substitution, and the ceiling of the whole subject is set by a missing theory rather than by any device, budget or surgeon.
• 100 / 250+ yr: Handwave Full bidirectional bandwidth in the sense the phrase is usually meant — a person and a machine sharing representations rather than exchanging symbols — requires a unit of measurement that does not currently exist, and a theory of shared representation that neither neuroscience nor machine learning has offered. Handwave It is not obviously impossible; it is not currently a specification. Handwave Anyone offering a date for it is offering a date for a sentence, not for an experiment.

• Depends on This brief waits on one result another brief on this map produces. Neuroplasticity Engineering owns the second learner in the loop: every closed-loop interface asks a nervous system to adapt to a decoder while the decoder adapts to it, and whether that joint process converges, oscillates or degrades is a plasticity question with no interface-side answer. That brief also holds the only long-follow-up case in which an implant left the nervous system permanently better than it found it, which is the difference between a prosthesis and a therapy and is currently drawn from a single patient. Beneath that sit dependencies that are not briefs: materials science for chronic electrode survival, systems neuroscience for the missing forward model of stimulation, and machine learning for calibration time, which is the one input that has reliably delivered.
• Requires (not on this map) Three dependencies bind here that no brief on this map will produce. The first is institutional: every approved neural implant in every major jurisdiction is a therapy for a diagnosed condition, and an augmentation device in a healthy adult has no predicate, no cleared comparator and no evidentiary standard — which is a harder brake on the exotic end of this subject than any physical limit in section 4. The second is scientific and nobody is commissioning it: without a model that predicts what cortex does with a novel multi-site stimulation pattern, the write side of a bidirectional interface cannot be specified in units, let alone built, and the field's headline objective remains a sentence rather than a target. The third is a shared task: every paper in this literature compares against its own prior, so there is no test on which invasive and non-invasive approaches can be made to disagree about the same measurement, and constructing one requires the field to accept in advance that some of its demonstrations will lose.
• Enables What this brief would unblock is not a device but a measurement. A published relationship between channel count and task performance would settle, for every field that borrows from this one, whether neural bandwidth is a resource worth buying. A catalogue of discriminable evoked percepts would give sensory substitution, cortical vision and closed-loop prosthetics their first shared unit. And a demonstrated, auditable gate against unintended inner-speech decoding would become the reference implementation for Cognitive Liberty, which currently has a legal framework and no engineering primitive to attach it to. Human Cognitive Augmentation and Distributed Cognition both inherit whatever this subject concludes about whether the pair or the channel is the right unit of analysis.
• Adjacent Alongside sit the fields solving adjacent halves of the same problem. Bioelectric Medicine supplies the closed-loop controllers and the regulatory precedent, and is the only neighbouring field to have run a sham-controlled trial large enough to believe unaided. Biological Computing is this subject with the human taken out, and shares its missing benchmark exactly. Robotics in Infrastructure holds the actuation and control-bandwidth problems that defeated Hardiman and still bound wearable exoskeletons. And the cochlear implant, which belongs to no brief here, is this field's most successful device and its most ignored lesson.

Established “Neuralink patients have 3,072 electrodes.” The human N1 carries 1,024 active electrodes on 64 flexible leads. The 3,072 figure is the 2019 animal research platform — 96 threads of 32 electrodes each — and it is a rig, not an implant. Frontier A related claim circulated in an earlier version of this Institute's own research pack — that the most-read encyclopaedia article prints 3,072 beside the human-trial narrative — was checked and withdrawn: that article attributes the figure explicitly to the research platform and gives no N1 count at all. The conflation is real in general coverage and should not be pinned on that page. It is recorded here because a brief that corrects other people's errors owes the same treatment to its own.

Established “BrainGate uses 100 electrodes.” The Utah array is a ten-by-ten grid of 100 silicon shanks; the clinical NeuroPort system wires 96 of them. One hundred is the physically present count, 96 is the recorded count, and the difference is exactly the kind of thing a technical reader checks. Write it as a 96-channel Utah-type array with 100 electrode sites.

Established “The handwriting BCI ran at about 90 characters per minute.” The figure as reported is 86 characters per minute and 18 words per minute. Worse, the DOI that circulates alongside it returns HTTP 404 at the publisher, so a reader who tries to check the number cannot reach the paper. Frontier This brief prints 86 and does not print the DOI.

