Brain-Computer Interfaces and Neuralink: Where the Clinical Evidence Actually Stands

Brain-computer interfaces (BCIs) establish a direct communication link between the brain and an external device, bypassing peripheral nerves and muscles altogether. For decades, that idea lived mostly in cyberpunk fiction or in university labs running low-bandwidth electroencephalography (EEG) setups. Over the past few years, however, a wave of venture-backed neurotechnology companies has pushed implantable systems into multi-site clinical trials.

Jacques Vidal coined the term "brain-computer interface" at UCLA in the 1970s and showed that humans could nudge a cursor using scalp EEG signals (Vidal, 1973). Because the skull smears electrical signals across centimeters of tissue, scalp EEG struggled with bandwidth and reliability. The real clinical turning point arrived with intracortical microelectrode arrays, particularly the Utah array (Cyberkinetics/Blackrock), which penetrated cortical tissue to record individual neurons. In 2006, the BrainGate consortium demonstrated that a tetraplegic participant could control a computer cursor using 96 microelectrodes implanted in the motor cortex (Hochberg et al., 2006).

That academic proof of principle laid the groundwork for today's commercial entrants. Neuralink, founded by Elon Musk in 2016, has drawn the most public attention. Since its first human implantation in January 2024, Neuralink has enrolled 21 participants across trials in the United States, Canada, and Great Britain, reaching cursor control speeds comparable to a standard computer mouse (8 to 10 bits per second) and a peak of 11.32 BPS (Neuralink, 2026). At the same time, competitors such as Synchron, Paradromics, Science Corporation, and Precision Neuroscience have taken very different engineering bets on how to balance signal resolution against surgical risk.

Inside Modern Intracortical BCIs

The Neuralink N1 Implant

Neuralink's current device, the N1 implant (marketed as "Telepathy"), records neural activity from the motor cortex. The hardware platform relies on three components: a flexible polymer electrode array, a hermetically sealed skull-mounted chip, and a surgical robot built specifically to insert the threads.

Electrode array. The N1 distributes 1,024 electrodes across 64 flexible polyimide threads, each thinner than a human hair and carrying 16 recording sites with thin gold or platinum conductors. Each site can detect single-neuron action potentials ("spikes"). While 1,024 channels is a major jump over a 96-channel Utah array, it is still a third of the 3,072-electrode, 96-thread specification Neuralink originally outlined in its 2019 white paper (Musk & Neuralink, 2019). The shift from rigid silicon needles to flexible polymer threads is meant to reduce mechanical shearing and scarring as the brain pulsates inside the skull.

Implanted chip. The N1 electronics sit inside a coin-sized (23 mm diameter, 8 mm thick) sealed package that replaces a matching piece of skull bone. Custom application-specific integrated circuits (ASICs) amplify, digitize, and transmit the recordings over Bluetooth, and the battery recharges inductively through the skin. Unlike earlier academic systems such as BrainGate2, which required a metal pedestal bolted through an open opening in the scalp, the N1 leaves no transcutaneous wires after the incision heals. Neuralink has announced plans for a next-generation implant with roughly 3,000 channels (Neuralink, 2026).

Surgical robot (R1). Because flexible polymer threads buckle if pushed into brain tissue by hand, Neuralink built the R1 surgical robot to thread them in using a fine needle. The robot uses optical imaging to map surface blood vessels and route each insertion around vasculature with a reported precision of roughly 50 micrometers. In the first human participant, threads were placed 3 to 5 mm deep into superficial cortical layers; after most of those threads slipped back out, the insertion depth was increased to 8 mm for subsequent surgeries.

