The problem a BCI solves

Your intentions — move a hand, say a word — exist as patterns of electrical activity long before they become muscle movement. Injury or disease can break the path from brain to body while leaving the intentions intact: people with spinal-cord injuries, ALS (motor neurone disease) or brainstem strokes often can still generate the neural patterns for movement or speech they can no longer perform. A brain-computer interface rebuilds that path in silicon: record the signal, decode the intent, and drive an external device with it. That framing matters — a BCI is a bridge around a broken connection, not a mind-reading device.

How the three parts work

Recording. The trade-off that defines the field is signal quality versus invasiveness. Scalp electrodes (EEG) record through skull and skin: safe, cheap, wearable — but summed across millions of neurons and noisy. Electrodes placed under the skull or on the brain's surface (ECoG) improve fidelity. Implanted arrays inside the tissue record individual neurons or small groups: the highest information rate, at the cost of surgery and long-term biocompatibility questions. Neuralink's N1 implant, with over a thousand electrodes on flexible threads, is the most cited recent example; older research systems from the BrainGate consortium have run for years with implanted Utah arrays.

Decoding. Raw neural data is not language; it becomes control through machine learning. Decoders — today almost always deep-learning models — map recorded patterns to intents. The user trains the decoder by attempting an action repeatedly while the model learns what that attempt looks like in their data. Modern speech decoders illustrate how far this has come: research groups have shown systems that translate attempted speech from paralysed participants into text or synthesised voice at practical rates in the best cases, using vocabularies that expanded from hundreds of words to thousands as the models improved. The decoder adapts per person, per session — personalisation is inseparable from accuracy.

Output. The decoded intent drives whatever the user needs most: a cursor and keyboard (the workhorse of BCI studies — restoring computer access restores communication, gaming, and control of smart-home devices), a robotic arm for reaching and grasping, functional electrical stimulation that activates the user's own muscles, or a speech synthesiser. Each output restores a different slice of independence, which is why cursor-and-keyboard remains clinically central even when arms look more impressive.

What has actually been demonstrated

Keep achievements anchored to studies, not to demo reels. Implanted research systems have let people with paralysis point, click and type at speeds competitive with able-bodied smartphone use in the best cases; control robotic arms to drink and eat in scripted tasks; and communicate through speech decoders when typing is impossible. Neuralink's first human participant, implanted in January 2024, played chess and controlled a computer by intent alone, and additional participants followed in the expanded study. Non-invasive systems, commercially available as EEG headsets, deliver coarser control — enough for simple interfaces, rehabilitation training and consumer applications, not for precise robotic control. Everything beyond these anchored results — always-on thought typing, two-way sensory feedback at scale, consumer neurotech with implant-grade fidelity — is research pipeline, not product.

AI is the field's accelerant on both ends: decoding models improve as machine learning does (our AI research coverage tracks the underlying methods), and the same robotic systems BCI users drive are cousins of the arms documented on our robotics desk. A modern BCI is best understood as a neurotechnology wrapper around an applied-AI problem — the electrodes have changed less in twenty years than the decoders have in five.

Limitations, risks and open problems

Engineering. Implants face the longevity problem: tissue response and micro-motion degrade signals over months to years, and every implanted system needs a plan for what happens when performance fades. Bandwidth remains low — even the best decoders transmit a fraction of natural speech or movement speed. Wireless power and data through skin sets further limits.

Safety and regulation. Implantation carries genuine surgical and infection risk, which is why trials enroll people with severe paralysis who judge the trade-off worth making. Devices proceed through medical-device pathways — the FDA's breakthrough-devices programme in the US — and every capability claim above is anchored to a supervised study. Consumer EEG devices sit outside those medical pathways entirely; treat their marketing accordingly.

Ethics and privacy. Neural data is intimate by nature. Who owns it, whether it can be sold, how consent works for people with severe disability, and whether emerging "neurorights" should be codified are active questions — UNESCO and national bioethics bodies have ongoing programmes on exactly this. The honest position: today's systems decode narrow, user-controlled signals; tomorrow's will be more capable, so governance is being built ahead of the technology rather than after it.

How to read BCI news critically

The field generates outsized headlines, so apply the checklist from our critical-thinking guide with special care here. Three questions separate signal from hype: Was the result demonstrated in a peer-reviewed publication or a registered study, or only in a video? How many participants, over how long? And is the claimed capability a trained decoder doing a specific task, or a system-level promise? When a claim fails those tests, file it as aspiration — the real, documented progress in this field is already remarkable without embellishment. Primary literature lives on Nature's BCI subject pages, and trial listings on public clinical-trial registries.

Frequently asked questions

What is a brain-computer interface?
A BCI is a system that records signals produced by the brain, decodes them with software, and uses the result to control something external — a cursor, a keyboard, a robotic arm, a speech synthesiser. It does not read thoughts in the conversational sense; it detects patterns of neural activity that a person learns to modulate, and maps those patterns onto actions.
How does a BCI actually work?
Every BCI has three parts. Recording: electrodes capture neural signals — non-invasively through the scalp (EEG), or invasively from electrodes placed on or inside the brain, which capture far more detailed activity. Decoding: machine-learning models translate raw signals into intent — the direction of an arm movement, an attempted word, an imagined cursor motion. Output: the decoded intent drives a device. Users train the system by repeatedly attempting the action while the decoder adapts.
Is Neuralink available to the public?
No. Neuralink's implant has been tested in a small, FDA-supervised clinical study that began recruiting in 2024, with participants with severe paralysis using it to control computers. That is research, not a product: participation is limited to specific medical criteria, and no commercial availability has been announced. Anyone can buy non-invasive EEG headsets, but those capture much weaker signals and cannot match implanted systems' precision.
Can a BCI read my thoughts?
No. Current systems decode specific, trained patterns — attempted movement, attempted speech, attention — not free-form thoughts, and the person controls what signal patterns they produce. The decoded output is also limited: speech decoders work from a restricted vocabulary or through slow spelling interfaces. Privacy debates about “neurorights” are nonetheless real, precisely because the technology improves and generalises.
What can BCI users actually do today?
The strongest demonstrated results come from implanted systems in research trials: people with paralysis have controlled cursors and keyboards at practical speeds, driven robotic arms, and — most life-changing — produced speech through decoders that translate attempted speech into text or a synthetic voice at conversational rates in the best cases. All of this remains in supervised studies with small numbers of participants.
What are the main risks and limitations?
Implants carry surgical risk and degrade or migrate over time — electrodes move millimetres and lose their signal, so longevity is an open engineering problem. Non-invasive systems avoid surgery but record much weaker signals, capping their precision. Beyond the hardware, the field faces bandwidth limits (how much information per minute), decoder robustness outside the lab, and genuine ethical questions about consent, data ownership and identity that bodies like UNESCO are now studying.