Brain-to-Brain Networking

Speculative If a direct link between two nervous systems ever carries something a mouth cannot — which no published experiment has shown, and 7 says what would settle it — the prize is a message that never has to be encoded into words at all. That is the only prize on offer here that speech does not already supply more cheaply. Everything else a brain-to-brain channel has so far delivered, a person could have said aloud in a fraction of the time.

Established The arithmetic against the baseline is brutal and it is the spine of this brief. Across 17 languages and 10 native speakers each, Coupé and colleagues measured human speech at roughly 39 bits per second, with syllable densities between 5 and 8 bits. The only direct human-to-human brain-to-brain channel that has ever published its own throughput ceiling — Rao and colleagues in 2014 — put that ceiling at 0.05 bits per second, set not by neuroscience but by the 20-second spacing that transcranial magnetic stimulation safety guidance imposes between pulses. The ratio is about 780 to one. Put the other way: one second of ordinary conversation takes that channel about thirteen minutes.

Frontier The one result that points at a prize speech cannot match is a rat feeling a whisker gap it never touched. In the 2013 Duke experiment, decoder rats that received cortical microstimulation derived from an encoder rat’s somatosensory activity discriminated aperture width at 62.34 ± 0.59 per cent correct, and selected levers in a motor task at 64.32 ± 1.1 per cent, both above chance. Those are low numbers, but they are the only ones in the literature where the thing crossing the link resembles a percept rather than a symbol the two parties agreed on beforehand. Every human demonstration to date has moved a pre-arranged bit, which is precisely what language already does well.

Speculative The people who would gain are a small and specific group, and the condition on their gaining anything sits on the write side of the link. A person with no reliable motor output has no cheap alternative channel, so for them even 0.05 bits per second is not competing with 39; it is competing with zero. For everyone else the direct channel would have to beat a keypress, and a single press chosen from a thirty-key set carries about 4.9 bits in well under a second. What has to become true is a way of writing into a receiving brain that can repeat faster than once every eight seconds without exceeding a safety envelope nobody has yet drawn for healthy volunteers — and the list in 5 starts there.

Established A brain-to-brain interface is not one device; it is a brain-computer interface bolted to a computer-brain interface, with a network in between. Brain A is read, usually by electroencephalography; the reading is classified into a symbol by software; the symbol travels as ordinary network traffic; and a stimulator writes something into brain B. Nam and colleagues, in the field’s only systematic review, describe exactly this two-stage architecture across every study they found.

Frontier The word doing the most work in the phrase is “direct”, and it is doing it dishonestly. In the 2014 University of Washington demonstration the signal left the sender’s scalp, was thresholded into a binary decision, crossed a mile of campus network, and fired a magnetic coil over the receiver’s motor cortex. The authors themselves measured the internet leg at about 10 milliseconds inside a total loop of roughly 650 milliseconds, of which some 627 milliseconds was the receiver’s own muscle responding. Calling that link direct is a claim about the absence of a keyboard, not about the absence of encoding.

Established The taxonomy is small because the field is small. Nam and colleagues screened 193 records and found 15 experimental studies published between 2013 and August 2020. Of those, 73.3 per cent (11 studies) were direct unidirectional links and 26.7 per cent (4) were indirect bidirectional, where the return path runs through the eyes rather than through the skull. They found zero direct bidirectional systems. Sender-receiver pairs accounted for 86.7 per cent of studies; exactly one used an N:N configuration and one an N:1.

Established The species mix tells you where the invasive work happens. In the same census, 46.7 per cent of studies (7) were human to human, 26.7 per cent (4) human to rodent, 20 per cent (3) rodent to rodent, and one study ran human to cockroach. Read side: EEG in 80 per cent of studies, intracortical electrodes in 13.3 per cent, optogenetics in 6.7 per cent. Write side: transcranial magnetic stimulation and intracortical microstimulation at 40 per cent each, transcranial focused ultrasound at 13.3 per cent, optogenetics at 6.7 per cent. Humans get the non-invasive read and the non-invasive write; animals get the electrodes.

Established Because the output is another person’s experience rather than a machine action, bit rate here is not one metric among several — it is the entire product. A brain-computer interface that types slowly still types, and its comparison class is a person who cannot type at all. A link between two able-bodied people competes against their mouths, and that competition can only be scored in information per unit time, which is why 3 is organised around throughput.

Frontier The receiving end has no addressing scheme, and that is a conceptual gap rather than an engineering one. Writing a bit into a brain means producing an event the recipient can notice: a phosphene, a twitch, a tingle. None is a natural symbol, so the recipient must be told in advance what it means — the semantics travel by ordinary language before the session starts, and only the selection travels over the link.

Established The field has a clean founding experiment and it was in rats. Pais-Vieira and colleagues at Duke, publishing in Scientific Reports in February 2013, paired encoder and decoder rats through cortical microstimulation. Experiment 1 used 3 encoders and 4 decoders; Experiment 2 used 2 encoders and 5 decoders. In the motor task the decoder rats chose correctly on 64.32 ± 1.1 per cent of trials (range 60–72 per cent); in the tactile discrimination task, 62.34 ± 0.59 per cent (range 60–64.58 per cent). Both exceeded chance at p < 0.05. Local transmission ran at about 20 milliseconds.

Established That same paper said plainly what the field has still not answered. Its authors wrote that “the limit for the amount of information that can be transferred by unit of time (i.e. capacity) is currently unknown for a BTBI channel”, and identified the binding quantities as encoder firing-rate range and decoder microstimulation threshold. Thirteen years later no published study has measured that capacity for any pairing. The same group also ran the link from Natal, Brazil to Durham, North Carolina, where mean transmission time rose from about 20 milliseconds to 232 ± 217.5 milliseconds without a comparable loss of decoder performance — a result about network latency, not about brains.