Signal Processing and Decoding Pipeline

Turning raw voltage spikes into smooth cursor movement follows a four-stage pipeline:

1. Spike detection. On-chip ASICs apply threshold detection on each channel. Rather than running computationally heavy spike-sorting to isolate single neurons in real time, practical BCI pipelines generally treat threshold crossings on a channel as multi-unit activity.
2. Feature extraction. Spike counts or firing rates are aggregated over short time windows, typically 20 to 50 milliseconds, producing a feature vector for each time step.
3. Decoding model. Classic BCIs mapped firing rates to cursor velocity using linear regression or Kalman filters. Modern pipelines use deep sequence models. As of October 2026, Neuralink uses encoders based on Mamba (selective state-space models) pretrained via self-supervised learning on more than 50,000 hours of unlabeled neural recordings (Neuralink, 2026).
4. Operating system integration. The decoded movement vector and click intentions are exposed to the host computer or tablet as a standard Bluetooth pointing device.

The Thread Retraction Problem

Shortly after Noland Arbaugh received the first human N1 implant in January 2024, roughly 85% of the 64 electrode-bearing threads retracted from his cortex, leaving only about 15% of the channels active (Reuters, 2024). Thread retraction is a well-known failure mode of flexible polymer probes: the brain moves several millimeters relative to the skull during breathing, heartbeat, and posture changes, which can tug flexible leads out of shallow cortical tissue before scar tissue anchors them.

Neuralink worked around the hardware loss in two ways. For Arbaugh, engineers rewrote the decoding pipeline to rely on local field potentials and reweighted the surviving channels, which recovered and eventually surpassed his initial cursor speed without a second surgery. For subsequent participants, the FDA cleared a deeper 8 mm insertion depth to anchor the threads more firmly. According to company updates, signal stability improved in 18 of the next 20 implantations (Neuralink, 2026).

Even so, the episode highlights a basic materials challenge. The brain is soft, wet, and constantly moving. Flexible polyimide threads reduce chronic tissue damage compared with rigid silicon shanks, but keeping a high-channel electrode-tissue interface stable over five to ten years remains the hardest unsolved engineering problem in invasive neurotechnology.

Clinical Trials and Real-World Outcomes

Neuralink's PRIME, CONVOY, and VOICE Trials

By early 2026, 21 participants had received N1 implants across clinical trials in the US, Canada, and the UK (Neuralink, 2026). The studies enroll patients with cervical spinal cord injury (SCI) causing tetraplegia or amyotrophic lateral sclerosis (ALS) causing severe paralysis and loss of speech.

Cursor and robotic control. Several participants have reached cursor control rates of 8 to 10 bits per second (BPS), roughly matching able-bodied mouse use, with a peak record of 11.32 BPS using the pretrained Mamba decoder (Neuralink, 2026). Participants use the system daily for web browsing, messaging, computer-aided design (CAD), gaming, and, under the CONVOY protocol, controlling an external robotic arm for tasks like self-feeding.

Day-to-day autonomy. Beyond grid-target benchmarks, several participants have used the implant to resume studies or work. Noland Arbaugh returned to university coursework in neuroscience, and another participant with a spinal cord injury returned to medical school. For patients who previously relied on eye-trackers or sip-and-puff switches, low-fatigue cursor control that works in any posture or lighting condition is a practical step forward.

Speech decoding. In September 2026, Neuralink reported early results from its VOICE study (ClinicalTrials.gov ID: NCT07224256), where an ALS participant communicated through a brain-to-speech pipeline at 97.5% word accuracy, targeting conversational speeds near 140 words per minute (Neuralink, 2026). This builds directly on academic breakthroughs from the BrainGate consortium, which published a 256-electrode real-time voice synthesis neuroprosthesis in Nature in 2025 (Wairagkar et al., 2025).

Visual prosthesis (Blindsight). In September 2024, the FDA granted Breakthrough Device Designation for Blindsight, a planned system designed to stimulate the visual cortex directly using camera input, bypassing damaged eyes and optic nerves. Breakthrough status speeds up FDA communication, though human trials for Blindsight had not started as of October 2026.

Context from Academic Benchmarks: BrainGate

Evaluating Neuralink's numbers requires comparing them against two decades of peer-reviewed work from the BrainGate consortium at Stanford, Brown University, and Massachusetts General Hospital.

