A brain-computer interface (BCI), also called a brain-machine interface, is a system that measures activity in the nervous system and translates it into commands for an external device, bypassing muscles and peripheral nerves entirely. The canonical definition comes from a 2002 review by Jonathan Wolpaw and colleagues, which described a BCI as a non-muscular channel for communication and control built from electrophysiological measures of brain function [1]. Every BCI has the same four stages: a sensor that acquires neural signals, a processing step that extracts features, a decoder that maps those features onto an intended output such as a cursor movement or a word, and feedback that lets the user adapt.

Nearly all clinical BCI work targets people who have lost movement or speech: spinal cord injury, amyotrophic lateral sclerosis (ALS), brainstem stroke, and related conditions [1]. The most striking recent results come less from new sensors than from new decoders. The 2024 UC Davis speech system, which held 97.5% accuracy over more than eight months of use, read from intracortical silicon microelectrode arrays, the same recording technology behind the first human intracortical BCI demonstration in 2006 [6][11]. What changed was the software: deep learning sequence models paired with language model priors pushed decoded speech to 62 words per minute by 2023, which the authors of that study noted begins to approach the roughly 160 words per minute of natural conversation [9].

The scale of the change shows up in information transfer rate. The 2002 review put the state of the art at 10 to 25 bits per minute [1]. In January 2026, Neuralink reported that one participant exceeded 10 bits per second within his first week of use, matching or beating the 8 to 10 bits per second the company measures for able-bodied people using a mouse [2]. That gain is real but narrowly held: it belongs to a small population of implanted research participants worldwide, and every chronically implanted BCI covered here is still an investigational device.

History

The term originates in a 1973 review by Jacques Vidal, then a computer science professor at UCLA, titled "Toward direct brain-computer communication" [3]; UCLA credits that paper with coining the phrase "brain-computer interface" and setting the foundation for the field [44]. Non-invasive electroencephalography (EEG) based systems dominated the next three decades: slow cortical potentials, P300 evoked responses, and sensorimotor mu and beta rhythms were all used to spell text or move a cursor, at rates that made a short sentence a multi-minute exercise [1].

The invasive branch depended on a specific piece of hardware. In 1991, Patrick Campbell, Richard Normann, and colleagues at the University of Utah published a manufacturing process for a silicon microelectrode array: a 4.2 by 4.2 by 0.12 mm substrate with 100 conductive needles, each about 1.5 mm long and 0.09 mm thick at the base, with platinum-coated tips [4]. The Utah array, as it became known, is still the workhorse of intracortical BCI research [39]. It reached the clinic through Blackrock Microsystems (now Blackrock Neurotech), whose NeuroPort microelectrode array system holds FDA 510(k) clearance for temporary (less than 30 days) recording and monitoring of brain electrical activity; a February 2011 clearance covered iridium oxide tipped 1.0 and 1.5 mm arrays [5]. The multi-year implants used in BCI trials run instead under investigational device exemptions.

The first demonstration that a human could control devices with intracortical signals came in 2006, when Leigh Hochberg, John Donoghue, and colleagues reported that a participant with tetraplegia used a 96-microelectrode array in primary motor cortex to move a cursor, open and close a prosthetic hand, and operate a robotic arm, three years after his spinal cord injury [6]. That study belonged to the BrainGate trial program, which has continued since and supplies most of the field's long-term reliability data [39].

Recording modalities

BCIs are usually sorted by how far the sensor sits from the neurons it is listening to. Signal quality and invasiveness trade off against each other almost monotonically.

Modality	Placement	Typical signal	Notes
EEG	Scalp electrodes	Summed cortical potentials	Non-invasive, low cost, blurred by skull and scalp; also the recording method used by consumer headbands [38][45]
MEG	External sensor array	Magnetic fields from cortical currents	Non-invasive and substantially better than EEG for text decoding, but not portable [36]
fMRI	Scanner	Hemodynamic response	Non-invasive, seconds-scale latency; used for offline semantic decoding rather than real-time control [37]
Endovascular (Stentrode)	Inside a cortical vein via catheter	Field potentials recorded from within a vein	No craniotomy; delivered through the jugular vein into the superior sagittal sinus [7]
Micro-electrocorticography	On the cortical surface, non-penetrating	Surface field potentials	Does not pierce tissue and is reversible; the basis of Precision Neuroscience's Layer 7 [21]
Intracortical arrays	1-1.5 mm into cortex	Single and multi-unit spiking, threshold crossings	Highest bandwidth; used by the handwriting BCI and the Stanford and UC Davis speech systems [9][10][11]

Surface and endovascular approaches have a practical advantage that is easy to underrate: they are reversible, and their surgical risk profile resembles procedures neurosurgeons already perform. Intracortical arrays give that up in exchange for access to individual neurons, which carry far more information per unit time than a summed field potential.