Established The first human-to-human demonstrations arrived in 2014, and both moved a single pre-agreed bit. Grau and colleagues, working out of a European FET Open project called HIVE, encoded the words “hola” and “ciao” as binary strings, sent them over the internet from India to France, and delivered them as phosphene flashes in the peripheral vision of receivers under robot-positioned transcranial magnetic stimulation. Four healthy participants aged 28 to 50 took part: one on the reading side, three on the writing side. The Beth Israel Deaconess release reporting the work puts the distance at about 5,000 miles and a second run between Spain and France at a total error rate of 15 per cent, split as 11 per cent at the decoding end and 5 per cent at the coding end.

Frontier What Grau and colleagues demonstrated was a telegraph with a very poor operator, not a thought transfer. The greeting was translated into binary before it left anyone’s head; the receiver counted flashes and applied a lookup table agreed in advance. No part of the semantic content of “hola” crossed the link. This brief could not open the primary paper — the publisher’s article endpoint served a different article on every attempt — so the figures here come from the European Commission news item and the hospital press release, both of which are interested parties, and the per-bit timing of that demonstration is therefore not stated in this brief.

Established Rao and colleagues published the number that governs the whole field. In their 2014 PLOS ONE paper, six participants formed three sender-receiver pairs about a mile apart on the University of Washington campus. Sender motor imagery was decoded from EEG and used to fire transcranial magnetic stimulation over the receiver’s motor cortex, involuntarily moving the receiver’s hand onto a touchpad to fire a cannon in a cooperative game. Hit rates across the three pairs were 83.3, 25.0 and 37.5 per cent. In control blocks with the link disabled, the rate was zero.

Established The ceiling is 0.05 bits per second, and it is a safety number rather than a physics number. The paper states that the enforced pause “prevented two consecutive TMS pulses from being delivered less than 20 seconds apart, thus reducing the maximum amount of magnetic stimulation delivered to the Receiver to a level well below the strictest safety guidelines, and setting the upper limit of information throughput to 0.05 bits per second”. Measured mutual information between sender and receiver response vectors ran from 0.25 to 0.81 bits per trial, totalling about 4 to 13 bits per experimental block of 10 to 16 trials. The whole loop took roughly 650 milliseconds, of which about 627 milliseconds was the receiver’s muscle.

Established BrainNet is the anchor result and it holds up, with one correction to how it is usually described. Jiang and colleagues, Scientific Reports 2019, networked five triads — 15 participants in total, two Senders and one Receiver per group — on a Tetris-like task in which the Receiver, who could not see the bottom of the screen, had to decide whether to rotate a falling block. Mean accuracy was 81.25 per cent, which is 13 of 16 trials, at p = 0.002 on a binomial test, with a mean area under the curve of 0.83 at p < 0.001. The commission for this brief described BrainNet as reaching “above 80 per cent”, which is right; the figure is 81.25 per cent and the session length is 16 trials, so the whole published human N:1 record is 80 trials.

Established BrainNet’s own numbers set its rate, and the paper does not state it. Senders selected between two steady-state visually evoked potential frequencies, 17 Hz for “rotate” and 15 Hz for “do not rotate”. The classifier tallied one-second epochs and, in the authors’ words, made the final decision “over a 10-second period”. On the receiving side, “an eight-second delay was enforced between the two pulses to remain within the strictest safety guidelines of TMS stimulation”, with phosphene threshold set by the PEST method and then adjusted in 5 per cent increments. A round therefore cannot run faster than roughly eighteen seconds for one bit of Receiver decision, which is under 0.06 bits per second — arithmetic on the paper’s stated timings, not a figure the paper reports.

Established The paper is explicit that the payload has never grown. Its own framing notes that across this literature “the level of information complexity has remained binary, i.e., only a bit of information is transmitted during each iteration of communication”. That sentence, from the field’s most-cited human result, is the most useful thing in it.

Frontier BrainNet’s most interesting finding is about trust, not about bandwidth. The experimenters corrupted one Sender’s channel with injected noise and found that Receivers learned to weight the reliable Sender more heavily, at Z = 5.87, p < 0.001 on regression beta weights and Z = 7.31, p < 0.001 on correlation coefficients. This is a real result about social learning through a degraded channel. It is not a result about the channel carrying more.

Established The multi-brain animal work shows pooling helps and says nothing about communication. Ramakrishnan and colleagues connected four rhesus macaques into what they called a Brainet, splitting control of a virtual arm between two or three animals. In the three-monkey 3D configuration performance rose “from 20% correct trials in the first session to 78% correct trials in the last session”, and the authors report that decoding accuracy “benefited from mixing the contributions from different brains as well as the overall neuronal mass”. Nothing passed between the monkeys; a shared decoder averaged their signals. That is ensembling, and it is the same trick a second electrode array in one brain would buy.

Established Cross-species links are the highest-accuracy results in the literature, because the receiver is an implanted animal. Zhang and colleagues, Scientific Reports 2019, had six human operators steer six implanted rats through a maze using motor-imagery EEG mapped to left, right and forward microstimulation commands. EEG decoding accuracy was 77.86 ± 12.4 per cent; turning accuracy under brain control reached 91.75 ± 3.85 per cent with their gradient model and 93.32 ± 1.73 per cent with their thresholding model, and instruction-generation delay fell from 494.70 ± 47.22 milliseconds to 155.01 ± 3.10 milliseconds, a reduction the authors put at nearly 70 per cent. The receiving brain in that experiment has wires in it and no vote.

Frontier The single result that most threatens the pessimistic reading changed the write mechanism rather than the read. Lu, Wang and Luo, in Science China Life Sciences in 2020, built an all-optical link between mice: fibre photometry read calcium signals from neuromedin-B neurons in the nucleus incertus of a “Master” mouse and optogenetics wrote them into the same nucleus of an “Avatar”. Their claim is that the link ran “with information transfer rate about two orders of magnitude higher than previous BtBIs”. Two orders of magnitude on 0.05 bits per second is 5 bits per second, which is most of the distance to a conversation — obtained in mice, with a viral construct, in one nucleus.

