Reading the Mind: The Brain-Computer Interface

Cochlear implants have restored hearing function to more than a million recipients worldwide. The mechanism is worth stating precisely, because it is the mechanism the whole field rests on: the device bypasses damaged hair cells and stimulates the auditory nerve directly, which is the proof that a machine can interface with the nervous system and be useful. The record carries dates. The first crude implant was placed in 1961 by Dr. William House and John Doyle in California. The FDA approved the first commercial device, the House and 3M single-channel implant, in 1984, the first device ever approved to replace a human sense. That approval is now more than forty years old, which is rarely where coverage of this subject starts.

Deep brain stimulation places electrodes in the basal ganglia and runs current through them to treat movement and psychiatric disorders, and its approval history has to be stated in two parts rather than one, because our own research file compresses them into an error. The FDA approved Medtronic's Activa Tremor Control System in July 1997 for essential tremor. The Parkinson's-disease-specific indication, bilateral stimulation of the globus pallidus or the subthalamic nucleus, followed in January 2002. Our own file dates the Parkinson's approval to 1997, which attaches a real year to the wrong indication, and this article states the corrected version because it was independently verified rather than because the file says so. Deep brain stimulation is also FDA-approved for dystonia, obsessive-compulsive disorder and epilepsy. Studies report an average symptom-severity reduction of more than 50 percent for Parkinson's patients, with medication dosage reductions of up to 80 percent in some cases. One figure that travels with those does not get stated here as a fact: the commonly repeated count of roughly 200,000 patients worldwide is a widely repeated industry estimate that this research pass could not pin to any single authoritative dataset, so it appears here as an estimate and nothing firmer.

Vision has been harder, and the one approved answer has already been through an entire commercial life cycle. The Argus II retinal prosthesis, made by Second Sight Medical Products, received FDA approval on February 14, 2013, through the Humanitarian Device Exemption pathway rather than a full premarket approval, as the first implanted device to treat adult patients with advanced retinitis pigmentosa. A 60-electrode array stimulates the retina, producing partial vision at roughly the equivalent of 20 pixels of resolution, on the evidence of a single-arm trial of 30 patients followed for as long as 38.3 months. More than 350 patients worldwide received the implant. Then Second Sight discontinued commercial operations for new patients in 2019 and wound down its broader operations in 2020, citing a small patient population and revenue shortfalls rather than any safety problem, and the successor company Cortigent took over the device's rights and support in 2023. Read that sequence carefully, because it is a failure mode this subject rarely discusses. The device worked, the approval stands, and the business underneath it did not survive.

Vagus nerve stimulation received FDA approval as an adjunctive therapy for drug-resistant epilepsy in 1997, and for treatment-resistant depression in 2005. Since that 1997 approval it has been implanted in more than 100,000 patients for epilepsy, including more than 30,000 children. Note what the epilepsy indication actually says, because the wording is the claim: adjunctive, and drug-resistant. It is an addition to a treatment that was not working, offered to people for whom the standard answer had already failed.

One approved device does something the four above do not: it listens before it acts. NeuroPace's RNS System, FDA-approved in November 2013, is closed-loop and bidirectional. It continuously monitors electrocorticographic activity through implanted leads, detects the abnormal electrical patterns that precede a seizure, and delivers brief stimulation to normalize activity before the seizure occurs, which makes it the first commercially available device to provide closed-loop responsive brain stimulation. Its approval rested on a 65-patient feasibility trial, a 191-patient double-blind randomized controlled trial, and a subsequent 230-patient open-label long-term trial running seven years. Our own research file carries this at Tier 2 rather than Tier 1, and this article keeps that grading rather than promoting it on the strength of an approval date. The approval is a fact. The tier is our file's judgment of the evidence, and quietly upgrading it here is exactly the move this site exists to refuse.