Signal decoding with machine learning

Neural decoding is a sequence modelling problem with an unusually hostile input. The signal is non-stationary, so decoders drift and periodically need recalibration; keeping one stable for more than six months without recalibration is reported as a result in its own right [35]. Ground truth is often missing, because a participant who cannot speak cannot produce the audio a model is meant to reproduce [14]. Training data is scarce, since each participant is effectively a separate dataset and collection is capped by how long they can work before tiring [12][17].

Current intracortical systems share a common shape. Neural features (binned threshold crossings and spike-band power per electrode) feed a recurrent neural network that emits a probability distribution over phonemes at each time step [9]. That output is aligned to text without frame-level labels, using connectionist temporal classification or a recurrent-transducer objective [9][13]. A language model then assembles phonemes into words: the 2023 Stanford system used a custom 125,000-word trigram model, and the 2024 UC Davis system applied two further open-source language models to turn the initial word sequence into the most likely English sentence [9][11].

Every stage contributes. In the 2021 handwriting BCI, an RNN decoding attempted pen strokes achieved 94.1% raw character accuracy online, rising above 99% offline once a general-purpose autocorrect was applied [10]. Vocabulary size is the other lever: the 2023 Stanford speech neuroprosthesis reported a 9.1% word error rate on a 50-word vocabulary and 23.8% on a 125,000-word one [9].

Calibration burden has fallen sharply. Early systems needed hours of supervised data per session; the 2024 UC Davis system reached 99.6% accuracy on a 50-word vocabulary after 30 minutes of calibration on the first day of use [11]. A 2026 BrainGate2 preprint went further, showing that pooling recordings from previous users (which the authors call cross-brain transfer) improves a new user's decoder when their own training data is under about 200 sentences, provided the model uses dataset-specific input layers [12]. That is transfer learning applied to a domain where every additional training example costs a disabled person real effort.

The other two active fronts are latency and expressivity. Accurate text is not the same as conversation, so a 2025 UCSF system used a deep recurrent transducer to synthesize speech continuously in 80 ms increments, producing audio while the participant was still forming the sentence [13]. A 2025 UC Davis system trained without ground-truth audio decoded paralinguistic features alongside phonemic content, letting the participant shape the intonation of the synthesized voice and produce melodic vocalizations [14].

Communication neuroprostheses

The clearest measure of progress is the sequence of intracortical communication results published since 2021. All involved single participants or very small cohorts, and all used machine learning decoders.

Year	Group	Approach	Headline result
2021	Stanford / BrainGate	Attempted handwriting, RNN decoder	90 characters per minute, 94.1% raw accuracy; comparable to smartphone typing in the participant's age group (115 cpm) [10]
2023	Stanford / BrainGate	Attempted speech, phoneme decoder	62 words per minute; 9.1% word error rate on 50 words, 23.8% on 125,000 words [9]
2023	UCSF (Chang lab)	High-density ECoG, text plus audio plus avatar	Median 78 words per minute, median 25% word error rate; personalized voice; decoders trained in under two weeks [15]
2024	UC Davis (BrainGate2)	256 intracortical electrodes in left ventral precentral gyrus	97.5% accuracy sustained over 8.4 months and 248 hours of use, at about 32 words per minute [11]
2025	UCSF / UC Berkeley	Streaming RNN-transducer voice synthesis	Continuous synthesis in 80 ms increments; generalized to single-unit and electromyography inputs [13]
2025	UC Davis	Voice synthesis without ground-truth audio	Intelligible speech with user-controlled intonation and singing [14]

A related 2025 result showed that the speech motor cortex, targeted for speech decoding, also supports cursor control and click at 2.90 bits per second with 40 seconds of calibration, meaning a single array placement can serve both functions [16].

Inner speech and mental privacy

A 2025 study in Cell examined whether speech BCIs can read imagined speech rather than attempted speech. Across four participants, the authors found that inner speech is robustly represented in motor cortex and that imagined sentences can be decoded in real time. The representation correlates strongly with attempted speech but is separated by what the authors describe as a motor-intent dimension. They also showed that some free-form inner speech could be decoded during recall and counting tasks, and then demonstrated safeguards designed to prevent a speech BCI from decoding private inner speech unintentionally [17].