Established The comparison ladder, in one place, all of it measured. Human speech: about 39 bits per second, from 17 languages at 10 speakers each, with syllable densities of 5 to 8 bits. The highest information transfer rate a clinical review in April 2026 could find for any brain-computer interface speller: 360 bits per minute, which is 6 bits per second, and that review still calls spellers “slow for everyday use”. An electrocorticographic matrix speller: a peak rate of over 100 bits per minute. A 96-electrode intracortical array in a tetraplegic participant: up to 10 correct characters per minute. Direct human-to-human brain-to-brain: 0.05 bits per second.

Established The gap between the direct channel and speech is about 780-fold, which is just under three orders of magnitude — not the millions often asserted. 39 divided by 0.05 is 780. The claim that the channel is “millions of times slower than speech” is a rate comparison, and as a rate comparison it overstates by roughly three orders of magnitude. Counted differently the gap is far larger than millions: in the eighteen seconds a BrainNet round consumes, speech emits about 700 bits and therefore selects among 2 to the power of 700 possible messages, while the link selects among two. Both framings are defensible; neither is “millions”, and a brief that wants the arithmetic to carry weight has to say which one it means.

Frontier Adjusting for the honest comparison class makes the case worse, not better. Against a brain-computer interface that decodes to a machine, the direct channel is about 120 times slower: 6 bits per second against 0.05. A single keypress from a thirty-key keyboard is about 4.9 bits, so one press beats the entire output of a BrainNet round fivefold in a fiftieth of the time. And against the best intracortical speech neuroprostheses — 62 words per minute at a 23.8 per cent word error rate on a 125,000-word vocabulary, 78 words per minute at 25 per cent on a 1,024-word set, each in one participant — it is not the same category of device.

Established The field’s own reviewers say the same thing in gentler language. Cinel, Valeriani and Poli, surveying neurotechnology for cognitive augmentation, conclude of the brain-to-brain literature that “the communication is restricted to very limited type of information, and that the ITR is very low”. Nam and colleagues report that the stimulation method “varied greatly between studies indicating no agreed upon neurostimulatory method for transmitting information”. Hildt, writing on the ethics of BrainNet specifically, notes that in it “there was no ‘mind reading’ or more complex information transfer”.

Frontier The most striking thing about the record is how thin and how old it is. The complete published human-to-human direct brain-to-brain corpus, as catalogued by the field’s only systematic review, is seven studies. The most-cited of them is 2019. The searches behind this brief found no human-to-human demonstration published after that date reporting a higher payload, and the 2026 clinical review of brain-computer interfaces does not mention brain-to-brain interfaces at all. That absence is evidence of a quiet field, not proof of one, and it is stated here as a search result rather than as a fact about the world.

Frontier The first open question is whether this channel is read-limited or write-limited, and the evidence points hard at the write side. EEG decoding in the human studies sits between about 78 and 83 per cent for a two-way choice, which is poor but not catastrophic. The write side imposes an eight-second inter-pulse delay in BrainNet and a twenty-second one in Rao and colleagues. If the read side improved to perfection tomorrow, the throughput ceiling would not move at all.

Frontier The Lu result is the strongest available test of that proposition and it agrees. Changing only the write mechanism — optogenetic stimulation in place of magnetic pulses — is claimed to have bought two orders of magnitude of information transfer rate in mice. Whether the same factor is available with any method a human volunteer would accept is the field’s central unanswered question, and no group has published an attempt.

Speculative The second open question is whether two brains share enough code for a neural pattern to mean anything in the other one. Every human demonstration sidesteps this by agreeing on a symbol in advance: a phosphene means rotate, a twitch means fire. Nothing in the literature shows that a cortical stimulation pattern derived from one person evokes a similar experience in another.

Frontier The rat tactile result is the closest thing to evidence that a code can be shared, and it is within a species with a stereotyped whisker map. Decoder rats discriminating aperture width at 62.34 ± 0.59 per cent were responding to microstimulation derived from another animal’s barrel cortex. The barrel cortex is among the most stereotyped cortical maps in any mammal. Nothing about human association cortex is stereotyped in that way.

Frontier The third open question is whether recipients adapt, and here there is a genuine positive. BrainNet’s noise-injection condition showed Receivers reweighting Senders by reliability at Z = 5.87 and Z = 7.31, both p < 0.001. That is learning about the channel. Whether recipients also learn to extract more bits per stimulation event has not been tested.

Frontier Inter-brain synchrony is a large, active literature and it is not a channel. Nazneen and colleagues reviewed 64 articles on brain-to-brain synchrony from non-invasive recording: 39 used EEG, 13 functional near-infrared spectroscopy and 7 fMRI, with phase-locking value in 16 studies, wavelet transform coherence in 11 and intersubject correlation in 6. Synchrony between interacting people is a correlate of joint activity, not a transmission. Confusing the two is the most common error in popular writing about this subject.

Frontier That literature also has a control problem it acknowledges. In the same review, 52 of 64 studies (82 per cent) flagged concerns about spurious synchronisation, but only 72 per cent implemented shuffled or random pair analysis as a control. Two people watching the same film synchronise because the film is the same. Establishing anything stronger requires the pseudo-pair control, and a quarter of the field omits it.

Speculative Whether more brains help at all is unresolved, and one measured answer is discouraging. Cinel, Valeriani and Poli report a collaborative brain-computer interface configuration that needed at least seven individuals’ signals to reach the behavioural performance of the average single observer. The monkey Brainet reached 78 per cent correct on a three-animal 3D task by the last session, up from 20 per cent, but that is a shared decoder pooling neurons, not brains negotiating.

