How Brain-Computer Interfaces Work: Neuralink, Synchron, Precision, and the Science Behind Them

A brain-computer interface (BCI) records electrical activity from neurons and converts it to numbers. A trained model then predicts what the person intends to do. Today's implanted BCIs read movement intent, such as an attempted hand movement or attempted speech. They do not read private thoughts. Neuralink is the best-known example, but it is one of at least six companies and labs with devices in people.

This guide explains each step of the system: the electrodes, the chips, the radio, the surgical robot, and the AI decoder. It uses Neuralink as the main case study. It compares Neuralink with Synchron, Precision Neuroscience, Paradromics, Blackrock Neurotech and the BrainGate labs, and a Chinese device that won market approval in 2026.

The key numbers

A BCI reads movement on its way to the muscles

A useful comparison is lip reading. A lip reader watches the mouth move and infers the words. The lip reader does not know what the speaker thinks but does not say. An implanted BCI works in a similar way. It watches the brain's motor commands and infers the intended movement.

Neuralink places its electrodes in the hand area of the motor cortex. Its technology page says that neurons in movement-related areas "represent intended movements." Before the first surgery, an fMRI scan located the part of the precentral gyrus that was active when the participant attempted to move his hand and arm.

Speech BCIs follow the same rule. The Stanford team behind Willett et al. (2023) recorded from area 6v, a premotor area that controls the face and mouth. Neurons there were tuned to speech articulators, such as the lips, jaw, and tongue. The decoder read attempted mouth movements, then a language model turned them into words.

This is why the people in these trials often say they "think about moving." The BCI needs the motor plan. A thought that never becomes an attempted movement leaves no pattern that the decoder was trained to find.

Why writing into the brain is harder than reading

Reading is a pattern recognition problem. The brain already makes the patterns, and a model learns to label them. Writing is a control problem. The device must create activity that the brain accepts as real sensation, such as a point of light.

Writing uses electrical stimulation. A pulse from one electrode excites the cells around its tip, not one chosen neuron. The person then has to learn what the new signals mean. Neuralink's 2019 chip could stimulate on every channel, but the paper did not demonstrate it. For Blindsight, its vision program, Neuralink showed an S2 chip with "over 1,600 channels of electrical stimulation" in its Summer 2025 update. The company said it planned a first Blindsight participant in 2026, with navigation as the goal.

The signal chain, step by step

Every implanted BCI follows the same chain. The details below use Neuralink's published numbers where they exist.

1. A neuron fires. A spike, or action potential, is a short electrical pulse that a neuron sends to other cells. An electrode near the cell picks up a small voltage change.
2. The electrode records it. An electrode is a small metal contact. Neuralink's electrodes sit on flexible polymer threads. The 2019 paper describes threads 5 to 50 µm wide and about 4 to 6 µm thick.
3. A chip amplifies and digitizes the signal. The 2019 chip had 256 amplifiers and sampled each channel at 19.3 kHz with 10-bit resolution. Its input noise was 5.9 µV RMS, and the whole chip used about 6 mW.
4. The implant detects spikes. Detection on the chip keeps the events and drops most of the raw waveform. The data-rate section below shows why this step is necessary.
5. A radio sends the data out. The N1 sends its data wirelessly to the Neuralink app. Its battery recharges by induction through the skin, so there is no cable or connector.
6. A decoder predicts intent. The decoder is a machine learning model trained on that user's own data. It maps neural features to cursor velocity, a click, a letter, or a word.
7. The computer acts. The cursor moves, or text appears, or a robotic arm moves. The person sees the result and corrects the next attempt, which closes the loop.

Where each BCI sits: signal quality against invasiveness

BCIs differ mainly in how close the electrodes get to neurons. Closer electrodes record single cells. Farther electrodes record the blurred sum of many cells but need less surgery. The 2019 Neuralink paper states the trade-off directly. Noninvasive methods "record the average of millions of neurons through the skull." Surface electrodes average "the activity of thousands of neurons."

The major BCI companies and labs compared

Three points from the table need care.

