Neural Interfaces

This brief is about the object, not the application. Brain-Computer Interfaces carries what these systems are used for and who uses them; Human-Machine Symbiosis carries the devices by name and channel count. What neither carries, and what this brief owns, is the engineering: what an electrode is made of, what happens to it in tissue, how many of its sites still record a spike after three years, why the ones that fail fail, and what a genuinely decade-stable thousand-channel implant would actually require.

The organising fact is how small the evidence base is. From the first human intracortical implant in the 1990s to the present, the entire published human chronic-yield literature amounts to one five-year single-subject retrospective, one explant study of six arrays from two people, one 1,500-day stimulation record in one person, one mixed-species longevity survey with two humans in it, and — since July 2025 — one twenty-year, twenty-array, fourteen-participant dataset that is still a preprint. Everything else is non-human primate work or vendor claim. A field that has been arguing about channel counts for twenty years has published five human durability studies.

Two findings reorganise the subject. First, the scaling curve the field said did not exist now does, and it is logarithmic: decoding signal-to-noise grows with the logarithm of electrode count, which is the mathematical form of diminishing returns. Second, human implants massively outlast the animal literature everyone quotes — a median recording life of 182 days in macaques against a mean implant duration of 1,063 days in humans — and nobody has explained the gap. If the explanation is the percutaneous connector, then a decade of electrode materials science was aimed at the second-largest problem in the device.

Established The founding chronic-failure dataset is Barrese and colleagues, 2013, and it is an animal study whose numbers have been quoted about humans ever since. Seventy-eight silicon microelectrode arrays implanted in 27 rhesus macaques from 1996 onward, analysed in Journal of Neural Engineering 10(6):066014. Established Recording duration ranged from 0 to 2,104 days — 5.75 years — with a mean of 387 days and a median of 182 days. Sixty-two arrays failed completely, at a mean of 332 days and a median of 133; nine experiments were electively terminated at a mean of 486 days; seven were still active at 753 days when the study closed. Established Fifty-six per cent of failures occurred within the first year. The single largest failure category was acute mechanical, at 48% of the total, and 83% of those were connector issues. Established Grossly observable biological failures accounted for 24%, of which the most prevalent single mode — 14.5% overall — was a progressive meningeal reaction that separated the array from the parenchyma. Frontier Where no acute interruption occurred, the slow decline in spike amplitude, noise and viable channel count extrapolated to signal loss at about eight years. Speculative That eight-year figure is the origin of most decade-scale-implant talk in the field, and it is an extrapolation from a decline curve rather than an observation of anything.

Frontier In July 2025 the human numbers arrived, and they are much better. Hahn, Stein, the BrainGate Consortium, Donoghue, Simeral, Hochberg and Willett assembled twenty years of BrainGate and BrainGate2 data — 2,319 recording sessions, 634 decoding sessions, 20 arrays, the first 14 participants, nineteen arrays in hand-knob motor cortex and one in middle frontal gyrus — and posted it as medRxiv 2025.07.02.25330310. Frontier Implant durations ran from 296 to 2,780 days, with a mean of 1,063 days, about 2.9 years, and a maximum follow-up of 7.6 years. Mean electrode yield — the fraction of sites recording neural spiking waveforms — was 35.6%. Frontier The decline across the whole enrollment period was 7%, from 41% yield in the first three months to 34% in the final three. More than half the arrays held at least 20% yield across the first three years. Frontier Eleven of fourteen arrays provided meaningful movement decoding throughout enrollment at a decoding SNR above 1; three peaked above 4.5 against an able-bodied computer-mouse benchmark of 6.29; and the average change in dSNR across recording duration was +0.34 units — performance drifted slightly up, not down. The failures cluster by array rather than by year: three arrays never consistently produced decoding signals, two showed large declines within the first year, four had consistently low yield throughout. Frontier This is a preprint from one consortium and has not been independently replicated, and that qualification should travel with every number in this paragraph.

Established The species gap is real, it is confirmed by two independent datasets, and nobody has explained it. Sponheim and colleagues, in Journal of Neural Engineering 18(6), reported over 6,000 recorded datasets spanning almost nine years across 55 Utah arrays in 17 rhesus macaques and 2 human subjects, with an average available recording lifespan of 622 days, several examples past 1,000 days and one reaching nine years. Established That paper states plainly that human implants lasted longer than non-human primate implants. Frontier Hahn 2025 says the same thing in its own words: longevity and reliability in the human cohort were better than in prior non-human primate studies. Speculative The candidate explanations are all untested — humans are less mechanically active around a head-mounted pedestal, human implants are handled more carefully and connected less often, human recording sessions are shorter and less frequent so connector wear is lower, human dura and pia are thicker and tougher, the surgical technique differs. The connector explanation fits Barrese’s own finding that 83% of acute mechanical failures were connector issues, and if it is right then the field’s central engineering effort has been misdirected for a decade. Established Sponheim also reported two findings the materials literature should have taken more seriously: shank length did not affect longevity or quality, and iridium oxide tip metallisation showed superior yield to platinum.