The finding cuts both ways. Decoding imagined rather than attempted speech would remove the physical fatigue that currently limits how long participants can use these systems. It also confirms that the privacy concern raised about speech BCIs is technically grounded rather than hypothetical.

Clinical programs

A ClinicalTrials.gov search for "brain-computer interface" returned 94 recruiting studies in July 2026 [46]. The implanted commercial programs are concentrated in a handful of companies.

Company	Device	Approach	Status
Neuralink	N1 Implant, R1 surgical robot	1,024 electrodes on 64 flexible threads inserted into cortex; fully wireless, inductively charged [18]	21 participants enrolled in trials worldwide as of January 2026 [2]
Synchron	Stentrode	Stent-mounted electrodes delivered endovascularly to a cortical vein; no craniotomy [7]	SWITCH (Australia) completed; COMMAND (US) reached primary completion September 2024; new studies in Australia, Canada, and the US registered in 2026 [19][33]
Precision Neuroscience	Layer 7 Cortical Interface	1,024 thin-film surface electrodes over a postage-stamp area, inserted through a micro-slit, non-penetrating [8][21]	Layer 7-T cleared by FDA 510(k) on 30 March 2025 for temporary use of less than 30 days [22]
Paradromics	Connexus	421 microelectrodes reaching 1.5 mm below the cortical surface, with a chest transceiver and near-infrared optical link [23]	FDA investigational device exemption November 2025; first long-term implant 17 June 2026 [24]
Blackrock Neurotech	Utah array, NeuroPort	Research-grade intracortical arrays; NeuroPort cleared in 2011 for temporary recording [4][5]	The recording hardware behind the BrainGate trials, including the 2021 handwriting and 2023-2025 speech results [39]

Neuralink

Neuralink, founded by Elon Musk, performed its first human implantation in January 2024 in the PRIME early feasibility study [18][25]; Barrow Neurological Institute in Phoenix, one of the study's two United States sites, states that the first participant to receive the N1 implant did so there [25][47]. The company calls the resulting capability Telepathy [18]. Its first participant, Noland Arbaugh, experienced thread retraction that temporarily reduced his BCI performance; the company implemented surgical mitigations (reducing brain motion during surgery and narrowing the gap between implant and cortex) and reported no retraction in the second participant [26].

In February 2025 Neuralink reported three participants, more than 670 cumulative implant-days, more than 4,900 hours of use, and an average of 6.5 hours per day of independent use in the preceding month [27]. In January 2026 it reported 21 participants enrolled worldwide, higher signal quality in 18 of the 20 participants implanted after the retraction mitigations, and no serious device-related adverse events. The same update described plans to increase electrode count from 1,000 to 3,000, to explore inserting threads through the dura rather than removing it, and reported participants typing at up to 40 words per minute by imagining ten-finger keyboard movements [2].

Registered Neuralink studies now include PRIME in the United States, CAN-PRIME in Canada, UAE-PRIME at Cleveland Clinic Abu Dhabi, GB-PRIME in the United Kingdom, CONVOY for control of assistive devices, and VOICE, a communication-restoration study that began in October 2025 and targets conversational speed of 140 words per minute [2][25][28][48]. The company received FDA Breakthrough Device Designation for a visual prosthesis called Blindsight in September 2024 and for speech restoration in May 2025, and raised a $650 million Series E in June 2025 [29][30][31].

Synchron

Synchron's Stentrode takes the opposite bet: less signal, far less surgery. The device is delivered by catheter through the jugular vein into the superior sagittal sinus, where stent-mounted electrodes record field potentials from the underlying motor cortex [7][32]. The SWITCH first-in-human study in Australia, published in JAMA Neurology in 2023, reported no serious adverse events, no vessel occlusion, and no device migration in four analyzed patients, with signal bandwidth of 233 Hz that stayed stable across the study [7]. The US COMMAND study enrolled six participants at Buffalo, Mount Sinai, and UPMC and reached primary completion in September 2024 [19]. A 2025 preprint reported that motor-related modulation in the 30-200 Hz band remained discriminable across five participants over 12 months post-implant [32]. Synchron describes the implant procedure in its trials as taking about two hours with no open brain surgery, and most participants going home the next day [43].

