Brain-Computer Interface
A brain-computer interface (BCI) is a system that records brain activity, translates it into commands or text, and sends the result to an external device — a cursor, a speller, a robotic arm, a speech synthesizer — without relying on the body's normal muscle pathways. The word "interface" is literal: signals flow from neurons through electrodes to a decoder, and in some designs feedback flows back through stimulation. Jacques Vidal used the term "brain-computer interface" in the early 1970s in his work on EEG-based communication, and the core idea has been constant since: neural activity carries intent, and a trained model can read some of it.
The field rests on the ability to measure electrical activity at different distances from the neurons, and the choice of sensor is the main dividing line. Hans Berger recorded the first human electroencephalogram in 1924, showing that brain electrical rhythms could be detected through the scalp. Non-invasive systems build on that: scalp EEG is safe and cheap but blurry, because the skull smears the signal, so it supports slow, low-bandwidth control such as the P300 speller or motor-imagery cursor control. Semi-invasive electrocorticography (ECoG) places electrode grids on the brain's surface beneath the skull, giving much cleaner signals. Fully invasive intracortical arrays, such as the Utah array, penetrate cortex to record individual neurons and offer the highest resolution at the cost of surgery, scarring and signal degradation over time. Endovascular approaches, such as the stent-mounted electrodes developed by Synchron, try to sit between the two by travelling through blood vessels.
Intracortical BCIs reached human users through the BrainGate research consortium, whose pilot clinical trial began in the mid-2000s. An early participant, Matthew Nagle, controlled a computer cursor and other devices with a cortical implant, reported in Nature in 2006. Later participants with paralysis used robotic arms to drink from a bottle and typed by moving a cursor across an on-screen keyboard. Work at Stanford and collaborating groups, including a 2021 Nature paper on decoding imagined handwriting, showed that rapid text entry was achievable by decoding attempted movements rather than just cursor motion.
The most striking recent results come from speech decoding. In 2023, Willett and colleagues at Stanford and other institutions reported in Nature a speech neuroprosthesis using intracortical arrays in a participant with ALS, decoding attempted speech to text at a markedly higher speed than earlier communication BCIs, with word error rates that depended on vocabulary size. Also in 2023, Metzger and colleagues at UC San Francisco reported in Nature a system using a large ECoG array on a participant with severe paralysis after a brainstem stroke, producing text, synthesized speech and a facial avatar. Both studies rely on machine learning that combines neural signals with a language model, which is why the output can be fluent even when the neural evidence for any single word is noisy. These were small studies with individual participants working with research teams, not consumer products.
Commercial and regulatory milestones have followed the research. Neuralink, founded in 2016, announced its first human implant in January 2024 as part of a clinical study, using a device with many thin flexible electrode threads placed by a surgical robot, and its early participant reported controlling a computer cursor. Other companies and academic groups pursue similar goals, and regulators such as the US Food and Drug Administration treat implanted BCIs as medical devices requiring staged trials. Open questions include device longevity, since scar tissue and shifting electrodes change signals; the need for recalibration; wireless power and data; and who owns and can access neural data, which has prompted early neural-privacy legislation in a few jurisdictions.
Common misconceptions are that BCIs read thoughts and that non-invasive headsets can do what implants do. Current systems decode narrow, trained signals — attempted movement or attempted speech in motor areas — not private inner speech or arbitrary ideas, and decoders must be fitted to each individual, often with hours of calibration. Consumer EEG headbands measure coarse rhythms that can indicate attention or drowsiness, not detailed intent. Performance numbers are also easy to misread: a high-accuracy result in one participant with a large implant does not transfer to everyone, and speed is typically reported in words per minute under constrained conditions rather than natural conversation.
Read brain-computer interface as a translation layer between a brain that still forms intent and a machine that needs commands, where the sensor sets the ceiling on information, the decoder sets the usefulness, and the burden of surgery and calibration sets who can benefit.