Handwave The proposition that a network of brains constitutes a new kind of computational substrate is the field’s recurring rhetorical move and it has no supporting measurement. The BrainNet authors write of “future brain-to-brain interfaces that enable cooperative problem solving by humans using a ‘social network’ of connected brains”. Nothing in the 80 trials they ran distinguishes that vision from three people with a very slow intercom.

Frontier Nobody has measured channel capacity for any brain-to-brain pairing, which is the gap the founding paper named. Mutual information per trial has been reported; a capacity in the Shannon sense has not, for any read-write combination. Without it, claims that a method sits near or far from its limit are assertions.

Established The binding bottleneck is the duty cycle of non-invasive stimulation, and it is imposed by safety guidance rather than by hardware. Rao and colleagues enforced 20 seconds between transcranial magnetic stimulation pulses and stated that this set the throughput ceiling at 0.05 bits per second. Jiang and colleagues enforced eight seconds. Neither interval reflects what a coil can do; both reflect what an ethics committee will approve for repeated stimulation of a healthy volunteer.

Frontier Focused ultrasound is the obvious replacement write method and its spatial resolution is an order of magnitude too coarse. A 2025 review of focal neuromodulation puts transcranial focused ultrasound at a spatial precision “in the range of 1–5 mm”, using a “frequency range of 250–700 kHz”, with acoustic intensity kept “below 100 W/cm²”. A cortical column is a few hundred micrometres across. Writing a distinguishable symbol requires addressing something smaller than the current focus.

Established The read side is slow because the decision window is long, not because the classifier is bad. BrainNet’s steady-state visually evoked potential discrimination required tallying one-second epochs over a 10-second period to separate 17 Hz from 15 Hz. Shortening that window trades directly against accuracy, and the accuracy was already only 81.25 per cent at the group level for a binary decision.

Established The receiver’s own body is a substantial part of the latency budget. In the 2014 loop, the internet leg was about 10 milliseconds, coil activation about 1.4 milliseconds and pulse generation about 4.0 milliseconds — but the motor response took roughly 627.1 milliseconds. Making the network faster, which is where engineering attention naturally goes, addresses about 1.5 per cent of the delay.

Frontier The symbol alphabet at the receiving end has about two members and no obvious route to more. Phosphene versus no phosphene; twitch versus no twitch. Grading phosphene intensity, position or count has been tried only in the Grau protocol’s flash-counting scheme, and the reported 15 per cent total error rate there was for a handful of characters. An alphabet of 32 symbols would need 32 reliably distinguishable percepts under repeated stimulation, and no study has demonstrated more than a few.

Established There is no agreed stimulation method, which means results do not compose. Nam and colleagues found transcranial magnetic stimulation and intracortical microstimulation each used in 40 per cent of studies, ultrasound in 13.3 per cent and optogenetics in 6.7 per cent, and concluded there is “no agreed upon neurostimulatory method for transmitting information”. A field with four incompatible write mechanisms and fifteen studies cannot accumulate.

Frontier The DARPA non-surgical targets state what would be needed and the programme closed without publicly meeting them. Next-Generation Nonsurgical Neurotechnology set out to read and write “16 independent channels within a 16mm3 volume of neural tissue” at “within 50ms” and at “sub-millimeter” specificity, across six performer teams announced in 2019 — Battelle, Carnegie Mellon, Johns Hopkins Applied Physics Laboratory, Palo Alto Research Center, Rice and Teledyne. The programme page now records it as complete. Sixteen independent binary channels at 50 milliseconds would be roughly 320 bits per second, eight times speech; that this was the target says how far from it the field sits.

Speculative Even a perfect write mechanism leaves the interpretation problem untouched. Delivering an arbitrary spatiotemporal pattern to 16 cortical sites does not make the recipient experience anything in particular. The bottleneck after bandwidth is semantics, and there is no measured result in humans showing that a pattern copied from one cortex evokes a matching percept in another.

Frontier Fatigue and habituation are unmeasured over the sessions any real use would require. Every published human protocol is a single session of 16 trials or fewer per participant, and nothing reports what happens to accuracy or throughput over an hour, a day or a week.

Frontier The result this subject most needs from elsewhere is a non-invasive write method with a millisecond-scale repeat rate and a safety envelope drawn for healthy people. Transcranial magnetic stimulation cannot supply it at any clinically accepted duty cycle; focused ultrasound might, at 1–5 mm resolution and intensities below 100 W/cm², if repetition-rate safety were characterised. That work belongs to the neuromodulation community and is not being done for this application.

Speculative It needs a demonstration that neural codes transfer between individuals at all. This is a basic-science result, not an engineering one: show that stimulation patterns derived from person A’s sensory cortex produce reportably similar percepts in person B, above what a matched random pattern produces. Nobody has published an attempt in humans, and the closest animal evidence is a rat whisker-map result at 62.34 per cent correct.

Frontier It depends on read-side advances it does not control. Every human-to-human study used EEG, whose spatial resolution is centimetres and whose single-trial binary accuracy in these protocols was 78 to 83 per cent. Higher-bandwidth reading is being pursued for brain-computer interfaces with different goals; this field inherits whatever that work produces, as Brain-Computer Interfaces sets out.

Established It depends on an information-theoretic reporting convention that the field does not have. Of the fifteen studies in the systematic census, the ones that report throughput at all report it differently: mutual information per trial, bits per block, a safety-imposed ceiling, or a claim of “two orders of magnitude” relative to unspecified predecessors. Without a common metric, the central quantity in the field is not comparable across papers.

Frontier It depends on a reason for anyone to fund it. The clinical case is served by brain-computer interfaces that decode to a machine, further ahead and already in trials. The augmentation case competes with speech. The remaining case — transferring something not verbally expressible — would justify the programme, and is the one nobody has run the experiment for.