Established The materials post-mortem overturned the field’s own folk model, and the field has mostly not noticed. Barrese, Aceros and Donoghue, in Journal of Neural Engineering 13(2):026003, put eight arrays explanted from non-human primates at 37 to 1,051 days post-implant, plus three never-implanted controls, under a scanning electron microscope. Established They found progressive platinum electrode corrosion with underlying silicon degradation, parylene insulation cracking and delamination, silicone elastomer peeling from the array edges, and substantial and increasing tissue encapsulation. Histology showed gliosis and disrupted cortical architecture with minimal neuronal loss, and high Iba-1 microglial reactivity. Established At 1,051 days the electrode tracts were absent or barely visible — the array was no longer in the tissue. Frontier The mechanistic claim the authors draw is that gliosis occurs within months but does not prevent several years of useful recording, while progressive meningeal fibrosis encapsulates the array and lifts it off the cortex and ongoing foreign-body reaction degrades the materials, probably through oxygen radicals released by immune cells. Established For twenty years the standard story was that the glial scar insulates the electrode. Frontier The largest histological study of chronic explants says the scar is not the main problem; mechanical extrusion and oxidative degradation of the insulation are.

Established There is exactly one quantified human explant dataset in existence and it covers six arrays from two people. Woeppel and colleagues, Frontiers in Bioengineering and Biotechnology 9:759711, analysed two platinum arrays of 96 wired shanks each from participant P1’s motor cortex, implanted for about 980 days, and from participant P2 two platinum arrays of 88 wired shanks in somatosensory cortex plus two iridium-oxide arrays of 32 wired shanks in parietal cortex, implanted 182 days. Established Peak-to-peak voltage rose for the first 30 to 40 days and then declined; P1’s median fell 52% across 550 days and P2’s fell 14% across 90, at rates between days 30 and 120 of −4.0 and −4.86 µV per month. On the platinum arrays, degraded tips ran 27.8% in P1 against 11.1% in P2, degraded shanks 15.2% against 1.8%, tissue encapsulation 72.4% against 50%, all at p < 0.001. Established The two arrays inside P1 differed more than the two participants did: the medial array showed 16.3% tip degradation and 56.4% encapsulation, the lateral 41.6% and 89.0%. The encapsulation was highly cellularised and primarily collagen-I, and for one array the team demonstrated by matching vascular pattern against pre-implant imaging that part of what looked like scar was in fact pre-existing pial membrane. Established Impedance fell over time in all six arrays and stabilised after about two years. Frontier A study with n=2 that finds a 2.5-fold within-subject difference by placement is telling the field that chronic yield may be a property of where the array went, not of what it is made of.

Established Stimulation and recording are different longevity problems, and the evidence points in opposite directions. Hughes and colleagues, Journal of Neural Engineering 18(4), followed one participant with cervical spinal cord injury for 1,500 days, with sputtered iridium oxide film arrays in somatosensory cortex regularly stimulated and platinum-tipped arrays in motor cortex not. Established Recording quality declined on both array types by every measure — waveform count above 100 µV peak-to-peak, peak-to-peak voltage, noise, SNR — but SIROF electrodes were more likely than platinum ones to keep recording high-amplitude signals. Detection thresholds for stimulus-evoked sensations fell from a median of 31.5 µA at day 100 to 10.4 µA at day 1,500, most of the change between days 100 and 500; sensitivity to stimulation improved. Established Against that, Woeppel found that after seven test sessions of microstimulation over roughly a month — 200 µs cathodal, 400 µs anodic, 20–300 Hz, 1–100 µA — one iridium-oxide array showed 41.1% damaged tips on stimulated electrodes against 5.6% on unstimulated and 33.3% damaged shanks against zero, with a mean minimum interphase voltage of −1.7 V against −1.1 V on the undamaged array. Recording performance was unaffected by the loss of iridium. Frontier The honest statement is that the coating degrades while the function does not, over the durations so far measured, and that a brief asserting either that stimulation destroys electrodes or that it is harmless is wrong in both directions.

Established Impedance is not the health metric the field treats it as, and the sign convention is backwards. Barrese 2013 found systematic early impedance increases that did not appear to affect recording quality, followed by slow falls consistent with insulation failure. Established Woeppel found impedance falling in all six human arrays, with the lower-impedance arrays also showing lower signal amplitude and worse material degradation. Frontier A falling impedance is a sign of insulation breakdown, not of a healthier interface. Established The one supporting human single-subject record, Colachis and colleagues in Journal of Neural Engineering 18(4):0460d7, followed 448 sessions from 2014 to 2019 and found rapid deterioration in the first year followed by slower decline, with BCI performance still usable at five years. For scale: the Utah array in its clinical NeuroPort configuration has 100 electrode sites and 96 recorded channels; the endovascular Stentrode has 16; Neuralink’s N1 human implant has 1,024 electrodes on 64 flexible leads; its 2019 animal platform had 96 threads × 32 electrodes, or 3,072; Neuropixels 1.0 addresses 960 titanium nitride sites through 384 recording channels on a 10 mm shank of 70 × 20 µm cross-section; Neuropixels 2.0 exceeds 5,000 sites; the Argus II retinal prosthesis had 60 electrodes with 55 enabled. Frontier The spread from a working clinical device to a research probe is two orders of magnitude, and the clinical outcomes do not span two orders of magnitude.

