Brain–Computer Interfaces: The Chip Inside the Skull and the AI Outside It
A person with paralysis moves a cursor by thinking about moving their hand. A woman with ALS "speaks" at 62 words per minute through an implant. These are brain–computer interfaces (BCIs), and in 2025–26 they moved from a few lab demonstrations to dozens of implanted patients, a first market approval in China and a fast-growing industry. This explainer covers how they actually work: the electrodes, the tiny chip that has to fit inside strict heat and power limits, the AI that turns brain signals into words, the companies and trials, and the privacy questions they raise.
1. What a BCI is, and what it isn't
A brain–computer interface records electrical activity from the brain and translates it into an action: moving a cursor, selecting letters, producing text or synthetic speech, or controlling a robotic arm. Today's medical BCIs are built for people who have lost movement or speech through spinal-cord injury, stroke or ALS.
Two common misunderstandings are worth clearing up. First, current BCIs do not "read minds" in general: they decode specific, trained signals, mostly from the motor cortex, such as attempted hand movements or attempted speech. Second, a BCI is not a new kind of computer for AI. It is an input and output device. The implant's chip does very little computation; the AI that decodes the signals runs on ordinary phones, laptops or servers. That is why BCIs were not on the scorecard in Beyond Silicon: they use AI compute rather than add to it.
2. The signal chain
Every BCI follows the same path from neurons to action. Click a stage to see what happens there and how much data passes through it.
From neuron to cursor
3. Three ways in (and one way around)
The closer the electrodes are to individual neurons, the clearer the signal, and the bigger the surgery. Choose an approach.
Signal quality vs invasiveness
4. The chip's hard limits
The implanted chip is a tiny, extremely constrained computer. Its job is to amplify microvolt signals, digitise them, find the "spikes" (neurons firing), compress the data and send it out wirelessly, all while sealed inside the body.
- Heat. Implant safety studies commonly use a limit of no more than 1 °C warming of surrounding tissue. In thermal simulations this allows only around 5 mW for a small 2 × 2 mm chip sitting in tissue, and roughly 70–75 mW for a larger unit mounted in the skull, where heat spreads more easily.
- Data. Raw recording is a firehose: about 1,000 channels × 30,000 samples per second × 16 bits ≈ 480 Mbps. Detecting spikes and compressing on the chip brings that down to roughly 10–100 Mbps or much less ⚠️ secondary estimate. Transmitting raw data would cost too much power.
- Power. Fully implanted devices recharge wirelessly through the skin by induction, which itself generates heat.
- Longevity. The body attacks foreign objects; electrodes can be encapsulated by scar tissue and signals can degrade over years.
Heat and data budget simulator
Illustrative model. Power per channel and radio cost are assumptions chosen to be in a plausible range; the point is the trade-off, not exact numbers.
5. Worked example: Neuralink's N1
Neuralink's N1 is a coin-sized, fully implanted device that replaces a disc of skull. It records from 1,024 electrodes spread across 64 flexible threads, each thinner than a human hair and too fine to place by hand, so a surgical robot (R1) inserts them while avoiding blood vessels. Custom low-power chips in the implant process the signals before sending them wirelessly to an app; the battery is charged inductively.
6. The AI that decodes
The part of a BCI that has improved fastest is the software. Speech BCIs record from the parts of motor cortex that control the lips, tongue and larynx. A neural network maps each short window of neural activity to probabilities of phonemes (speech sounds); a language model then picks the most likely word sequence, correcting many errors, much as autocorrect does.
From noisy phonemes to words
Illustrative. Increase the decoder noise and see how much the language model recovers.