Speculative It depends on shared representation-learning, which is currently an artificial-intelligence problem. If neural codes do not transfer natively, the fallback is a learned mapping between two individuals’ representations trained on paired data — which converts a brain-to-brain link into two brain-computer interfaces with a translation layer, which is what it already is.

Frontier The decisive experiment is a head-to-head trial in which the same two people perform the same task through the direct neural channel, through speech, and through a single keypress, with information transfer measured identically in all three arms, plus a fourth arm carrying content the sender cannot put into words. No published brain-to-brain study has run it. Every control condition in the literature is the channel switched off — Rao and colleagues report 0 per cent of rockets destroyed in control blocks — which establishes that the link does something, and says nothing about whether it does anything a mouth would not do better.

Established That experiment needs no new hardware. Every rig required has existed since 2014: an EEG cap, a classifier, a network link, a stimulator, and a task with a scoreable outcome. Adding a spoken-channel arm and a keypress arm is a protocol change, not an instrument programme. This is the cheapest decisive test in this brief and the reason its absence is informative.

Frontier The fourth arm is where the subject either earns its own category or collapses into brain-computer interfacing. Candidate payloads that resist verbalisation: a somatosensory percept such as the aperture width the Duke rats discriminated; a motor plan with timing finer than speech can specify; a subjective magnitude such as perceived effort or confidence. If the direct channel beats speech on any of these under matched scoring, the field has a purpose. If it does not, the honest conclusion is that this is a slow, one-bit intercom.

Speculative The second experiment is the code-transfer test, and it is a basic-science result rather than a demonstration. Take stimulation patterns derived from person A’s sensory cortex during a known stimulus; deliver them to person B; ask B to report. Compare against matched random patterns and against patterns derived from a third party. If A-derived patterns do not beat random-derived ones, there is no shared code to exploit and every future system will need a learned translation layer.

Frontier The third experiment is a duty-cycle safety study, and it is the one that would move the ceiling. The 0.05 bits per second figure follows from a 20-second inter-pulse interval, and BrainNet used eight. Characterising what repetition rate is acceptable for repeated phosphene induction or motor-cortex stimulation in healthy volunteers is an ordinary neuromodulation safety study, and its result would multiply or divide the field’s headline number directly.

Frontier The fourth experiment is the ultrasound write test. Replace the magnetic coil with transcranial focused ultrasound at 250–700 kHz and 1–5 mm focus, measure how many distinguishable percepts can be induced per minute within intensity limits below 100 W/cm², and report the alphabet size. Nam and colleagues found ultrasound used as the write mechanism in only 13.3 per cent of the fifteen catalogued studies, so this is a thinly explored branch with a specific, answerable question attached.

Speculative The fifth experiment asks whether the Lu result generalises off mice. An all-optical link claimed two orders of magnitude over prior systems in animals with an engineered opsin in one nucleus. The test is whether any read-write pairing acceptable in a human buys a comparable factor. Absent that, 5 bits per second is a mouse number and should be reported as one.

Frontier The sixth experiment is longitudinal and nobody has run it. Every human protocol in the record is a single session with 16 or fewer trials per participant. Run one pair for twenty sessions and measure accuracy, throughput and subjective effort. Recipients demonstrably learn to weight senders by reliability; whether they learn to extract more per pulse decides whether this technology has a learning curve at all.

Frontier The seventh is a capacity measurement, which the founding paper asked for in 2013. Pais-Vieira and colleagues wrote that channel capacity “is currently unknown for a BTBI channel”. Estimating it for one fixed read-write pairing — by sweeping input distributions and measuring mutual information — would tell the field whether existing demonstrations sit near their limit or two orders below it. It has not been done for any pairing.

Handwave The experiment most often proposed in public — scale the network to many brains and see what emerges — is not a test of anything. BrainNet’s N:1 configuration with two senders produced 81.25 per cent on a binary decision. Adding senders without raising per-link capacity produces a slower committee, and the one measured multi-user result available reports needing seven people to match one average observer. An emergence claim needs a prediction that fails if emergence does not occur, and none has been offered.

Established A negative result here would be worth as much as a positive one, and cheaper. If the head-to-head trial shows the direct channel losing to a keypress on every payload including the non-verbal ones, that settles the augmentation case and redirects the remaining interest to people with no motor output — where the comparison class is zero and 0.05 bits per second is a real number.

Established The hardware on both ends is commodity, which is why the field’s slowness is not an equipment problem. Consumer-grade EEG caps, an open-source classifier, a standard TCP/IP link and a clinical transcranial magnetic stimulation system are the entire bill of materials for every human-to-human result in the record. The 2014 demonstration ran across a mile of ordinary campus network and the 2013 rat work ran from Brazil to North Carolina.

Frontier A usable device would need a write stage with a duty cycle at least two orders of magnitude faster than anything currently approved. Going from one bit per eight seconds to one bit per eighty milliseconds is the difference between 0.125 and 12.5 bits per second, which would put the channel within a factor of three of speech. Nothing in the approved neuromodulation toolkit does that.

Frontier Coil positioning is a precision problem the Grau protocol solved with robotics and nobody has miniaturised. That demonstration used robot-assisted, image-guided stimulation to hold phosphene induction stable; a wearable version needs sub-centimetre repeatability against head motion, the same constraint that keeps neuronavigated TMS in a clinic.

Established The receiver’s latency dominates and cannot be engineered away. With about 627 milliseconds of the 650-millisecond loop spent in the receiver’s muscle, any system whose output is a body movement inherits a floor near half a second. Systems whose output is a percept rather than a movement avoid this, which is one concrete argument for phosphene-style writing over motor-cortex writing.