Frontier The first published scaling relation between electrode count and decoding performance exists as of July 2025 and it is logarithmic. Hahn 2025 reports, in the authors’ own framing, that dSNR increases logarithmically with the number of electrodes, providing a pathway for scaling performance. Both halves of that sentence are true and they point in opposite directions: logarithmic growth is a pathway, and logarithmic growth is the mathematical form of diminishing returns. Frontier If the relation generalises, the marginal value of the ten-thousandth electrode is a small fixed increment over the value of the thousandth, and the channel race is a manufacturing contest rather than a capability contest. Speculative If it does not generalise — if some task family has a threshold rather than a slope — the high-channel programme keeps its case. Frontier The discriminating experiment is cheap and nobody has run it: take an existing high-channel human dataset, subsample electrodes, and publish accuracy against channel count on a fixed task with confidence intervals. Established It is worth recording that the sibling brief on Human-Machine Symbiosis states that no such curve has been published, and that this was correct when written.

Frontier Iridium oxide against platinum is now settled on the read side and open on the write side. Established Sponheim found iridium oxide tip metallisation gave superior yield; Hughes found SIROF electrodes more likely to keep recording high-amplitude signals over 1,500 days. Woeppel found stimulated iridium-oxide tips showing 41.1% damage after a month. Frontier The reconciliation the authors offer is that stimulation compromises the coating’s adhesion without detaching the metal until explantation, which is a mechanism, not a lifetime.

Frontier Ultraflexible probes have cleared the histology bar and have not cleared the years bar. Established Luan and colleagues, Science Advances 3(2):e1601966, showed nanoelectronic thread electrodes of subcellular dimensions with a cellular surgical footprint that reliably detected and tracked individual units for months, with impedance, noise, single-unit yield and signal amplitude stable, and with two-photon imaging and post-mortem histology showing seamless subcellular integration, fully recovered capillaries with an intact blood-brain barrier, and complete absence of chronic neuronal degeneration and glial scar. Frontier That is rodent work over months, and “glial-scar-free” is a histological claim about tissue, not a yield claim about an electrode over years. Speculative The extrapolation from no scar in a mouse at months to decade-stable in a human is the single largest unearned inference in the neural-interface literature. Frontier The design principle — match the mechanical compliance of the tissue, minimise the insertion footprint — is well supported and is the right principle. The nearest thing to a human test of it is Neuralink’s flexible threads, and the only public durability datum is that 85% of them retracted in the first participant.

Frontier Part of chronic stability turned out to be a software problem. The Neuropixels 2.0 paper pairs its 5,000-plus sites with newly designed analysis algorithms giving automatic post-hoc correction for brain movement, which is what allows recording from the same thousands of neurons across sessions in rodents at six laboratories. Frontier The same logic explains why Neuralink’s first participant recovered most performance in software after losing the great majority of threads. Speculative How far decoder adaptation can substitute for physical stability is a live and unquantified question, and it is the one place where the device layer and the application layer genuinely trade against each other.

Established Accelerated-lifetime testing exists, is cheap, and is almost absent from the intracortical literature. Lei and colleagues, Journal of Neural Engineering 13(4):046016, soaked candidate insulators in saline at 87 °C and measured silicon nitride dissolving at 18.3 ± 0.3 nm per day, thermally grown silicon dioxide at 0.104 ± 0.008 nm per day, and thin silicon carbide films showing no quantifiable change after 112 days, which they state is equivalent to about ten years in vivo; optimal SiC thickness was about 560 nm and defects appeared chiefly over rough surfaces and complicated topography. Frontier Silicon nitride dissolves roughly 176 times faster than thermal oxide and is disqualified for decade-scale implants on chemistry alone. No comparable published protocol exists for Utah arrays, Neuropixels probes or flexible polymer probes. Established Barrese’s parylene cracking and Woeppel’s tip degradation are the same problem observed in vivo, more expensively and more slowly.

Frontier The field’s frontier is defined by what it does not measure. Established There is no published human chronic yield from any manufacturer of a high-channel flexible implant. There is no published revision or replacement of a failed human intracortical array. Established There is no published comparison of two electrode technologies in the same person. There is no pre-registered chronic endpoint in any BCI trial. Established The entire human explant literature is six arrays from two people, and every array that comes out of a person and is not sectioned is a lost data point. Frontier The number of such arrays is now in the dozens.

Frontier The workback target is a thousand-channel implant that is stable for a decade in a human, and the chain to it has ten links of which one is scientific and two are institutional. Link one was, until July 2025, the binding link: publish yield-versus-time curves per subject, per array, per year. Frontier That required no new experiment, only reporting what was already collected, and it took twenty-nine years from the first implant to happen once, in a preprint. Speculative Everything downstream depends on it becoming routine rather than remaining singular.