Synchron has drawn heavily on outside AI infrastructure. Its own announcement index records work with NVIDIA Holoscan in January 2025, an in-house model it calls Chiral (described by the company as a cognitive AI brain foundation model) in March 2025, and integration with Apple devices through Apple's BCI human interface device protocol, demonstrated on an iPad in August 2025 [20]. The company raised a $200 million Series D in November 2025, and registered new studies in Canada, Australia, and the United States in 2026 [20][33].

Precision Neuroscience and Paradromics

Precision Neuroscience builds a thin-film surface array rather than a penetrating one. Its Layer 7 Cortical Interface carries 1,024 electrodes across an area roughly the size of a postage stamp, is inserted through a micro-slit incision, and is designed to conform to the cortical surface without piercing it and to be removed and replaced [8]. A 2025 paper in Nature Biomedical Engineering described the 1,024-channel array delivered without craniotomy in porcine models and cadavers, recording and stimulating from the same electrodes, and reported an intraoperative pilot in five neurosurgical patients that characterized how sensorimotor activity and speech are represented at the cortical surface [21]. The company's Layer 7-T received FDA 510(k) clearance under the cortical electrode classification on 30 March 2025, for temporary (less than 30 days) recording, monitoring, and stimulation at the brain surface rather than for chronic implantation [22].

Paradromics is pursuing the opposite end of the bandwidth spectrum. Its Connexus device places 421 microelectrodes 1.5 mm below the cortical surface and routes data through a chest-mounted transceiver over a near-infrared optical link. After an FDA investigational device exemption in November 2025, the company and the University of Michigan completed the first long-term implant on 17 June 2026, in a participant with motor neuron disease who will be followed for six years; the Connect-One study also runs at UC Davis and Massachusetts General Hospital [23][24].

Minimally invasive work in China

A Tsinghua University group with Bo Hong as senior author has pursued an epidural approach using only eight platinum-iridium electrodes placed over the sensorimotor cortex and powered wirelessly, in a system called NEO [35]. In a participant with complete spinal cord injury, nine months of home use produced an average grasp-detection F1 score of 0.91 and a 100% success rate on object transfer tests when paired with a wearable hand exoskeleton. The same participant showed measurable neurological recovery, including a 27-point gain on the Action Research Arm Test [34]. A 2025 preprint from the group reported two-dimensional cursor control at Fitts information transfer rates of 36.7 and 30.0 bits per minute on two tasks, with hit rates above 91%, recordings stable beyond 18 months, and the decoder holding performance for over six months without recalibration [35].

Non-invasive decoding

Non-invasive BCIs remain far behind implanted ones on communication rate, though deep learning has narrowed the gap. A model from Meta AI called Brain2Qwerty decoded overtly typed Spanish sentences from 35 healthy volunteers. In the peer-reviewed analysis published in 2026, the mean character error rate was 29% with magnetoencephalography and 65% with EEG; the best participant reached 18% with MEG [36]. Those figures are well short of usable daily communication, and the better of the two modalities requires a stationary MEG scanner, so the practical path to a wearable device is not obvious.

A separate line of work decodes meaning rather than motor intent. A 2023 fMRI decoder from the University of Texas at Austin reconstructed continuous language from cortical semantic representations, producing word sequences that recovered the gist of perceived speech, imagined speech, and silent video. The authors specifically tested and confirmed that decoding failed without the subject's cooperation, which they framed as a mental-privacy safeguard [37]. A 2025 review surveys the recording methods, decoding algorithms, and open-source toolchains in current non-invasive use [38].

Consumer EEG headbands belong to a different category. Products such as FocusCalm, sold by BrainCo, use scalp EEG for neurofeedback training around states like focus and calm rather than decoding a specific intended output, and they are not regulated as implanted devices are [45]. UNESCO notes that while medical uses of neurotechnology are strictly regulated, consumer devices such as connected headbands remain largely unregulated, and most recent neural-data rulemaking is aimed at that tier rather than at clinical trials [42].

Limitations

Electrode longevity is the least-discussed constraint. A 2025 analysis of 14 BrainGate participants across 2,319 sessions and 20 Utah arrays found that arrays recorded neural activity on about 35.6% of electrodes, declining only about 7% over enrollment periods of up to 7.6 years, and that useful movement decoding persisted in 11 of 14 arrays. The same analysis found that decoding performance scales logarithmically with electrode count, which sets expectations for how much a higher channel count actually buys [39].