Established “The speech BCI record is 97.5% accuracy on a 125,000-word vocabulary.” That pairing does not appear in the source. The abstract states 99.6% at 50 words on day one, 90.2% at 125,000 words on day two, and 97.5% sustained over 8.4 months with no vocabulary size attached to the sustained figure. The defensible large-vocabulary pairing is 90.2% at 125,000 words. Established And the framing is wrong anyway, because that system ran at approximately 32 words per minute — roughly half the 62-wpm paper's rate. There is an accuracy record and a rate record, they belong to different papers with different participants, and any sentence containing “the speech BCI record” has chosen a metric without saying so.

Established “These results are validated across many patients.” Willett 2023: one participant. Metzger 2023: one participant. Card 2024: one participant. Adaptive DBS's randomised blinded evidence: four. The inner-speech study, at four participants, is unusually large for this literature. Frontier The improvement from one word in four wrong to one in ten happened across different people in different laboratories in the space of a year, which is encouraging and is not replication.

Established “More channels means more bandwidth means more capability.” The public record contains a patient who lost 85% of implant threads and recovered most performance through software, and a paper noting accurate speech decoding from only a small region of cortex. Frontier No published scaling curve relates channel count to task performance in either direction, which makes the channel race a marketing metric with no measured relationship to what it claims to predict.

Established “Non-invasive BCI just needs better electrodes or better AI.” The limits are the skull, optical scattering and water diffusion: 1–2 mm of optical depth, a 100 ms magnetic-resonance temporal floor, a 40-micrometre spatial floor, a thermal cap near 40 milliwatts in a 500 mg brain. Frontier The best reported non-invasive information transfer rate in the clinical literature is 360 bits per minute, after five decades. Better algorithms do not move a diffusion constant.

Frontier “Adaptive DBS is proven better than conventional DBS.” It is approved, which is different. The randomised blinded evidence is a four-person crossover feasibility trial whose own authors call it that; a clinical review scores it at a 16.3% reduction in bothersome symptoms on Level 3 evidence; a real-world series found 3 of 8 individually improved. Established The frequently quoted “40% reduction in motor symptoms” traces back to the four-person pilot and is not an established effect size. Frontier And the 98% of ADAPT-PD participants who chose to continue measures preference, not efficacy.

Established “DBS is a benign procedure.” Aggregated reported complication rates include intracranial haemorrhage around 5%, infection around 8%, lead migration around 10%, and neuropsychiatric effects in subthalamic cohorts reported as high as 70%. Frontier Those come from a tertiary aggregation, several are implausibly wide, and they are given here as the range a reader will encounter rather than as measured values. The point survives wherever in the range the truth sits: this is brain surgery, and any argument for implanted augmentation has to carry it.

Established “Exoskeletons will replace wheelchairs.” Hardiman managed 0.76 m/s at 680 kg in the 1960s and never worked. Sixty years later, the pooled randomised evidence for robot-assisted gait training after spinal cord injury is 8 trials and 241 people, with standardised mean differences of 0.39 to 0.57 on function, 0.03 on strength, and GRADE certainty from low to moderate. Frontier Current devices are slower than a wheelchair, need flat ground and often an attendant, and are reimbursed poorly. The defensible clinical case is physiological rather than ambulatory, and nobody has powered a trial to measure it.

Frontier “BCIs read your thoughts.” This used to have a clean answer and no longer does. Deployed systems decode attempted movement or attempted speech — motor intent, from motor cortex. But inner speech is robustly represented in the same cortex, imagined sentences can be decoded in real time, and some aspects of free-form inner speech were decodable during structured tasks in four participants. Established The correct statement is no longer “they cannot”; it is “they can, partially, in a research setting, and the field published a gate before anyone shipped the leak.” Anyone still using the old reassurance is a year out of date in the direction that matters.

Established Sceptic-side: “BCIs are a rich person's toy with no clinical value.” A person with ALS communicating at 62 words per minute after losing intelligible speech is a clinical outcome by any standard, and four patients have used an endovascular implant at home for over a year with no serious adverse events. Frontier The correct scepticism is not about clinical value but about scale, replication, and the gap between an approved device and a proven one.

Frontier And a misconception about the literature itself. A 2026 peer-reviewed clinical review attributes the 78-words-per-minute avatar result to “Willett et al. (2021)” — a paper that is the handwriting study, by an author whose relevant speech paper is 2023, describing work by a different group. Established The number is right, the citation is wrong, and it passed peer review. Handwave The lesson the Institute should take furthest: in a field where every headline is one participant and every figure travels through three layers of coverage, provenance is worth checking as carefully as the number.