Speculative An implanted version changes the engineering entirely and changes the indication with it. Intracortical microstimulation supports the fastest and most accurate links in the literature — 91 to 93 per cent turning accuracy in the human-to-rat work — because the receiving brain is wired. Putting electrodes in a healthy human to raise a communication rate that speech already exceeds is not a proposition any regulator has been asked to consider, and the device longevity and explant record in Neural Interfaces is why.

Frontier An all-optical architecture is the only demonstrated route to a large rate gain and it is not available in humans. Fibre photometry plus optogenetics in one nucleus produced the claimed hundredfold improvement in mice. Both ends require genetic modification of the target neurons and physical optical access. No human pathway exists, and the engineering question is whether any non-genetic method reproduces the gain.

Established The software stack is where most published effort has gone and it is the least interesting part. The Zhang group’s near-70 per cent cut in instruction delay, from 494.70 to 155.01 milliseconds, came from changing the command model. A real gain, and still inside a budget dominated by stimulation safety spacing.

Established The nearest neighbour is the brain-computer interface, and it is ahead on every metric this brief cares about. Spellers reach 360 bits per minute at their best reported rate; intracortical speech decoding reaches 62 and 78 words per minute in single participants. Brain-Computer Interfaces owns that record, including its small denominators. The distinction that justifies a separate brief is the destination: a brain-computer interface decodes to a machine, and the machine can then do anything, including speak.

Frontier Hyperscanning shares the vocabulary and none of the mechanism. The 64-study synchrony literature measures correlated activity between interacting people using EEG, functional near-infrared spectroscopy and fMRI. It is an observational science of joint action. No synchrony study transmits anything, and treating synchrony findings as evidence for a communication channel is a category error that Distributed Cognition and Collective Intelligence both have cause to guard against.

Established Neuromodulation supplies the write side and has its own clinical agenda. Transcranial magnetic stimulation, transcranial focused ultrasound at 250–700 kHz, and implanted stimulation are developed for depression, tremor, epilepsy and pain. Communication is not on their roadmap, so this field consumes their safety envelopes rather than shaping them, as Bioelectric Medicine describes from the therapeutic side.

Speculative Memory transfer is the adjacent claim most often confused with this one. Writing a remembered episode into another brain is a different and much harder proposition than writing a bit, and it depends on results that Memory Engineering treats. Hildt’s 2015 analysis names the specific problem it would create: a recipient unable to distinguish their own memories from transferred quasi-memories.

Frontier Cross-species links are the highest-performing systems in the census. Four of the fifteen catalogued studies ran human to rodent and one human to cockroach, with turning accuracies above 90 per cent because the receiver is implanted and has no competing agenda. Interspecies Communication covers the signal-meaning question those systems do not address.

Speculative Augmentation framings pull this subject toward a literature with different standards of evidence. Cinel, Valeriani and Poli place brain-to-brain work inside cognitive augmentation with the caveat that its transfer rate is very low; Human Cognitive Augmentation and Cognitive Enhancement hold the wider claims.

Established The institution that actually governs the throughput number is the research ethics committee. The 0.05 bits per second ceiling exists because the investigators chose a 20-second inter-pulse interval to stay “well below the strictest safety guidelines”. The headline quantity in this field is therefore set by a conservative reading of stimulation safety guidance, not by an instrument specification — which means an institutional change could move it without any new science.

Frontier No regulator anywhere has a pathway for a communication device whose output is another person’s neural stimulation. Stimulators are regulated as therapeutic devices with an indication. A two-person link has two users, one of whom is being stimulated on the authority of the other, and no existing device class contemplates that. Trimper, Wolpe and Rommelfanger noted in 2014 that there is “no legislation regulating informed consent and protecting personal data extracted via BCIs, much less BTBIs”, and this brief found nothing since that closes the gap.

Established The defence funder that did set numeric targets has closed its programme. DARPA’s Next-Generation Nonsurgical Neurotechnology aimed at “high-performance, bi-directional brain-machine interfaces for able-bodied service members” with 16 channels in 16 mm³ at 50 milliseconds, funded six teams from 2019, and is now listed as complete. No successor programme with comparable targets is named on the same page.

Established The neuroethics literature has moved faster than the engineering, which is unusual and worth noting. Ienca and Andorno proposed four neurorights — cognitive liberty, mental privacy, mental integrity and psychological continuity — in 2017, with mental integrity defined around unauthorised alteration of neural computation causing harm. That framework was written for a technology that transmits one bit every eight seconds.

Established The ethics literature on brain-computer interfaces does not cover this case at all. Burwell, Sample and Racine’s scoping review coded 42 articles, with user safety in 24, justice in 20, and privacy and security in 19, and only 11 of the coded articles were empirical. The review contains no discussion of brain-to-brain interfaces. The two-person case is a gap in the field that is supposed to have thought about it.

Frontier There is no buyer, and that is the most reliable predictor of the next decade. The clinical market is served by devices that decode to machines; the consumer market is served by speech; the military interest was framed around machine control. A technology with no indication, no payer and no consumer proposition advances at the rate of curiosity-driven grants, which is roughly the rate the record shows.

Speculative Neural-data protection law now being written will govern this case by accident. Instruments drafted around data from a single wearer will apply to a link in which one person’s decoded intention becomes another’s stimulation. Whose data the stimulation record is — sender’s, receiver’s or operator’s — is settled nowhere this brief could find; Cognitive Liberty carries the rights argument.

Established The distinctive ethical fact is that the receiving party is acted on, not merely read. In the 2014 protocol the receiver’s hand moved involuntarily; in BrainNet the receiver saw a flash they did not generate. Trimper and colleagues put the resulting problem crisply: information is “introduced to a receiving brain, presumably without the ability of that brain to refuse or inhibit the impulse”.