FDA-approved and FDA-cleared are commercial-marketing statuses. They mean a device is legally sellable, which is what every device in section 01 is. An Investigational Device Exemption is not that. It is permission to run a clinical trial, and a device operating under one is legal to study in enrolled participants and not legal to sell. Breakthrough Device Designation is a third thing again, an expedited-review status, and it is neither an approval nor a trial permission. Synchron's regulatory history, in section 05, carries the second and third of those and not the first, and compressing them together is exactly how the research document behind this article ended up with the wrong chronology. Every device in the next three sections is investigational. Every one of their results below is real, published and peer-reviewed. Neither of those sentences cancels the other, and this file will not let either one be used to imply the other.

In 2012, Cathy Hutchinson, paralyzed by a brainstem stroke, reached out with a robotic arm, picked up a bottle of coffee and drank from it without help. She did it by thinking about it. The system was BrainGate: a 96-electrode Utah array implanted in her motor cortex, decoding the intention to move and passing it to the arm. It was published by Hochberg and colleagues in Nature on May 16, 2012, as Reach and Grasp by People with Tetraplegia Using a Neurally Controlled Robotic Arm. One number needs correcting on the way past. Our own research file says it was the first time she had done that in 14 years, while the contemporary press coverage of the paper itself, including Nature's own news report, describes it as nearly 15 years since her stroke. This article states roughly 15 years, on the primary coverage, and flags the discrepancy rather than passing it along.

Nine years later the same programme turned intention into text. BrainGate's 2021 Nature paper, by Willett, Avansino, Hochberg, Henderson and Shenoy, had a participant with a cervical spinal cord injury imagine handwriting individual letters while a Utah array in motor cortex recorded the attempt. Decoded, that produced typing at 90 characters per minute with 94.1 percent accuracy in real time, rising above 99 percent offline once a general-purpose autocorrect was applied to the output. That was more than double the previous brain-computer interface typing speed record. The result is worth reading twice for what it does not say. Nothing was read out of the participant's mind. What was decoded was an attempted movement of the hand, recorded at the place in cortex where hand movements are planned, and the fact that letters came out the other end is a property of what he was attempting, not of the machine's access to his thoughts.

One question decides whether any of this is practical: how long the hardware survives inside a living brain. In 2025 it got an answer covering 20 years. A long-term safety and durability analysis of the BrainGate and BrainGate2 pilot trials examined 2,319 recording sessions across 20 Utah arrays in the first 14 participants, enrolled between 2004 and 2021, with an average implant duration of 872 days. Across that record the arrays recorded neural spiking on an average of 35.6 percent of their electrodes, with only a 7 percent decline over as much as 7.6 years. Two caveats travel with those figures and both belong in the same paragraph. The first is that these are averages across arrays rather than a guarantee for any one of them. The second is provenance: this circulated as a 2025 preprint with a PubMed Central-indexed companion, so it should be read as a recent study rather than as decades-old settled consensus, and this file states it that way rather than dressing it up.

Neuralink's N1 implant carries 1,024 electrodes distributed across 64 ultra-thin polymer threads, each thread about 5 micrometers thick, thinner than a human hair, placed by a purpose-built surgical robot rather than by hand. The first human recipient was Noland Arbaugh, paralyzed by a diving accident in 2016, implanted on January 28, 2024 at the Barrow Neurological Institute in Phoenix as part of the PRIME Study. Within weeks he was moving a computer cursor by thought and playing online chess, Civilization VI and Mario Kart on a Nintendo Switch, for the first time since his injury. That is the achievement, and this file does not shrink it. It is also a clinical trial under an Investigational Device Exemption, and the next paragraph is why that distinction is not bookkeeping.

About three months after the surgery, roughly 85 percent of the N1's threads had retracted from Arbaugh's brain tissue. The likeliest explanation offered was that his brain moved inside the skull more than Neuralink had anticipated. The consequence was immediate: sharply fewer electrodes able to record a signal, and slower cursor control. Neuralink's response was to modify its decoding algorithms rather than remove the implant, and the device stayed in. Arbaugh went on using it 8 to 10 hours a day, and at 18 months after surgery was still using it, still playing games including Mario Kart. Both halves of that belong on the record. A serious hardware problem appeared in the very first patient, and the recovery from it was real.