In August 2025, BrainGate researchers demonstrated an intracortical speech neuroprosthesis in Nature that synthesized a participant's voice in real time with closed-loop audio feedback from 256 microelectrodes in the speech motor cortex (Wairagkar et al., 2025). Earlier BrainGate studies established 90-character-per-minute decoded handwriting (Willett et al., 2021), rapid decoder calibration (Brandman et al., 2022), and multi-year single-neuron decoding in tetraplegic users (Brandman et al., 2023).

Those academic milestones show that high-speed cursor and speech decoding are not unique to the N1 chip. Much of the performance jump across the field comes from replacing linear filters with recurrent neural networks, transformers, and state-space models, combined with Neuralink's genuinely important packaging advance: making the implant wireless and flush with the skull.

Limitations of the Clinical Evidence

Three caveats stand out when reading current commercial BCI reports:

• No peer-reviewed clinical papers from Neuralink yet. While BrainGate publishes detailed methodology, failure modes, and raw metrics in Nature and NEJM, Neuralink's clinical data as of late 2026 comes almost entirely from company blog posts and livestreams. Company updates naturally highlight peak sessions and best-performing participants rather than median daily performance or signal dropout rates.
• Small, self-selected cohorts. Twenty-one participants is large for an intracortical BCI study, yet still small by Class III medical device standards. Early trial volunteers are unusually motivated and carefully screened.
• Short follow-up. No human has lived with an N1 implant for more than roughly two and a half years. Academic recordings with Utah arrays document steady declines in spike amplitude and channel yield over multi-year spans due to gliosis and insulation wear (Barrese et al., 2016).

Pretraining on 50,000 Hours of Brain Data

Self-Supervised State-Space Encoders

In October 2026, Neuralink described a self-supervised pretraining pipeline designed to tackle one of the biggest headaches in intracortical BCIs: daily recalibration (Neuralink, 2026). Because individual neurons shift slightly relative to the electrodes from day to day, traditional decoders go stale within hours, forcing users to spend 10 to 20 minutes every morning staring at calibration targets.

To build a drift-resistant decoder, engineers trained a Mamba-based state-space encoder (Gu & Dao, 2023) on more than 50,000 hours of unlabeled neural activity collected across trial participants, including 9,000 hours (22.4 billion spikes) from the first participant alone. Each channel receives a learned embedding that triggers when a spike occurs, fed through a Perceiver-style input layer. During pretraining, half of the recording channels are masked out and the model learns to predict their Poisson spike counts from the remaining half.

Once pretrained on this population dynamics task, the encoder needed far less supervised calibration: some participants went more than three weeks without recalibrating, and weekly calibration time dropped to roughly 10 minutes. The team also reported early cross-participant transfer, where an encoder pretrained on prior users accelerated setup for a newly implanted participant.

What Neural Foundation Models Can and Cannot Solve

Learning stable low-dimensional manifolds of motor cortex activity is a well-established idea in computational neuroscience, and scaling it to 50,000 hours of human recordings is a genuine engineering milestone. Still, calling it a universal "foundation model for the brain" stretches the analogy to text LLMs.

Human languages share an explicit discrete vocabulary across speakers; cortical microcircuits do not. Electrode placement varies by millimeters from surgery to surgery, and spinal cord injury or ALS reorganizes motor cortex representations differently in each patient. Moreover, algorithmic drift correction only works while enough physical electrodes still record neural activity above the noise floor. Software can compensate for a shifting neural population, but it cannot read from a dead or encapsulated wire.

Competing Hardware Architectures

Synchron: Endovascular Stentrode

Synchron avoids open brain surgery altogether. Its Stentrode is a 16-electrode self-expanding stent delivered through the jugular vein via catheter and deployed inside the superior sagittal sinus, a blood vessel running alongside the motor cortex. Because interventional neuroradiologists already perform thousands of venous stenting procedures every year, implantation takes place in a standard angiography suite without a craniotomy.