Frontier Authorship of the resulting action is genuinely ambiguous, and this is not a hypothetical. Hildt’s 2015 analysis notes that a recipient may “erroneously ascribe authorship to themselves” for a movement the link caused. In the published protocols the receiver knows exactly what is happening, so the ambiguity is contained. It is contained by the experimental design, not by anything in the technology.

Frontier Responsibility has no owner once two people share an action. Trimper, Wolpe and Rommelfanger pose the military version directly: if a monitoring soldier transmits a threat assessment to a soldier on the ground who then fires, “who would be responsible for the soldier or civilian deaths?” At one bit per eight seconds the question is academic. It stops being academic at the first rate that makes the system operationally useful, which is the point of raising it now.

Speculative Autonomy in a network is a different problem from autonomy in a pair, and BrainNet is the first data point. Hildt argues that recipients in larger networks face difficulty maintaining autonomous decision-making when leaning on multiple senders of uncertain reliability, raising what she calls the risk of “fake news or brain hacking”. BrainNet’s own noise condition showed receivers learning to discount an unreliable sender, which is evidence in the opposite direction and should be counted.

Established Consent has to cover both roles and the literature has noticed. Hildt argues for consent forms addressing technical detail alongside privacy, agency and identity, while conceding this “may be difficult to achieve”. The sender consents to having intentions decoded; the receiver consents to cortical stimulation on someone else’s schedule. Those are different consents, treated as one.

Speculative Mental privacy is invoked constantly here and applies less than it appears to. The decoded quantity in every human study is trained motor imagery or a gaze fixation on a flickering square: a voluntary act with a public correlate, closer to pressing a button than to reading a mind. The real risk would be a system reading a signal the sender did not intend to send, and no study has shown one.

Frontier Coercion is the risk that scales worst with improvement. A channel that writes into a brain is also a channel that writes into a brain without agreement. Ienca and Andorno’s right to mental integrity is defined precisely around direct access to neural signalling without informed consent producing harm. The current technology cannot do this — a coil must be positioned on a cooperative head — and the wearable versions the engineering section describes would weaken that protection.

Handwave The claim that networked brains would dissolve individual identity is the field’s most-repeated ethical worry and it has no purchase on anything built. Trimper and colleagues ask “how is individuality defined when a brain is synched with another’s, or perhaps many others’?” Three people passing one bit every eighteen seconds have not merged. Raising a dissolution worry about an 81.25 per cent binary classifier misallocates the field’s limited ethical attention, which would be better spent on consent asymmetry between sender and receiver.

Established On the measured record this technology changes nothing at civilizational scale, and saying so is the honest finding. Seven human-to-human studies, a maximum published payload of one bit per iteration, and a throughput three orders below speech do not constitute a communications technology. Any civilizational claim has to be conditioned on a rate nobody has achieved.

Speculative If the rate did reach speech, the first thing it would change is not communication but verification. A channel carrying a percept rather than a report removes the sender’s option to misdescribe. Courts, medicine, journalism and contracts are all built around the gap between what a person experienced and what they say; narrowing it is an epistemic instrument before it is a convenience.

Speculative The same property is why the first serious deployment would be coercive if it were ever cheap. The value of an unfalsifiable channel accrues to whoever can compel its use: employers, security services, courts. The sender’s protection today is that the signal must be deliberately generated through trained motor imagery, and a system that removed that requirement would remove the protection with it.

Handwave The claim that linked brains would compute better than separate ones has one piece of evidence and it points the other way. A collaborative brain-computer interface configuration needed at least seven people to match the average single observer. The monkey Brainet improved because a shared decoder pooled neurons, which is an ensembling result available without any inter-animal link. Nothing measured supports a superlinear return to connecting brains.

Frontier The distributional question is unusually easy to answer here, and the answer is that almost nobody gains. For people with no motor output the comparison class is zero, so even a poor channel has value; that population is small and already the target of better-funded brain-computer interface work. For everyone else the channel loses to speech by a factor of about 780 and to a keypress by about 100. There is no wide constituency for this technology at its current performance.

Speculative Language is the incumbent and it is very good. Across 17 languages with sharply different syllable rates, information rates converge near 39 bits per second at 5 to 8 bits per syllable — a system that trades speed against density to land in the same place everywhere. The one honest argument for displacing it is content it cannot carry.

Frontier The most likely civilizational effect of this subject over the next decade is regulatory rather than technical. Neurorights frameworks, neural-data statutes and device rules are being drafted now, partly in response to demonstrations like BrainNet, and will bind whatever arrives later. A field with 80 published trials is shaping law that will govern technologies with different physics.

These horizons track the one quantity that decides this subject: bits per second across a direct brain-to-brain link, against the 39 bits per second of speech and the 0.05 bits per second the field has published.

• 10 yr: Frontier The head-to-head trial against speech and a keypress is run by somebody, because it costs a protocol change rather than an instrument. A duty-cycle safety study for repeated non-invasive stimulation of healthy volunteers is the other cheap, decisive item. Expect the published human rate to remain below 1 bit per second unless the write mechanism changes; expect Brain-Computer Interfaces to continue setting every communication-rate record in neurotechnology.
• 25 yr: Speculative If focused ultrasound or a successor write method is characterised for repetition rate and reaches an alphabet larger than two symbols, rates in the 1–10 bits per second range become conceivable, which is the band the mouse optical link already claims. That is still below speech and within reach of a keyboard, so the use case would remain people with no motor output.
• 50 yr: Speculative A channel competitive with speech requires the code-transfer question to have been answered positively — that patterns derived from one cortex mean something in another without a lookup table. Nothing currently under way tests this in humans, so the horizon is set by whether anyone starts, not by how long the work takes.
• 100 / 250+ yr: Handwave Networks of many brains exchanging non-verbal content at conversational rates, with the attendant collapse of testimony as an institution, is a coherent picture resting on at least three results nobody has produced. It is listed because the field invokes it, not because anything in the record points there.