The second PRIME Study participant, publicly known as Alex, was implanted in July 2024 with a revised architecture: 128 thinner threads carrying 8 electrodes each, the same 1,024 electrodes in total, engineered specifically to reduce displacement of brain tissue. Alex experienced zero thread retraction. He was discharged the day after surgery, was controlling a cursor within minutes of his first session, with that first day's cursor speed reported as passing the previous record held by a brain-computer interface outside Neuralink, and within two days was using Fusion 360, a computer-aided design package, to design a mount for his own charger. One patient's hardware failure, one engineering change, and a second patient without it. That is what iteration looks like when it works, and it is also a good reason not to read any single patient's result as a settled capability.

The current count is newer than the snapshot our own research file carries, and it is stated here because it was independently verified rather than because the file says so. As of January 28, 2026, two years to the day after Arbaugh's surgery, 21 people across four countries had received an N1 implant: the United States, Canada, Great Britain and the United Arab Emirates, across the PRIME, CAN-PRIME, UK-PRIME and UAE-PRIME studies, with no serious device-related adverse events reported to date. Our own file still describes a first patient and a second patient and stops there. Twenty-one people is a much larger number and a very small one at the same time, and the studies they are enrolled in are feasibility studies. Not one of those 21 people is a customer, because there is nothing to buy.

Everything above is documented clinical work. What follows is not, and the two must never be read in the same register. Neuralink has publicly stated ambitions well beyond cursor control, including restoring vision through an announced programme it calls Blindsight, and eventually restoring speech. Those are the company's own stated roadmap items. They are not achieved results, they are not independently verified, and nothing in the PRIME Study record above establishes them. This pipeline's own tier discipline grades company timeline promises at Tier 2 to 3; this article takes the more cautious end of that band and files them at Tier 3, attributed to Neuralink by name, which is the only honest place for a corporate intention the record has not yet met. If Blindsight delivers, the evidence will look like the paragraphs above it: a named patient, a date, a published outcome, and a complication reported alongside it.

Synchron's Stentrode is a mesh electrode array delivered by catheter through the jugular vein and pushed into a blood vessel that runs alongside the motor cortex, the superior sagittal sinus. No open-brain surgery is involved. What happened next matters in its exact order, because our own research file compresses it into one line, "FDA approved for clinical trial (2020). First US patient: 2022," and that line is wrong twice over. The device's first human implants were not in the United States at all. Four patients received one in Australia's SWITCH trial at the Royal Melbourne Hospital, with enrollment beginning in 2019, and the results, published by Oxley and colleagues in the Journal of NeuroInterventional Surgery in 2021, showed patients texting, shopping online and managing their finances from home using thought alone. The FDA's action in August 2020 was a Breakthrough Device Designation, an expedited-review status rather than a trial approval. The actual Investigational Device Exemption for the US COMMAND trial came in July 2021, and Synchron describes that as the first IDE the FDA has awarded for a permanently implanted brain-computer interface. The first US patient was implanted in 2022, at trial sites including Mount Sinai and the University of Pittsburgh. Three separate events, three separate years, and our own file had them as one date and a country that was not first.

The COMMAND early feasibility study, which is newer than our own file, evaluated the Stentrode in six patients with severe paralysis against a 12-month primary endpoint. No serious adverse events were reported. The device reliably captured brain signals and converted them into what Synchron calls digital motor outputs, letting patients control a mouse cursor, an Apple Vision Pro headset and Amazon Alexa hands-free. In November 2025 the company raised $200 million in Series D funding to run a pivotal trial in 2026, which is the step it has to clear before it can seek premarket approval, the status that would make the device legally sellable. Which fixes the position exactly, as of this research pass in July 2026: Synchron's device is investigational. It is not FDA-approved for market, and the pivotal trial that would open that door has not concluded.

In 2021, Flesher, Downey, Weiss and colleagues reported in Science that pressure signals from the sensors of a robotic hand had been fed back into a paralyzed patient's somatosensory cortex, and that the patient could feel the hand's grip strength in real time. That is the loop closing. The brain sends an intention outward, and the world sends a sensation back in through an electrode, which is a different and harder claim than moving a cursor, because it requires the machine to write something the brain will accept as sensation rather than merely read something the brain was already doing.