1. "First approved" depends on the definition. NEO is the first implanted motor BCI approved for routine use. Its electrodes sit on the dura, so it is less invasive than Neuralink or the Utah array.
2. Precision's clearance is for temporary use. The 510(k) covers up to 30 days, for uses such as brain mapping during surgery. It is not a permanent communication device yet.
3. Paradromics' 200 bits per second is from animals. Do not compare it with human cursor results.

China has more programs. A Chinese Academy of Sciences team implanted a wireless intracortical system in a man with tetraplegia in June 2025. By January 2026 he could steer a wheelchair and direct a robot dog, according to a Xinhua report hosted by the institute.

Inside the Neuralink N1 implant

The N1 is a sealed disc that replaces a small piece of skull. Neuralink's technology page lists four parts: a hermetic enclosure, a battery charged "via a compact, inductive charger," custom low-power chips, and the threads.

Neuralink's published specifications changed between the 2019 research system and the 2024 clinical implant. The table compares them.

The 2019 system was a research platform for animals, so it could use a cable and more power. The paper said the clinical version would need "on-board signal compression, reduced power consumption, wireless power transmission, and data telemetry through the skin." The N1 is that version, with one third of the electrodes.

Neuralink's roadmap is to grow again. Its January 2026 post says the company is moving "from 1000 to 3000 electrodes." In the Summer 2025 update, it set a goal of more than 25,000 channels per implant by 2028.

Several outlets describe the N1 radio as Bluetooth. Neuralink's own pages say only that the implant transmits "wirelessly." This guide uses Bluetooth LE as a reference radio for the math below, not as a confirmed N1 specification.

The R1 robot: why a machine inserts the threads

Each thread is thinner than a hair and too flexible to push into tissue by hand. Neuralink built the R1 robot to place each thread at a chosen site while it avoids blood vessels on the brain surface. The robot head holds five camera systems and an optical coherence tomography (OCT) system, which images tissue depth with light, per the technology page.

The 2019 paper gives the mechanics. The needle is tungsten-rhenium wire etched to 24 µm in diameter. A linear motor drives it, with retraction acceleration of up to 30,000 mm/s², so the thread separates from the needle. Under 405 nm light, the polyimide thread fluoresces, and computer vision finds its 16 × 50 µm loop. In those animal surgeries, the average insertion success rate was 87.1% over 19 surgeries.

Two speed numbers seem to disagree. The 2019 robot could insert "up to six threads (192 electrodes) per minute," or one thread every 10 seconds. The human R1 robot needs 17 seconds per thread in the best case, according to the Summer 2025 update. Human surgery adds safety margins, brain motion, and a thicker brain, so a slower clinical robot is plausible. Neuralink says its next-generation robot needs 1.5 seconds per thread. In the same update, Neuralink said the new robot reaches more than 50 mm below the brain surface. These are company figures, and no independent measurement exists yet.

What Neuralink's patents claim

A patent describes what a company claims as its invention. It does not prove that the product uses that design. The six granted US patents below are assigned to Neuralink Corp. on Google Patents. They match the chain in FIG 01.

Three of the six patents are about cutting data on the chip before transmission. The next section shows why.

The data-rate math: why the implant must compress

A wireless implant cannot send every raw sample. The worked example uses the 2019 chip's published numbers. Neuralink has not published the N1's sampling rate, so treat the result as an estimate.

1. Count the channels. The N1 has 1,024 electrodes.
2. Multiply by the sampling rate. The 2019 chip sampled at 19.3 kHz, so 1,024 × 19,300 = 19.8 million samples per second.
3. Multiply by the bits per sample. At 10 bits, the raw stream is 197.6 million bits per second (Mbit/s).
4. Compare it with the radio. The Bluetooth 5 LE 2M PHY runs at 2 Mbit/s at the physical layer. Real application throughput is lower.
5. Divide. 197.6 ÷ 2 = 98.8. The raw data is about 99 times what that radio can carry.
6. Send spike counts instead. As an illustration, count spikes per channel in 20 ms bins (50 bins per second) with 4 bits per count. That is 1,024 × 50 × 4 = 0.2 Mbit/s.