Speculative Link seven is the scientific binding link: show that the tissue response does not scale badly with channel count. Frontier The dominant failure modes in the largest dataset are mechanical and material, and each has an engineering answer — wireless telemetry removes the connector, silicon carbide or thermal oxide removes insulation dissolution, compliance matching removes the extrusion driver, iridium oxide beats platinum on yield. Speculative But every additional electrode adds foreign surface, insertion trauma and vascular disruption, and there is no reason to assume the response stays constant as device volume grows. Whether encapsulation scales sublinearly with implanted surface area per cubic millimetre has never been measured, and the entire high-channel programme presupposes an answer to it.

Frontier Links two, three and four are cheap, use data that already exist, and would probably move the field further per dollar than any new material. Explain why 41% initial yield means six electrodes in ten never record a spike from the beginning. Frontier Explain the human-versus-primate gap with connector-handling frequency as a measured covariate. Predict chronic yield from pre-implant imaging by retrospective correlation across the existing fourteen-participant cohort. Speculative None of the three has been attempted.

Established Links nine and ten are institutional and bind on deployment regardless of any device result. No intracortical group has published a revision — an explant, a replacement, and a recovery of prior decoding performance. Established No instrument exists anywhere that obliges a manufacturer to support an implanted device after it stops selling them. Frontier A decade-stable implant inside a company that lasts five years is not a decade-stable implant. Frontier And what is not binding, despite twenty years of belief that it was, is the glial scar: the human data now say it is not what limits multi-year recording, and a research programme organised around it is organised around the wrong mechanism. Speculative The candid summary of the whole chain is that one link needs an animal experiment nobody has funded, three need an afternoon with data that already exist, two need a manufacturer to publish something unflattering, and two need a regulator to invent an instrument that does not exist in any jurisdiction. Speculative Only the first is a bottleneck in the scientific sense.

Established This brief depends first on a reporting norm rather than on a result. Frontier Yield-versus-time curves are collected by every group running a chronic implant and published by almost none, and until they are a routine secondary endpoint no comparison across devices, materials or surgical techniques is possible even in principle. The dependency is unusual in that the data exist, the analysis is trivial, and the obstacle is entirely a matter of what journals ask for and what sponsors permit.

Frontier It depends second on the application layer telling the device layer what to optimise. If the logarithmic scaling relation holds, the right engineering target is stability and reliability at existing channel counts rather than more channels; if some task family has a threshold, the target inverts. Speculative That question belongs to Brain-Computer Interfaces and its answer determines what this brief should be trying to build.

Established It depends third on ordinary materials and packaging science that is not neuroscience at all. Hermetic feedthrough design, thin-film insulator qualification, corrosion-resistant tip metallisation, implantable-grade wireless telemetry and low-power on-chip digitisation are all mature disciplines with their own literatures, and the neural-interface field mostly consumes them rather than advancing them. Frontier It depends fourth on the foreign-body-response literature reorganising itself around the mechanisms that the explant data actually implicate — meningeal fibrosis, oxidative degradation, placement variance — rather than around the glial scar. Speculative And it depends fifth on a regulatory pathway that permits a design change without restarting the investigational clock, which is discussed under institutional requirements and is not a scientific dependency at all.

Frontier The one dependency that is genuinely scientific runs through immunology rather than through materials. If the residual tissue response is to be managed rather than engineered around, somebody has to show that pharmacological suppression of fibroblasts and leukocytes — which the 2016 explant paper explicitly recommends — improves chronic yield without an unacceptable systemic cost. Speculative That is a transplant-medicine question asked of a brain implant, it has never been run in a human, and it is the only route by which the biological half of this problem gets solved rather than avoided.

Frontier The subsampling study. Take an existing high-channel human dataset, subsample electrodes at a spread of counts, and publish decoding accuracy against channel count on one fixed task with confidence intervals. Cost: an analyst and a month. Frontier It would either confirm the logarithmic relation on a second dataset or find the threshold that rescues the channel race.

Speculative The density study, which is the scientific binding link. Implant the same electrode material at four densities in one species and measure yield at two years as a function of implanted surface area per cubic millimetre. Pass condition: sublinear scaling of encapsulation with surface area. Speculative It is straightforward, expensive, unglamorous and has never been done.

Frontier The placement study, which is the cheapest high-value experiment available. Correlate pre-implant MRI and angiography against three-year yield across the existing fourteen-participant BrainGate cohort. Pass condition: a placement rule — cortical thickness, vascular density, pial adherence, distance from a sulcus. Frontier The data already exist and nobody has published the analysis.

Frontier The connector study. Publish a failure-mode analysis for the first cohort of fully wireless human implants, with acute mechanical failures falling from roughly 48% of the total as the pass condition. Wireless intracortical parity with wired was demonstrated in 2021 and no failure analysis has followed. Established The insulation study. Run accelerated soak testing at elevated temperature on an intracortical array with dissolution rates in nanometres per day, plus in-vivo confirmation at two years; the protocol exists and was published for retinal implants a decade ago. Frontier The comparison study. Implant a flexible high-channel design and a Utah array bilaterally in the same person against a common endpoint. The revision study. Explant a failed array from a human, replace it, recover prior decoding performance, and publish the case report. Speculative Nobody wants to be first to publish an explantation, which is why the single most useful clinical datum in the field is missing.