The same year, Moses, Metzger, Liu and colleagues reported in the New England Journal of Medicine that a neuroprosthesis had decoded attempted speech from a paralyzed man with anarthria, the loss of the physical ability to articulate, directly into text and a synthesized voice. He had no functional way to speak, and the device gave him a channel. This is a result our own research file lists in its bibliography and never narrates in its own text, which is a strange omission in a document about brain-computer interfaces, so it is stated here in full rather than left as a reference nobody reads. Neither this nor the touch result above appears on the approved-device list in section 01, and this file does not put them there. They are published clinical research, and that is the honest description of them.

In 2023, Tang, LeBel, Jain and Huth reported in Nature Neuroscience that continuous language a subject had heard or was imagining could be reconstructed from blood-oxygen-level-dependent fMRI signal alone. What came out was the semantic gist, not a word-for-word transcript. A related approach, from Takagi and Nishimoto at CVPR the same year, used diffusion models to reconstruct images a person was viewing from fMRI data, at low but recognizable resolution. And Meta's magnetoencephalography-based system, reported by Defossez and colleagues in 2023, demonstrated real-time, word-by-word decoding, using millisecond-resolution equipment that fills a room. Then the number that governs all three. An fMRI voxel is about 2 millimeters on a side and contains roughly 500,000 neurons. An implanted array reads individual neurons. The non-invasive methods are working from a signal averaged over half a million cells at a time, and every capability and every limit in this paragraph follows from that one fact.

Non-invasive EEG headsets sold to the public, such as those from Emotiv and OpenBCI, can detect basic mental states and drive simple interfaces. Their spatial and signal resolution is far below implanted devices, and their reliability in real-world conditions outside laboratory settings is limited. Our own file carries this at Tier 2, and it is a useful correction to any assumption that a headset somebody can buy today is a small version of the array in section 03.

One public programme aimed squarely at writing to a brain was run by a defence agency. DARPA's SUBNETS programme, roughly $70 million between 2014 and 2019, pursued closed-loop treatment for mental illness: detect the neural biomarkers of depression or anxiety, and stimulate in real time to normalize them. The model is the brain pacemaker of section 01 pointed at psychiatric rather than movement disorders. Our own file carries it at Tier 2, and the honest reason for that grading is that its clinical outcomes are considerably less publicly documented than the approved devices above.

Memory is the other target, and there are two named results. DARPA's Restoring Active Memory programme, running from 2017 to 2020, tested a targeted memory prosthesis for traumatic brain injury: intracranial electrodes identified the neural patterns associated with encoding a memory, and stimulation was applied to strengthen the encoding, with reported improvements of 15 to 25 percent on word-recall tasks. Separately, a hippocampal prosthesis developed at the University of Southern California, reported by Hampson, Song, Robinson and colleagues in the Journal of Neural Engineering in 2018, recorded the neural patterns produced during learning and played them back to strengthen consolidation, with a first human trial showing a 37 percent improvement in short-term memory. Both are real, published, and modest. A percentage gain on a word-recall task in a first trial is a long way from a memory prosthesis anybody could be offered.

Then the result this file will not allow to become a capability claim. Optogenetic techniques have selectively erased specific memories in mice, reported by Bhatt and colleagues in 2020, using light-activated proteins to deactivate the synapses that encoded one particular memory. In mice. This has not been tested in humans, and nothing in this article should be read as suggesting that it is close or imminent. What the result does earn is a question asked early rather than late: if a specific memory could eventually be removed, who decides which ones, and for what conditions. Post-traumatic stress disorder is the compelling case, and it is genuinely compelling. A memory that is politically inconvenient, or merely inconvenient to somebody with authority over the person carrying it, is not. This file poses that question and does not answer it, because the answer is not a scientific finding, and pretending it is one would be the dishonest move here.