That surgical simplicity comes at the cost of spatial resolution: 16 vascular electrodes record bulk local field potentials rather than the 1,024 single-neuron channels of an intracortical array. Even so, across its SWITCH (Australia), COMMAND (US), and FOCUS-CAN (Canada) trials, Synchron has shown reliable one-year safety and digital device control, including native integration with Apple's BCI Human Interface Device (HID) profile on iPadOS (Synchron, 2025; Ognard et al., 2025). Backed by over $200 million in Series D funding, Synchron is preparing for a pivotal trial that could make the Stentrode the first permanently implanted BCI to win FDA premarket approval (PMA).

Paradromics, Science Corporation, and Precision Neuroscience

• Paradromics targets the high-bandwidth intracortical tier with its Connexus Direct Data Interface, using a rigid microwire array engineered for high-throughput speech and communication prostheses. The company completed its first human research implantation at the University of Michigan in June 2026 (Paradromics, 2026).
• Precision Neuroscience places an ultra-thin flexible micro-electrocorticography (ECoG) film (Layer 7 Cortical Interface) on the cortical surface through a narrow cranial slit, recording high-density surface potentials without puncturing brain tissue.
• Science Corporation, founded by former Neuralink president Max Hodak, focuses on its PRIMA photovoltaic subretinal implant for macular degeneration, stimulating retinal bipolar cells rather than visual cortex.

Analysts estimate the broader BCI market (still dominated by non-invasive EEG) at $2.3 to $3.2 billion in 2025 (BCC Research, 2025). Within the implantable segment, Neuralink's $650 million Series E in June 2025 valued the company at roughly $9 billion (Reuters, 2025), ahead of Science Corporation ($1.5B) and Synchron ($1B+).

Ethics, Neural Privacy, and Regulation

Mental Privacy and Inferred Neural Data

High-channel neural recordings capture sensitive biological data. While a motor cortex decoder only looks for intended hand velocity or phoneme articulation, raw cortical signals also reflect arousal, cognitive load, and pathology. When BCI telemetry streams through consumer phones and cloud backends, standard medical privacy frameworks leave major gaps. In the United States, HIPAA applies to healthcare providers and insurers, not necessarily to consumer neurotech device makers or cloud telemetry pipelines.

Chile amended its constitution in 2021 to protect brain activity and neurorights (Ruiz et al., 2024), and UNESCO adopted its global Recommendation on the Ethics of Neurotechnology in November 2025 to set baseline standards around consent and mental integrity (UNESCO, 2025). As a 2026 Stanford Law analysis points out, however, statutory protections written around raw neural voltage recordings often leave inferred mental data, the intentions and states reconstructed by AI decoders, in a legal gray zone (Stanford Law School, 2026).

Consent, Explantation, and Long-Term Support

Patients with tetraplegia or late-stage ALS face severe daily limitations, which requires careful informed consent to avoid therapeutic misconception during early feasibility trials (Ienca et al., 2025). Unlike taking a pill, participating in an implant trial raises long-term hardware questions: Who pays for maintenance, software updates, or explantation if a startup pivots, runs out of funding, or deprecates an early hardware generation? Regulatory approval through the FDA's Class III PMA pathway will require not only multi-year durability data, but also binding commitments for long-term patient support and secure local-first data architectures.

Where the Field Goes Next

Over the next three to five years, implantable BCIs will be judged on three practical tests rather than demo videos:

1. Five-year hardware survival. Can polymer threads, microwire bundles, or endovascular stents maintain usable signal-to-noise ratios across five years in the human body without requiring repeat craniotomies?
2. Bidirectional safety. Adding stimulation for tactile feedback in robotic prostheses or phosphene generation in visual cortex (such as Blindsight) introduces new constraints around charge density, tissue heating, and seizure thresholds.
3. Peer-reviewed clinical transparency. Moving from investigational device exemptions (IDEs) to approved medical therapies will require publishing complete cohort data, including non-responders, surgical complications, and hardware revisions, in peer-reviewed journals.

For patients with severe paralysis or ALS, restoring fluent speech and independent computer control is already a life-changing clinical achievement. Turning that achievement into a durable, trustworthy medical standard will depend on materials science, transparent clinical reporting, and strict safeguards around neural data.