Frontier The yield-at-implantation study, which is the one nobody has framed as an experiment at all. Six electrodes in ten never record a spike from the first session, and the field has no model of what distinguishes a recording site from a non-recording one. Frontier The measurement is an acute large-animal study correlating per-electrode recording success against tip geometry, insertion depth, local vasculature and cortical layer position, with post-hoc histology. Speculative Pass condition: a model that predicts which sites will record. Speculative It is the difference between a 96-channel device and a 40-channel one, and it costs less than a single human trial.

Established Remove the connector first. The percutaneous pedestal is the largest single failure category in the largest dataset and the main chronic infection route, and a fully implanted subcutaneous transmitter has already run for years in a human without one. Frontier Wireless telemetry at the data rates a hundred-channel broadband array produces is a power and thermal problem before it is a radio problem, and on-chip spike detection or compression is the standard answer.

Established Then qualify the insulation stack against a ten-year target rather than a two-year one. Silicon nitride at 18.3 nm per day is disqualified; thermal oxide at 0.104 nm per day and silicon carbide with no measurable loss after 112 days at 87 °C are candidates; optimal SiC thickness is about 560 nm and defect density rises over rough topography, which constrains the shank geometry the coating has to cover. Frontier Parylene cracking and delamination, silicone peeling from array edges and platinum corrosion with underlying silicon degradation are the observed in-vivo failure set, and they are the set a soak protocol should be designed to predict.

Established Choose the tip metallisation for the job. Iridium oxide beats platinum on recording yield in two independent datasets; iridium oxide under chronic stimulation shows adhesion loss that platinum does not, without measurable effect on recording. Frontier A bidirectional device therefore has a genuine materials trade-off that a read-only device does not, and no published protocol resolves it. Speculative Beyond that the requirements are compliance matching to reduce micromotion, an insertion footprint small enough to leave the vasculature intact, and a thermal budget that respects the roughly 40 mW ceiling estimated for a rodent brain and scaled accordingly — a set of constraints that pull against channel count at every point.

Frontier And design for removal, which nothing currently does. An array whose tracts have vanished at 1,051 days because meningeal fibrosis lifted it off the cortex is an array that came out of the tissue by accident rather than by plan. Speculative A device intended to be revised needs an anchoring strategy that survives the dura, a lead route that can be cut and rejoined, and a connector geometry that a second surgeon can mate to; none of those is a research problem and none of them appears in any published intracortical design. Speculative The engineering requirement that would change the field’s risk calculus most is not a better electrode but a device somebody is willing to take out.

Frontier The strongest adjacent technology is the one that avoids the problem entirely. Established The endovascular stent-mounted array reaches motor cortex through a vein, carries 16 electrodes, and has a four-patient one-year safety record with no serious adverse events — one sixtieth the channel count of the flagship penetrating device and a risk profile a cardiologist would recognise. Frontier Subdural thin-film arrays that never break the pia are the other non-penetrating route, and the fully implanted ECoG systems already in humans have run for years without a percutaneous connector.

Established The mature manufacturing comparators are cochlear and retinal implants, and they are more instructive than any intracortical result. The cochlear implant is the only mass-deployed neural prosthesis — of order 600,000 recipients worldwide as of 2016 — with device failure requiring reimplantation estimated at 2.5–6% and revision surgery rates of 3.8% to 8%. Frontier Those are the only mature revision statistics for any neural implant anywhere, and the intracortical field has no equivalent because it has never published a revision.

Frontier Two adjacent fields supply capability the device layer cannot. CMOS-integrated probe manufacturing, which is what makes a 960-site shank possible at all, is semiconductor foundry work rather than neuroscience, and the packaging that turns a foundry die into an implant is where the discipline’s actual difficulty sits. Frontier Drift-correction and decoder-adaptation software is the second, and it now does part of the job the hardware was supposed to do. Speculative Further out, non-penetrating deep-target modulation by temporal interference has been demonstrated in the human hippocampus with cadaver field validation, and free-floating distributed sensors reporting out by ultrasound or molecular means have been analysed as physically feasible while requiring synthetic-biology advances nobody has.

Speculative The strongest version of the non-penetrating argument is that the field has been solving chronic stability for a class of device it does not need for its best applications. Frontier What would have to be true is that the information available without breaking the pia suffices for the target application, and for communication and cursor control the electrocorticography record says it probably does — the 78-words-per-minute avatar result was an ECoG result, and the deployed fully implanted system that ran autonomously for years used subdural electrodes. Established For single-neuron work it certainly does not. Frontier The honest position is therefore a split verdict by application rather than a verdict on the technology, and it is the clearest case in this brief of a device debate that is really an application debate in disguise.

Established The most complete device failure in the deployed record was a corporate decision. The Argus II retinal prosthesis carried 60 electrodes with 55 enabled in a 6×10 grid, took a CE mark in March 2011 and an FDA Humanitarian Device Exemption in February 2013 authorising use in up to 4,000 patients a year, and was supported by a single-arm pivotal trial of 30 patients with a longest follow-up of 38.3 months and nine serious adverse events. Established It cost about $150,000 in the United States excluding surgery and training. In 2020 Second Sight ceased technical support with devices still implanted in patients, and an investigation reported that users risk, and in some cases had already experienced, a return to blindness. Established The company merged with Nano Precision Medical in August 2023 with a commitment to restore support. Frontier Implant longevity failed corporately, not materially, and no technical review of neural-interface durability includes this case.