Other constraints are structural. Every published headline result rests on one participant or a handful, so generalization claims are weak, and the participant pool is selected for particular conditions, ages, and anatomy. Decoders drift and need periodic recalibration [35]. Attempted speech is physically tiring, which caps how long a session can run [17]. No chronically implanted BCI described here is cleared for commercial sale in the United States: Paradromics labels Connexus an investigational device [23], Synchron states that its system is not approved for commercial use in any geography [43], and the one 510(k) clearance in this group, Precision's Layer 7-T, covers temporary use of less than 30 days rather than a permanent implant [22]. Neuralink's own January 2026 update acknowledged continuing variance in BCI performance correlated with anatomical differences such as intracranial spacing and with the participant's stage of ALS [2].

Regulation and neural data

Implanted BCIs in the United States proceed through the FDA's investigational device exemption pathway for clinical studies [24], with Breakthrough Device Designation available to accelerate review [29][30]. Components with established predicates, such as cortical electrode arrays, can clear through 510(k), though those clearances so far cover temporary rather than chronic use [5][22].

Neural data has attracted separate legislation. Colorado's HB24-1058, signed in April 2024, extended the state privacy act to biological data including neural data, defined as information generated by measuring the activity of an individual's central or peripheral nervous system that can be processed by or with the assistance of a device [40]. California's SB 1223, approved in September 2024, classified neural data as sensitive personal information under the state consumer privacy law [41]. In November 2025, UNESCO member states adopted a Recommendation on the Ethics of Neurotechnology in Samarkand, the first global standard for the field, which entered into force a week later and calls for explicit consent and transparency around neural data and advises against non-therapeutic use in children [42].

The asymmetry these statutes address is straightforward. A research participant in a clinical trial is covered by a trial protocol and institutional review; a consumer wearing an EEG headband is not, which is why UNESCO's recommendation singles out consumer neurotechnology and the need for explicit consent and full transparency [42].