The spike-count stream is 197.6 ÷ 0.2048 = 965 times smaller than the raw stream. It fits inside the radio budget with room for headers and retries. The decoder also needs only the spike information, not the full waveform. This is why the 2019 paper calls for on-board compression, and why the patents above detect spikes and drop silent channels on the chip.

The heat limit

Power creates heat, and brain tissue is sensitive to temperature. ISO 14708-1 is the international standard for active implants. It limits an implant's outer surface to 2 °C above the normal body temperature of 37 °C, as the ITRUSST consensus paper summarizes. Some BCI designers use a stricter 1 °C rise as their own target.

One study modeled a skull-mounted BCI unit with the 1 °C target. Serrano-Amenos et al. (2023) estimated a budget of about 70 mW, with a range of 47 to 81 mW. The 2019 Neuralink system used 750 mW through a cable. That is about 11 times the modeled budget, though the two devices have different shapes and positions. A fully implanted device must stay within a budget of tens of milliwatts. That budget must cover the amplifiers, the spike detection, and the radio.

How the decoder turns spikes into a cursor or a sentence

The decoder is where AI does most of the work. It is a machine learning model that learns the link between neural activity and intended action for one person.

Cursor decoders

A cursor decoder maps a vector of spike counts to a cursor velocity and a click signal. Classic BCI labs used a Kalman filter, which is a statistical model that blends a prediction of the cursor's motion with each new measurement. Newer systems use neural networks. The first BrainGate participant used a decoder to open e-mail and operate a television, as Hochberg et al. (2006) reported. By 2012, two BrainGate participants controlled a robotic arm to reach and grasp, in a second Nature paper.

Speech decoders and language models

Speech decoders add a second AI model. In Willett et al. (2023), a recurrent neural network (RNN) output the probability of each phoneme every 80 ms. A phoneme is the smallest unit of sound in a word. A 125,000-word trigram language model then picked the most likely sentence. The authors note that the language model often corrected phoneme errors.

Results improved fast:

The UC Davis participant in the 2026 study used the system without researchers present. He kept full-time employment. He also used the speech BCI as a keyboard and the cursor BCI as a mouse.

Neuralink entered speech in 2025. The FDA gave it a Breakthrough Device designation for speech in May 2025. Its VOICE study aims for 140 words per minute. In an October 2026 post, Neuralink engineers described a phone app for speech and automatic detection of attempted speech. They also built a prototype that starts a Grok computer-use agent by imagined speech.

Drift, recalibration, and the move to foundation models

Neural signals drift. Electrodes move slightly, cells die or change, and the person's state varies. A decoder trained on Monday loses accuracy by Friday. Users must recalibrate by doing a structured task that creates new labeled data.

Neuralink measured that cost. In its October 2026 post, users spent an average of 55 minutes per week on calibration. The company's answer borrows from large language models:

1. Collect unlabeled data. Participants had streamed more than 50,000 hours of freeform neural data.
2. Pretrain an encoder. A Mamba-based model, which is a sequence model with low and constant latency, learned to predict half of the channels from the other half.
3. Train a small decoder on top. The decoder reads the encoder's stable embeddings instead of raw spikes.

The results were company-reported. Thirty seconds of labeled embeddings matched about 3.5 minutes of raw spikes. Some users went from 10 minutes of calibration per day to 10 minutes per week. One participant set the 11.32 bits per second record. Models pooled across participants did not yet beat single-person models in live use.

How BCI speed is measured: bits per second

Bits per second measure how much information a person can send through the interface. The standard formula comes from Wolpaw et al. (2002). For N choices and accuracy P, the bits per selection are:

B = log₂N + P·log₂P + (1−P)·log₂((1−P)/(N−1))

A worked example with four targets and 90% accuracy:

1. Compute log₂N. log₂4 = 2 bits.
2. Compute the correct-choice term. 0.9 × log₂0.9 = 0.9 × (−0.152) = −0.137.
3. Compute the error term. 0.1 × log₂(0.1 ÷ 3) = 0.1 × (−4.907) = −0.491.
4. Add the terms. 2 − 0.137 − 0.491 = 1.372 bits per selection.
5. Multiply by the selection rate. At 30 selections per minute, 1.372 × 30 = 41.2 bits per minute, or 0.69 bits per second.