Two brains have been connected directly, and the honest measure of that achievement is how much passed between them. BrainNet, reported by Jiang, Stocco, Losey and colleagues in Scientific Reports in 2019, let three people collaborate on a Tetris-like game using brain signals alone: two senders wearing EEG caps transmitted whether a block should be rotated, and one receiver took that decision in through transcranial magnetic stimulation, at roughly 81 percent accuracy. Earlier, in 2014, Grau and colleagues sent single words, hola and ciao, brain to brain between India and France over the internet, using the same EEG-to-TMS pairing. Both are genuine proof-of-concept demonstrations of direct brain-to-brain information transfer. Both run at bits per minute. That is closer to Morse code than to telepathy, and the distance between the demonstration and the fantasy is the thing worth carrying out of this section.

A short aside before the closing section, labeled as one and kept deliberately brief, because the deep version of this argument belongs to a different file. Ray Kurzweil has predicted, in the 2029 to 2035 window, nanoscale wireless sensors distributed throughout the brain, neural dust, connecting every neuron to the cloud and merging human cognition with machine intelligence. There is one real proof of concept underneath that idea, and it is smaller than the idea by a wide margin. Seo and colleagues at UC Berkeley published ultrasonic-powered, sub-millimeter wireless neural sensors in Neuron in 2016, demonstrated in the peripheral nerves of rodents. Not the cortex. Not humans. Getting from that experiment to millions of cortical sensors with microsecond latency and wireless bandwidth far exceeding current WiFi is, in our own research file's own words, decades away at minimum, and may require breakthroughs nobody currently has.

The two ideas that follow from it get this paragraph and no more. The first is progressive consciousness migration: the hypothesis that a mind could be moved to a digital substrate gradually, neuron by neuron, as artificial equivalents replace biological ones, preserving continuity of experience at every step and so avoiding the copy problem that instantaneous whole-brain scanning creates. Whether that would preserve a person depends entirely on whether consciousness is substrate-independent, which is an open philosophical question rather than a technical one, and if consciousness turns out to require specific biological processes then silicon replacement would not carry it across. The second is collective consciousness: if brain-to-brain bandwidth ever scaled far past the bits per minute above, networked minds could in principle share thought or experience directly, with the risks worth stating as plainly as the promise, loss of individual identity, vulnerability to manipulation at the level of the network, and extraordinary power for whoever operates it. No current technology is remotely close to either. Both belong to our research file S_1_02, The Singularity and Transhumanism, linked in the sources below, and to the separate article that carries that subject in this wing. This file records that the ideas exist, records that they are genuinely unresolved, and stops there.

"We can read minds today" is misleading. What current systems decode are statistical patterns: an intended movement, read from motor cortex, or a broad semantic category, read from fMRI signal. The nearest thing in this whole article to reading language off a brain is the fMRI work in section 07, and what it recovers is the gist of language a subject heard or was deliberately imagining, not a transcript, from voxels holding roughly 500,000 neurons each. No system in this file reads thoughts in any colloquial sense. They cannot access memories. They cannot access an internal monologue. They cannot access subjective experience. That is our own file's Tier 4 statement, and every genuinely impressive result in the nine sections above sits underneath it.

There is no credible evidence of a deployed government mind-control brain-computer interface programme. The pieces people assemble into that claim are individually real, which is what gives it its grip. MKUltra was a genuine CIA programme, running from 1953 to 1973. DARPA genuinely funds brain-computer interface research, and two of its programmes are named in section 08 of this article. What does not follow is the conclusion. Current brain-computer interface resolution is nowhere near what anything resembling mind control, as popularly imagined, would require. The obstacle is not secrecy. It is resolution, and this entire article is a measurement of it.

Claims that 5G networks, injectable microchips or vaccines function as covert brain-computer interfaces are false, and they are false in a way that can be answered from physics rather than from trust. A brain-computer interface requires physical contact with neural tissue or very close proximity to it. Every device in this file is in exactly that kind of contact: implanted in the brain, threaded into a vessel beside it, implanted to stimulate the retina, wired to a nerve, worn on the scalp, or housed in room-scale equipment the person has to be inside. Electromagnetic signals attenuate, and that attenuation rules out a remote injectable device that reads a brain, with current or foreseeable technology. Everything that works in this article had to be implanted, threaded, placed, wired or worn, and that requirement is not an incidental detail. It is the answer to the claim.