| Study | Participant | Result |
|---|---|---|
| Willett et al., Nature 2023 (Stanford) | Woman with ALS | ~62 words per minute; 9.1% word error on a 50-word vocabulary, 23.8% on 125,000 words |
| Card et al., NEJM 2024 (UC Davis) | 45-year-old man with ALS; 256 electrodes in 4 arrays | 99.6% accuracy on 50 words on day 1; 90.2% on 125,000 words on day 2; 97.5% sustained beyond 8 months; ~32 words per minute over 248+ hours of use |
| Kunz et al., Cell 2025 (Stanford/Emory) | 4 participants with ALS or stroke | Decoded inner (imagined) speech with up to 74% accuracy; attempted and inner speech produce similar patterns |
| Meta, Nature 2025 (non-invasive) | Thousands of volunteers | Wristband reading muscle signals (sEMG): handwriting at 20.9 words per minute with decoders that work across people without calibration |
7. Who's where (late 2026)
| Company | Approach | Status |
|---|---|---|
| Neuralink (US) | 1,024 electrodes on 64 threads in tissue, robot-inserted, fully implanted | 21 participants (Jan 2026); speech and vision programmes |
| Synchron (US) | Stentrode inside a blood vessel, delivered by catheter via the jugular vein | COMMAND early-feasibility study (6 patients): met its 1-year safety endpoint, 100% successful deployment, median 20 minutes; ~100 patients by mid-2026 ⚠️ secondary |
| Precision Neuroscience (US) | Layer 7: thin film with 1,024 electrodes on the brain surface | FDA 510(k) clearance, April 2025, for implants of up to 30 days; fully implantable system in development |
| Paradromics (US) | Connexus: high-channel-count array in tissue | First trial implant reported June 2026 ⚠️ secondary |
| Blackrock Neurotech (US) | Utah array, used in academic BCI research for two decades | Research use (e.g. many of the speech studies above used Utah-type arrays) |
| Neuracle (China) | NEO implant for quadriplegia from cervical spinal-cord injury | Market approval in China, March 2026, reported as the first implantable BCI approved as a Class III medical device |
| NeuCyber / StairMed (China) | Semi-invasive (Beinao-1) / ultra-flexible 64-channel | Beinao-1: 7 implants; StairMed: 3 implants in 2025; China names BCI a strategic industry |
The regulatory ladder
What the common terms mean, from least to most. Click a rung.
8. Risks, ethics and law
- Medical risk. Brain surgery carries infection and bleeding risk; endovascular devices trade some signal quality for avoiding open surgery.
- Durability and abandonment. Electrodes degrade, and patients depend on companies for support; if a company fails, users can be left with unsupported implants.
- Mental privacy. Neural data can reveal more than intended, as the inner-speech study showed.
- Law. Colorado became the first US state to classify neural data as sensitive personal data (HB 24-1058, effective August 2024); California followed (AB 1008 and SB 1223, effective January 2025), and Montana in 2025. In November 2025 UNESCO adopted the first global Recommendation on the Ethics of Neurotechnology, which is not legally binding but calls for mental-privacy safeguards and freely given, informed consent.
9. What's next
🔢 Reading the numbers
- Peer-reviewed: Willett et al. (Nature 2023), Card et al. (NEJM 2024), Kunz et al. (Cell 2025), Meta (Nature 2025); thermal limits from engineering papers.
- Company and regulator milestones: Neuralink's January 2026 update, Synchron's COMMAND results, Precision's FDA clearance, Neuracle's approval (as reported by Scientific American and Fudan University).
- ⚠️ Secondary: Synchron's ~100 patients, Paradromics' first implant, Neuralink's international trial sites, and the compressed data-rate range.
- Illustrations: the signal-chain data rates, budget simulator and decoder demo.
📚 References
- Clinical and scientific
- Stanford Medicine: Brain implant enables speech at 62 words per minute (Willett et al., Nature 2023)
- Card et al.: "An Accurate and Rapidly Calibrating Speech Neuroprosthesis" (NEJM 2024)
- NIH NINDS: Decoding inner speech brain signals (Cell 2025)
- Meta Reality Labs: "A generic non-invasive neuromotor interface for human-computer interaction" (Nature 2025)
- Thermal analysis of a skull implant in brain–computer interfaces
- Companies and regulators
- Neuralink · Neuralink (trial history)
- Synchron: Positive results from the US COMMAND study
- Fierce Biotech: FDA clears Precision Neuroscience's implant
- Scientific American: China approves its first brain implant for commercial use
- NeuCyber says it is three years behind Neuralink
- Law and ethics
- KFF Health News: States pass privacy laws to protect brain data
- Inside Privacy: UNESCO adopts first global framework on neurotechnology ethics
- Related on this site
- Beyond Silicon · AI at the Edge