Frontier Nothing obliges a manufacturer to outlive its implants. Speculative The candidate instruments are an escrow of manufacturing and firmware assets, an open interface standard that lets a third party build a replacement external unit, or a condition of approval requiring funded support for the expected device life. None exists in any jurisdiction, and a device regulator that approves a ten-year implant against a two-year trial is making an implicit promise it has no mechanism to keep.

Speculative Regulation may also be why the field is still measuring a thirty-year-old design. The Utah array’s geometry is essentially unchanged since the 1990s, and every group that wants human data uses the device that already holds an investigational device exemption, because changing it restarts the pathway. Speculative The field is therefore characterising the chronic performance of the only design it is allowed to implant, which is an institutional constraint that presents as a scientific finding. Established A related concentration is worth naming: a large fraction of the world’s human chronic-interface knowledge — the 1,500-day stimulation record, the six-array explant study, the tactile-feedback result — comes from one centre and a handful of participants at Pittsburgh, and the largest interventional dataset in existence is fourteen people over twenty years.

Established The consent problem in this field is about the end of the study, not the beginning. Frontier A participant enrolls in a trial with a defined duration and receives a device with an undefined one, and the published record contains no case in which an intracortical array was electively removed and replaced with recovery of function. Speculative The honest disclosure is that the device may stop working, that nobody knows what happens next, and that the answer will be improvised.

Frontier Abandonment is a distinct harm from failure and it has happened. Established The Argus II cohort lost a working therapy to a business decision rather than to gliosis, which means the risk a participant carries includes the solvency and attention span of the sponsor. Speculative A right to continued support, or at minimum to an orderly explantation funded by someone, is the plainest missing entitlement in the area and does not require any new science.

Frontier The research burden falls on an extraordinarily small number of people. Fourteen BrainGate participants over twenty years carry the entire human chronic record, several of them contributing to multiple landmark results, and each additional experiment competes for the time of a person who has substantial care needs and a finite implant. Speculative Explant analysis, which is the single most informative measurement available, requires an array to come out, which sets the researcher’s interest against the participant’s. And the escalation logic of the field — more channels, larger craniotomies, more implanted surface — increases surgical risk for a capability gain that the one published scaling curve says is logarithmic. Speculative That is a risk-benefit argument that the device layer, not the application layer, is obliged to make.

Speculative There is also a quieter question about what the participant is entitled to know. Frontier A prospective participant asking how long the device will work is asking a question whose only honest answer, until July 2025, was that the field had never published the number. Speculative Consent to a procedure whose durability is unmeasured is a different act from consent to one whose durability is known and poor, and the difference is not usually drawn in the documents. Speculative The corresponding obligation is not a new ethical principle but an old one applied properly: publish the chronic outcome, per array, so that the next person can be told.

Speculative If the logarithmic relation holds, the long-run shape of this technology is not bandwidth. A civilisation with a mature neural-interface industry would be building reliable, revisable, low-channel devices for large populations rather than high-channel devices for a few, because the marginal channel buys a shrinking increment and the marginal decade of reliability buys everything. Frontier The cochlear implant is the existence proof and also the schedule: about fifty years from first demonstration to mass deployment, with the manufacturing, reimbursement and revision practice arriving long after the physics.

Speculative The deeper implication is about maintenance rather than capability. A society that implants devices meant to last a decade acquires an obligation with no current institutional home: someone must keep the firmware, the external units, the surgical expertise and the spare parts alive across corporate lifetimes that are shorter than the implants. Handwave At population scale that is closer to a utility than to a medical-device market, and no jurisdiction has begun to build the instrument. Speculative The failure mode of a mature neural-interface civilisation is therefore likely to look like the Argus II episode multiplied, rather than like a materials failure.

Speculative There is a third implication that the field’s own framing obscures. The interface that reaches the largest number of people will probably be the one with the fewest channels, because the applications with the largest addressable populations — communication in severe paralysis, switch access, environmental control — sit well below a hundred channels, and the delivery routes that reach them are the ones a surgeon can justify to a healthy-enough patient. Handwave A civilisation in which neural implants are ordinary would then be one in which the exotic high-channel device remains a research instrument indefinitely, in the way that a synchrotron remains a research instrument while X-rays became a clinic. Speculative That is not a pessimistic conclusion about the technology; it is a claim about which part of it scales, and the one published scaling curve is the only quantitative evidence anybody has for it.

These horizons track the device layer — longevity, yield, failure modes and the instruments that would measure them — rather than the applications, which move on their own schedule.