See also

References

Wolpaw, J. R., Birbaumer, N., McFarland, D. J., Pfurtscheller, G., Vaughan, T. M. "Brain-computer interfaces for communication and control." Clinical Neurophysiology 113(6), June 2002, 767-791. https://pubmed.ncbi.nlm.nih.gov/12048038/
Neuralink. "Two Years of Telepathy." 28 January 2026. https://neuralink.com/updates/two-years-of-telepathy/
Vidal, J. J. "Toward direct brain-computer communication." Annual Review of Biophysics and Bioengineering 2, 1973, 157-180. https://pubmed.ncbi.nlm.nih.gov/4583653/
Campbell, P. K., Jones, K. E., Huber, R. J., Horch, K. W., Normann, R. A. "A silicon-based, three-dimensional neural interface: manufacturing processes for an intracortical electrode array." IEEE Transactions on Biomedical Engineering 38(8), August 1991, 758-768. https://pubmed.ncbi.nlm.nih.gov/1937509/
US Food and Drug Administration. 510(k) Premarket Notification K110010, NeuroPort Cortical Microelectrode Array System, Blackrock Microsystems, decision date 9 February 2011. https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K110010
Hochberg, L. R., Serruya, M. D., Friehs, G. M., et al. "Neuronal ensemble control of prosthetic devices by a human with tetraplegia." Nature 442(7099), 13 July 2006, 164-171. https://pubmed.ncbi.nlm.nih.gov/16838014/
Mitchell, P., Lee, S. C. M., Yoo, P. E., et al. "Assessment of Safety of a Fully Implanted Endovascular Brain-Computer Interface for Severe Paralysis in 4 Patients: The Stentrode With Thought-Controlled Digital Switch (SWITCH) Study." JAMA Neurology 80(3), March 2023, 270-278. https://pubmed.ncbi.nlm.nih.gov/36622685/
Precision Neuroscience. "Our Technology." Accessed July 2026. https://www.precisionneuro.io/our-technology
Willett, F. R., Kunz, E. M., Fan, C., et al. "A high-performance speech neuroprosthesis." Nature 620(7976), August 2023, 1031-1036. https://pubmed.ncbi.nlm.nih.gov/37612500/
Willett, F. R., Avansino, D. T., Hochberg, L. R., Henderson, J. M., Shenoy, K. V. "High-performance brain-to-text communication via handwriting." Nature 593(7858), May 2021, 249-254. https://pubmed.ncbi.nlm.nih.gov/33981047/
Card, N. S., Wairagkar, M., Iacobacci, C., et al. "An Accurate and Rapidly Calibrating Speech Neuroprosthesis." New England Journal of Medicine 391(7), 15 August 2024, 609-618. https://pubmed.ncbi.nlm.nih.gov/39141853/
Levin, A. D., Avansino, D. T., Kamdar, F. B., et al. "Cross-brain transfer of high-performance intracortical speech and handwriting BCIs." bioRxiv preprint, 14 January 2026. https://pubmed.ncbi.nlm.nih.gov/41648134/
Littlejohn, K. T., Cho, C. J., Liu, J. R., et al. "A streaming brain-to-voice neuroprosthesis to restore naturalistic communication." Nature Neuroscience 28(4), April 2025, 902-912. https://pubmed.ncbi.nlm.nih.gov/40164740/
Wairagkar, M., Card, N. S., Singer-Clark, T., et al. "An instantaneous voice-synthesis neuroprosthesis." Nature 644(8075), August 2025, 145-152. https://pubmed.ncbi.nlm.nih.gov/40506548/
Metzger, S. L., Littlejohn, K. T., Silva, A. B., et al. "A high-performance neuroprosthesis for speech decoding and avatar control." Nature 620(7976), August 2023, 1037-1046. https://pubmed.ncbi.nlm.nih.gov/37612505/
Singer-Clark, T., Hou, X., Card, N. S., et al. "Speech motor cortex enables BCI cursor control and click." Journal of Neural Engineering 22(3), 14 May 2025, 036015. https://pubmed.ncbi.nlm.nih.gov/40280150/
Kunz, E. M., Abramovich Krasa, B., Kamdar, F., et al. "Inner speech in motor cortex and implications for speech neuroprostheses." Cell 188(17), 21 August 2025, 4658-4673. https://pubmed.ncbi.nlm.nih.gov/40816265/
Neuralink. "PRIME Study Progress Update." 12 April 2024. https://neuralink.com/updates/prime-study-progress-update/
ClinicalTrials.gov. NCT05035823, "COMMAND Early Feasibility Study: Implantable BCI to Control a Digital Device for People With Paralysis," Synchron Medical. https://clinicaltrials.gov/study/NCT05035823
Synchron. "News." Accessed July 2026. https://synchron.com/news
Hettick, M., Ho, E., Poole, A. J., et al. "Minimally invasive implantation of scalable high-density cortical microelectrode arrays for multimodal neural decoding and stimulation." Nature Biomedical Engineering 10(6), 2026, 1206-1221 (published online 2 October 2025). https://pubmed.ncbi.nlm.nih.gov/41039113/
US Food and Drug Administration. 510(k) Premarket Notification K242618, Layer 7-T, Precision Neuroscience Corp., decision date 30 March 2025. https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K242618
Paradromics. "Connexus Brain-Computer Interface." Accessed July 2026. https://paradromics.com/connexus/
Paradromics. "Paradromics and University of Michigan Complete First Connexus BCI Implantation for the FDA-Approved Connect-One Clinical Study." 17 June 2026. https://paradromics.com/news/paradromics-completes-first-human-brain-computer-interface-bci-implantation/