Neuralink uses a grid game called Webgrid. Its formula multiplies the information per selection, log₂(clicks × grid size² − 1), by net correct selections per second. On a 30 × 30 grid with one click type, each selection carries log₂(899) = 9.81 bits. Fifty-eight net correct selections in a minute give 9.81 × 58 ÷ 60 = 9.5 bits per second.

Neuralink's reference for an able-bodied person with a mouse changed over time. In May 2024 it cited its engineers at about 10 bits per second. In January 2026 it cited 8 to 10 bits per second on average. Its first participant reached 4.6 bits per second in his first session and 8.0 within about 100 days.

Neuralink's trials and regulatory status

Neuralink's first human implant was in January 2024 at Barrow Neurological Institute in Phoenix. The FDA had approved the PRIME study under an investigational device exemption (IDE) in May 2023. An IDE lets a company test an unapproved device in a clinical study. It is not approval to sell.

The program grew quickly after the first year.

Neuralink says it has seen higher signal quality in 18 of the 20 participants after the first, following its retraction fixes. It also reports zero serious device-related adverse events. Both claims come from company posts, not from a peer-reviewed safety report.

The counts also need care. Neuralink reported five participants in June 2025 and seven at its June 2025 update. In January 2026 it reported 21 participants, but its own chart of surgeries by half-year sums to 20.

What is still hard

1. Longevity. The body covers foreign materials in scar tissue. Neuralink's safety post calls this encapsulation the historical "key limitation of BCIs." Paradromics states a goal of stable recordings for more than 10 years. No company has public data for that span in many people.
2. Mechanical stability. Neuralink's first participant lost signal from retracted threads. The company is testing mechanical features to improve thread retention.
3. Variation between people. Neuralink reports that performance varies with anatomy, such as intracranial spacing, and with disease stage in ALS.
4. Regulatory proof. Neuralink states that its devices "are investigational and have not been approved by the FDA" in its October 2026 post. Synchron says on its technology page that its system is not approved for commercial use anywhere. A device for permanent use needs larger trials and a premarket approval.
5. Writing. Stimulation for vision or touch is at an earlier stage than reading.

Which approach fits which patient

There is no single best BCI. The right choice depends on the goal and on the risk the patient accepts.

1. If the goal is a few reliable commands with low surgical risk, an endovascular device such as the Stentrode fits. It gives fewer channels.
2. If the goal is fast typing or speech, penetrating arrays have the strongest published results. The UC Davis, Stanford, and Neuralink data come from electrodes inside the cortex.
3. If the use is temporary, such as brain mapping during surgery, a surface film fits. Precision's Layer 7 is cleared for up to 30 days.
4. If the person is in China and has a C2 to C6 cervical spinal cord injury, NEO is approved there for routine clinical use.

What BCIs teach about AI in business

A BCI is an AI system that turns a noisy, drifting signal into an action, with a person in the loop. The same problems appear in business AI.

For your own AI project, do one thing first. Pick the one metric that defines success for the task, as Webgrid does for the cursor. Then measure it before and after every change.

Bring AI From Noisy Data to Reliable Action in Your Company

In a BCI and in a business, the value is in the decoder, the evals, and the loop that keeps them accurate over time. OpenNash builds AI systems for business operations, not medical devices. We put forward-deployed engineers on your team, and we work with Fortune 2000 companies.

1. Define the signal and the action. Map which data predicts which business decision, and who approves it.
2. Build the eval first. Set one task metric with a pass threshold, like bits per second for a cursor.
3. Pick the right model for each step. Use small models for narrow, high-volume steps and frontier models for open-ended ones.
4. Plan for drift. Monitor quality in production, and retrain or recalibrate on a schedule.
5. Keep a person in the loop. Add approval steps for high-risk actions, and train your team to use them.

At the end, your company owns the code, the evals, the data pipelines, and the runbooks.

Book a call with an OpenNash forward-deployed engineer or email us at hello@opennash.com. In the first session, we find the workflow where reliable AI decisions return the most value, and we estimate its ROI.