• 10 yr: Frontier Per-array yield curves become a routine secondary endpoint, or conspicuously do not; the Hahn dataset is either replicated by a second consortium or remains the only one. The subsampling analysis is published and the logarithmic relation is either confirmed on independent data or bounded to one task family. Frontier A fully wireless human cohort accumulates enough follow-up for a failure-mode analysis, and the connector hypothesis is tested for the first time. Speculative At least one flexible high-channel implant reports multi-year human yield, most plausibly under commercial rather than academic auspices and therefore with the disclosure problems that implies. An accelerated soak protocol for intracortical arrays is published by somebody, because it costs almost nothing.
• 25 yr: Speculative Chronic yield above 60% at five years in humans is either routine or has failed to arrive, and the difference is decided by the density study that nobody has yet run. Revision becomes an ordinary clinical procedure with published outcomes, or the field continues to implant devices it does not know how to replace. Speculative A device-support obligation exists in at least one jurisdiction, prompted by a second Argus-style abandonment rather than by foresight. Frontier Non-penetrating routes take the large-population applications and penetrating arrays retreat to the applications that genuinely need single neurons.
• 50 yr: Speculative A thousand-channel implant with a documented ten-year human service life exists, or the field has settled on a hundred-channel device with a thirty-year one and considers that the better trade. The foreign-body response is managed pharmacologically as a matter of routine, in the way transplant rejection is, rather than engineered around. Handwave Manufacturing has moved from custom fabrication to something resembling a foundry process, which is the precondition for the cost falling to where a payer stops arguing.
• 100 / 250+ yr: Handwave Either the interface has stopped being a foreign object — grown, recruited or biologically integrated rather than inserted — or the penetrating electrode is a historical curiosity replaced by something that reads from outside the parenchyma entirely. Handwave The materials question resolves either way, and the institutional question — who maintains an implanted device across a human lifetime — does not resolve at all unless somebody builds the instrument for it.

• Depends on This brief sits underneath the application layer rather than after it, and its dependencies are unusually mundane. It waits on a reporting norm: yield-versus-time curves per subject, per array, per year, which every chronic-implant group collects and almost none publishes, and without which no comparison of materials, geometries or surgical techniques is possible even in principle. It waits on ordinary packaging and materials science — hermetic feedthroughs, thin-film insulator qualification, corrosion-resistant tip metallisation, implantable-grade telemetry and low-power on-chip digitisation — none of which is neuroscience and all of which the field consumes rather than advances. It waits on the foreign-body-response literature reorganising around the mechanisms its own explant data implicate, which are meningeal fibrosis, oxidative degradation of insulation and placement variance, rather than around the glial scar. And it waits on one animal experiment nobody has run: the same electrode material at four implant densities, with yield at two years measured against implanted surface area per cubic millimetre.
• Requires (not on this map) Two institutional constraints bind this brief harder than any material does. The first is a reporting standard: chronic yield, per array and per year, published as a matter of course. It took twenty-nine years from the first human implant for one consortium to assemble that curve, it is still a preprint, and every downstream question in this brief — which material, which geometry, which placement, which species difference — is unanswerable until the measurement is routine rather than exceptional. The second is a device framework written for the timescale these implants actually occupy. Approval regimes evaluate a two-year trial and license a device that will sit in tissue for a decade, with no mechanism obliging the manufacturer to still exist at the end of it; the Argus II cohort lost a working prosthesis to a corporate decision rather than to a biological one. An escrow, an open interface standard or a funded support condition would each close the gap, and none of them requires a scientific result.
• Enables What this brief enables is the application layer above it. Every claim about what a brain-computer interface will do for a patient over years is a claim about the device, and the honest version of that claim can only be written once the yield curve, the failure distribution and the revision practice exist. The concrete enabling results are three: a connector-free architecture, which would remove the largest single failure category and test the hypothesis that has organised this brief; an insulation stack qualified against a ten-year target rather than a two-year one, which is the difference between an implant and an experiment; and a published revision, which converts a permanent decision into a reversible one and changes what it is reasonable to offer a patient. None of the three requires a scientific advance. All three require somebody to do the unglamorous work and publish it.
• Adjacent The adjacent work is mostly outside neuroscience. Endovascular delivery borrows the entire interventional-cardiology stack and reaches cortex without a craniotomy at one sixtieth the channel count. Subdural thin-film arrays borrow flexible-electronics manufacturing. CMOS-integrated probes borrow semiconductor foundry process. The cochlear and retinal implant industries own the only mature reimbursement, manufacturing and revision practice for any neural device, and are the correct model for what this field will need rather than a curiosity beside it. Decoder-adaptation and drift-correction software is adjacent in a stranger way: it now performs part of the stability job the hardware was designed to do, and the boundary between a stable electrode and a well-corrected recording is no longer clean.

Frontier “Utah arrays last five to ten years.” And “Utah arrays fail within a year.” Established Both are true, of different species, and the confusion is systematic. In non-human primates the median recording duration was 182 days and 56% of failures occurred within the first year. Frontier In humans the mean implant duration is 1,063 days, the maximum 7.6 years, and the yield decline across the whole enrollment period 7%. Established Enthusiast writing quotes the human maxima; sceptical writing quotes the primate medians; neither is dishonest and the pair is misleading. Frontier The honest sentence is that in humans most arrays keep recording usefully for several years, and in monkeys most fail within one.