ClinicalTrials.gov. NCT06429735, "PRIME: An Early Feasibility Study of a Precise Robotically Implanted Brain-Computer Interface for the Control of External Devices," Neuralink Corp. https://clinicaltrials.gov/study/NCT06429735
Neuralink. "PRIME Study Progress Update: Second Participant." 21 August 2024. https://neuralink.com/updates/prime-study-progress-update-second-participant/
Neuralink. "A Year of Telepathy." 5 February 2025. https://neuralink.com/updates/a-year-of-telepathy/
ClinicalTrials.gov. NCT07224256, "VOICE: An Early Feasibility Study of a Precise Robotically Implanted Brain-Computer Interface for Communication Restoration," Neuralink Corp. https://clinicaltrials.gov/study/NCT07224256
Neuralink. "Neuralink Receives Breakthrough Device Designation for Blindsight." 17 September 2024. https://neuralink.com/updates/neuralink-receives-breakthrough-device-designation-for-blindsight/
Neuralink. "Neuralink Receives Breakthrough Device Designation for Speech." 1 May 2025. https://neuralink.com/updates/neuralink-receives-breakthrough-device-designation-for-speech/
Neuralink. "Neuralink raises $650 million Series E." 2 June 2025. https://neuralink.com/updates/neuralink-raises-650m-series-e/
Chetty, N., Kacker, K., Feldman, A. K., et al. "Signal properties and stability of a chronically implanted endovascular brain computer interface." medRxiv preprint, 25 September 2025. https://pubmed.ncbi.nlm.nih.gov/41040697/
ClinicalTrials.gov. Studies sponsored by Synchron, Inc., first posted in 2026: NCT07446114 (Functional Outcomes and Control Using Synchron BCI, Canada), NCT07533903 (same, Australia), and NCT07543367 (INTENT early feasibility study, United States). https://clinicaltrials.gov/search?spons=Synchron
Liu, D., Shan, Y., Wei, P., et al. "Reclaiming Hand Functions after Complete Spinal Cord Injury with Epidural Brain-Computer Interface." medRxiv preprint, 6 September 2024. https://doi.org/10.1101/2024.09.05.24313041
Yao, R., Zhou, W., Liu, D., et al. "Fine grained two-dimensional cursor control with epidural minimally invasive brain-computer interface." medRxiv preprint, 10 October 2025. https://doi.org/10.1101/2025.10.06.25337264
Lévy, J., Zhang, M., Pinet, S., Rapin, J., Banville, H., d'Ascoli, S., King, J.-R. "Noninvasive decoding of typed sentences from human brain activity." Nature Neuroscience, published online 29 June 2026. https://doi.org/10.1038/s41593-026-02303-2
Tang, J., LeBel, A., Jain, S., Huth, A. G. "Semantic reconstruction of continuous language from non-invasive brain recordings." Nature Neuroscience 26(5), May 2023, 858-866. https://pubmed.ncbi.nlm.nih.gov/37127759/
Edelman, B. J., Zhang, S., Schalk, G., Brunner, P., Muller-Putz, G., Guan, C., He, B. "Non-Invasive Brain-Computer Interfaces: State of the Art and Trends." IEEE Reviews in Biomedical Engineering 18, 2025, 26-49. https://pubmed.ncbi.nlm.nih.gov/39186407/
Hahn, N. V., Stein, E., BrainGate Consortium, Donoghue, J. P., Simeral, J. D., Hochberg, L. R., Willett, F. R. "Long-term performance of intracortical microelectrode arrays in 14 BrainGate clinical trial participants." medRxiv preprint, 2 July 2025. https://pubmed.ncbi.nlm.nih.gov/40630584/
Colorado General Assembly. HB24-1058, "Protect Privacy of Biological Data," signed 17 April 2024. https://leg.colorado.gov/bills/hb24-1058
California Legislature. SB 1223, "Consumer privacy: sensitive personal information: neural data," chaptered 28 September 2024 as Chapter 887, Statutes of 2024. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202320240SB1223
UNESCO. "Ethics of neurotechnology: UNESCO adopts the first global standard in the cutting-edge technology." 5 November 2025. https://www.unesco.org/en/articles/ethics-neurotechnology-unesco-adopts-first-global-standard-cutting-edge-technology
Synchron. "Technology: The Stentrode BCI." Accessed July 2026. https://synchron.com/technology
UCLA Samueli Newsroom. "Cerebral connections: UCLA engineers tap into rich legacy of brain-computer interface technology." 19 April 2019. https://samueli.ucla.edu/brain-computer-interface/
BrainCo. "The Science Behind FocusCalm." Accessed July 2026. https://focuscalm.com/pages/how-it-works
ClinicalTrials.gov. Search for "brain-computer interface" filtered to recruiting studies, 94 results retrieved 24 July 2026. https://clinicaltrials.gov/search?term=brain-computer%20interface&aggFilters=status:rec
Barrow Neurological Institute. "Barrow Patient Begins to Reclaim Independence With Neuralink Robotic Arm After Spinal Cord Injury." Accessed July 2026. https://www.barrowneuro.org/about/news-and-articles/patient-stories/barrow-patient-begins-to-reclaim-independence-with-neuralink-robotic-arm-after-spinal-cord-injury/
ClinicalTrials.gov. Studies sponsored by Neuralink Corp: NCT06429735 (PRIME), NCT06700304 (CAN-PRIME), NCT06992596 (UAE-PRIME), NCT07127172 (GB-PRIME), NCT06710626 (CONVOY), NCT07224256 (VOICE). https://clinicaltrials.gov/search?spons=Neuralink

Editing: Brain-Computer Interface