• Depends on every read-side advance in Brain-Computer Interfaces, whose decoders and electrode counts this field inherits without contributing to, and on the device-longevity and stimulation record in Neural Interfaces. Nothing else on this map blocks it: the limiting quantities are a stimulation duty cycle and an unanswered question about shared neural codes, and neither is a result another brief is waiting to produce.
• Requires (not on this map) a matched-condition experiment in which the same pair does the same task through the link, through speech and through a keypress, which no published study has run; a non-invasive write method that can repeat on millisecond rather than multi-second timescales, since the 0.05 bits per second ceiling is a duty-cycle number; a demonstration in humans that stimulation patterns derived from one cortex evoke matching percepts in another, without which every system needs a learned translation layer; a channel capacity in the Shannon sense for one fixed read-write pairing, which the founding 2013 paper asked for and nobody has supplied; a characterised safety envelope for repeated stimulation of healthy people, because an ethics committee currently sets the field’s headline number; and someone willing to pay for a link with no clinical indication and no consumer proposition.
• Enables nothing at present rates that speech does not enable more cheaply. Conditional on the code-transfer result, it would enable the transfer of content that resists verbalisation — percepts, motor timing, subjective magnitudes — which is the only capability in this brief that language does not already cover.
• Adjacent to Distributed Cognition and Collective Intelligence, which own the question of whether joined minds compute better; to Memory Engineering, which owns the harder payload; to Interspecies Communication, which owns the cross-species case these systems perform best in; and to Cognitive Liberty, which owns the rights framework being drafted around it.

Handwave “Scientists have achieved telepathy.” No part of any published experiment transmits a thought. Grau and colleagues transmitted “hola” after a human being converted it to binary and the receiver counted phosphene flashes against a lookup table agreed beforehand. The semantic content travelled by ordinary agreement; only the selection travelled over the link. This is a telegraph, and a telegraph with an error rate reported at 15 per cent for a handful of characters.

Established “BrainNet let three people play Tetris with their minds.” It let two people vote on one binary question and a third person act on the vote. Senders stared at a 17 Hz or 15 Hz flicker to signal rotate or do not rotate; the receiver saw a phosphene or did not; mean accuracy was 81.25 per cent over 16 trials, in five triads, for a total of 80 trials. The authors themselves write that the payload “has remained binary”.

Frontier “Brain-to-brain interfaces are a kind of brain-computer interface, so the same progress curve applies.” The curves are different because the write stage is different. A brain-computer interface improves when decoding improves; its output is a machine action with no safety duty cycle. A brain-to-brain link improves only when the stimulation side improves, and the stimulation side is limited by what may be done repeatedly to a healthy person. That is why spellers reach 360 bits per minute while the direct channel sits at 0.05 bits per second.

Frontier “Inter-brain synchrony shows brains are already communicating directly.” Synchrony is a correlation, and the field knows it. Across 64 reviewed synchrony studies, 52 acknowledged the risk of spurious synchronisation, and only 72 per cent ran the shuffled-pair control that distinguishes genuine coupling from two people being exposed to the same stimulus. Correlated activity in two people watching one film is a fact about the film.

Speculative “It is only a matter of bandwidth.” Bandwidth is necessary and nothing shows it is sufficient. A channel delivering an arbitrary pattern to sixteen cortical sites every 50 milliseconds would still face the question of what the recipient experiences. The code-transfer experiment in 7 has never been run in humans, so that assumption does unexamined work in every roadmap resting on it.

Established “The 2013 Duke experiment showed a rat transmitting knowledge to another rat.” It showed decoder rats performing slightly above chance. 64.32 ± 1.1 per cent on the motor task and 62.34 ± 0.59 per cent on the tactile task are real, significant, and roughly one animal in three still choosing wrongly on a two-alternative task. The result matters because of what crossed the link, not because of how well it crossed.

Established “Monkeys have been networked into a collective brain.” Four macaques shared a decoder; nothing passed between them. The Brainet partitioned control of a virtual arm and improved from 20 per cent to 78 per cent correct over training in the three-animal case. The improvement came from mixing contributions and from total neuronal mass, which is what more electrodes in one animal would also provide.

Frontier “The military already has this.” The programme that set the targets has closed, and its targets were about machines. DARPA’s Next-Generation Nonsurgical Neurotechnology aimed at bidirectional brain-machine interfaces for able-bodied service members controlling unmanned systems. Person-to-person links were not its objective.

Speculative “The channel is millions of times slower than speech.” As a rate comparison this is wrong by about three orders of magnitude. 39 bits per second against 0.05 is a factor of 780. The larger number appears when you count distinguishable messages instead of bits, where eighteen seconds of speech selects among 2 to the power of 700 alternatives and one BrainNet round selects among two — a gap far beyond millions. The correct criticism of this technology is not that it is unimaginably slow but that it is ordinarily slow and has no compensating capability yet.

Frontier “Non-invasive means safe, so rate limits are temporary.” The rate limit is the safety margin. If repeated transcranial magnetic stimulation at shorter intervals were shown acceptable in healthy volunteers, the published ceiling would rise immediately with no new hardware. That study has not been done for this application, which makes the ceiling an institutional artefact that could move in either direction.

Handwave “Connecting more brains will produce emergent group intelligence.” The only quantitative result on multi-user configurations found the opposite direction. A collaborative brain-computer interface needed at least seven individuals to match the behavioural performance of the average single observer. An emergence claim requires a prediction that could fail, and the literature has not offered one.

Established “This is a fast-moving field.” The only systematic review found fifteen studies in seven years. Nam and colleagues screened 193 records and included 15, spanning 2013 to August 2020, of which seven were human to human; the most-cited is from 2019, and an April 2026 clinical review of brain-computer interfaces does not mention brain-to-brain interfaces at all. A small field that produced several memorable demonstrations and then largely stopped.