Established “The glial scar kills the electrode.” This is the field’s own folk model and its own data contradict it. Established The largest histological study of chronically explanted arrays found gliosis with minimal neuronal loss and concluded that it does not prevent several years of useful recording; the failure it identified was meningeal fibrosis lifting the array out of the cortex, together with oxidative degradation of the materials. A human explant study later showed that part of the apparent scar on one array was pre-existing pial membrane, identified by matching its vascular pattern against pre-implant imaging. Frontier Human yield declines 7% over years. Speculative The steelman for the other side is real and comes from the same literature: the 2017 Nature Biomedical Engineering reframing of glia argues that they are not a passive barrier but an active determinant of implantation outcomes, remodelling neuronal networks and acting as an effector of stimulation-based therapy — on which reading glia matter more, not less, but in ways the scar model never captured. Handwave It is also worth recording that the founding review of this literature carries 1,942 citations and its companion 909, that neither abstract could be retrieved through any route available to this brief’s research pack, and that the frequently repeated neuronal-density figure from the second is therefore not printed here. Frontier A claim repeated 2,800 times whose primary text is hard to reach is exactly the kind of claim that drifts.

Frontier “More channels, more capability.” The one published curve is logarithmic. Frontier Six electrodes in ten never record a spike even at implantation. A patient who lost 85% of his threads recovered most performance in software. Frontier At the counts currently in contention, channel count is a manufacturing achievement rather than a capability metric. Speculative The genuine counter-case is that the tasks so far measured — cursor control, communication, movement intent — may all live below the threshold, and that dexterous individuated finger and wrist control with force, or full-field high-acuity vision, might need far more read and write channels than communication does. Frontier Nobody has published the curve for either.

Frontier “Flexible polymer probes have solved chronic stability.” Established They have demonstrated glial-scar-free integration with stable yield over months, in rodents. There is no multi-year human yield curve for any flexible probe, and the only public human durability datum for a flexible high-channel implant is 85% thread retraction in the first patient. Frontier “Neuralink’s yield beats the Utah array.” Established Neuralink’s published spiking yield of 45.60 ± 0.03% comes from 19 animal surgeries at a stated detection threshold; the Utah array’s 35.6% comes from 20 arrays in 14 humans over twenty years. Frontier Different species, different thresholds, different analysis pipelines. The comparison cannot presently be made, and that is itself the finding: no human chronic yield has been published for any high-channel flexible implant by anyone.

Established “Rising impedance means the electrode is failing.” Impedance rises early without affecting recording quality, then falls as insulation degrades; in the human explant study the lowest-impedance arrays had the lowest signal amplitude and the worst material degradation. Frontier The intuitive sign is backwards. Established “Chronic stimulation destroys the interface.” Over 1,500 days of regular intracortical microstimulation in a human, detection thresholds fell from 31.5 to 10.4 µA — sensitivity improved — and stimulated iridium-oxide electrodes were more likely than unstimulated platinum ones to keep recording large signals. Established Separately, one month of stimulation damaged 41.1% of stimulated tips without affecting recording performance. Frontier The coating degrades; the function does not.

Frontier “Neuropixels probes are used in human brains.” Established They are, acutely, during surgery, for minutes to hours, yielding over 200 well-isolated cortical single units simultaneously with eight separable waveform classes. No chronic human Neuropixels implant has been reported, and the “stable, long-term” claim in the Neuropixels 2.0 paper refers to mice and rats. Frontier “Nobody knows how long these things last, so all the claims are hand-waving.” As of July 2025 somebody does: 20 arrays, 14 people, 2,319 sessions, a 7.6-year maximum, 35.6% mean yield, a 7% decline. Frontier The correct criticism is not that the number is unknown but that it took twenty-nine years to publish and is still a preprint.

Established “n=1” is doing more work in this literature than any other phrase. The following are all single-participant results that circulate in public writing without the qualifier: the 62-words-per-minute speech interface, the 78-words-per-minute avatar interface, the 97.5%-over-8.4-months speech result, the tactile-feedback robotic-arm result, the cortical visual prosthesis, the brain-spine interface, the closed-loop depression device, the completely-locked-in speller, the fully implanted ECoG typist, and the five-year array-longevity retrospective. Frontier The individual corrections are less persuasive than the pattern, and the pattern is that the largest interventional dataset in the entire field is fourteen people.

Speculative “The device is the hard part.” Established The largest single failure category in the largest dataset was connectors. The most complete device failure in the deployed record was a company withdrawing support from patients who still had the implant inside them. Frontier The three cheapest available improvements — predicting placement from pre-implant imaging, explaining the 41% initial yield, qualifying an insulation stack by accelerated soak — are analyses of data or protocols that already exist. Handwave And at the far end of the hypothesis space sits the genuinely exotic proposal: an interface that recruits glia rather than repelling them, conductive polymers that grow into tissue, engineered cells acting as the transducer. What would have to be true is control over a chronic tissue response at a level nobody has in any organ, plus a regulatory route for a living implant. Handwave It is cheerfully far from evidence, and the nearest real thing to it is a 2017 review arguing that glia should be recruited rather than excluded, which is a research programme rather than a device.