What gets read out is never a "thought" — it's an order that was already on its way to a muscle, stopped halfway.
Someone has been paralysed for over a decade. A small opening is made in the skull and an electrode array smaller than a fingernail goes in. He looks at a screen, thinks right, and the cursor slides right. It looks like mind-reading. The truth is far more sober, and far more interesting: what the machine picks up is not a thought but the command motor cortex was already sending to the hand. The line to the muscles is cut in the spinal cord; the command is still being issued, with nobody to receive it. A BCI collects it. And the counter-intuitive part comes later: what usually makes these systems usable is not a cleverer decoder, but a user who has learned how to be read.
Neurons speak in electrical pulses — a voltage jump lasting about a thousandth of a second, called a spike. To read them you have to measure electricity, and the first decision is not which algorithm to use. It's how far away to stand.
Scalp EEG is entirely non-invasive, but the skull is a thick smearing layer: it wipes out detail and leaves a coarse outline. It's like hearing a crowd roar through the stadium wall — you can tell someone scored, not what anyone is shouting. Spatially you get about centimetres. Open the skull and lay a thin sheet on the cortical surface (ECoG), without entering the tissue, and resolution jumps to millimetres. Push dozens or hundreds of fine needles a few millimetres into cortex — the classic Utah array carries 96 channels — and only then do you hear individual neurons spike.
So the question "can non-invasive BCIs catch up?" is mostly physics, not software: the skull has already destroyed the information, and destroyed information cannot be computed back. motor system & cerebellum
In the 1980s, recordings from monkey motor cortex showed that each neuron has a preferred direction: it fires hardest when the arm reaches one particular way, and less as the reach swings away from it. But the tuning curve is extremely broad — so broad that watching a single neuron tells you almost nothing about where the hand is going.
Weight each neuron's preferred direction by how fast it is currently firing, add the little arrows tip-to-tail, and the resultant points accurately at the true direction. That is the population vector, and it illustrates a principle that runs through this whole site: information lives in the distribution, not in the cell. Which is why a few dozen extra channels often buys more than a new algorithm.
Today's decoders skip the hand-computed vector and fit a statistical model instead. You have the person imagine a batch of known movements repeatedly (or "shadow" a cursor that moves on its own), collect pairs of "how much each channel fired right now" and "what the target velocity was", then fit the mapping from the first to the second — Kalman filters early on, recurrent neural networks now. Put plainly, neural decoding is a supervised regression problem whose training set is your brain × known answers. The best recent results all come from this route: imagined handwriting at roughly 90 characters per minute, imagined speech at 60–80 words per minute — well past half of conversational speed.
A speech BCI is architecturally almost identical to the speech recognition on your phone: a network first emits probabilities over phonemes (sound units smaller than a word), and a language model — the same family as an LLM — picks the most plausible sentence. The decoder never has to be perfect; the language model patches the holes with prior knowledge. That also explains something often misread: the smaller the vocabulary, the more miraculous the accuracy looks. The same neural signal, given a large vocabulary, does markedly worse.
One fact gets skipped almost every time: a BCI is not an offline readout, it is a closed loop. The user sees the cursor move and immediately corrects; correcting changes the distribution of neural activity; and once that shifts, the ground the decoder was fitted on has moved. Both sides are learning each other.
One experiment made this vivid. Monkeys first learned to control a cursor through a BCI; then the decoding map was quietly swapped. If the new map lay inside the activity patterns those neurons already produce together — researchers call it the intrinsic manifold; a hundred neurons may sound like a hundred free dials, but in practice they only co-vary in a handful of combinations, like an orchestra that knows a limited set of chords — the animals relearned in minutes. If the new map demanded combinations outside that set, days of practice weren't enough.
Worth pausing on: plasticity has a shape. The new mapping you can pick up quickly has to be something your network already knows how to say. Asking it for unfamiliar grammar costs so much that it barely learns at all. (The manifold idea is rich enough for its own instalment — Topic 37.) synaptic plasticity
There is a moving side-effect, too. Expert users quietly abandon the strategy of "imagining my hand moving", because directly modulating the handful of recorded neurons is simply easier. At that point the cursor stops being an object being operated and starts being part of the body.
Two learning systems updating inside one loop — the decoder learning the person while the person learns the decoder — has a name in machine learning: co-adaptation. The instability it creates is the same one behind two GAN networks chasing each other, and behind policy and reward model drifting apart in RLHF: each side's training data is produced by the other, so there is no fixed target. The engineering fixes rhyme as well — slow one side down, or periodically freeze one and recalibrate against it, rather than letting both run loose at once.
Reading out is by now a mature line of work. Writing into the brain is far harder — and the difficulty is beautifully physical.
Electrical stimulation shouts, it doesn't speak. When an electrode fires, every cell within its radius is driven at once, excitatory or inhibitory, whatever its role in the circuit — while the information itself is carried in fine spatiotemporal patterns. Reading lets you listen selectively. Writing gives you one megaphone.
The genuinely successful case of writing has existed for decades under a different name: the cochlear implant. It bypasses dead hair cells and stimulates the auditory nerve directly, and hundreds of thousands of people use one. It works because the cochlea carries a clean map — frequency is laid out by position, high at the base, low at the apex. With a map, you know where to write. Visual prostheses have it far worse: stimulating one spot in visual cortex buys one blob of light, and assembling a face from blobs needs resolution and timing precision nobody has yet.
Deep brain stimulation for Parkinson's is a third category: what it writes is not information but a train of high-frequency pulses, acting more like a forced interruption of a pathological synchronised rhythm — effective, with a mechanism still unsettled. basal ganglia
As for where the field actually stands: the academic line (consortia such as BrainGate) has had human implants since 2004, and what those two decades bought is less "can we read it" than can we still read it in five years. Electrodes micromove; the immune response builds a glial scar that pushes electrodes away from neurons; signals fade channel by channel. glia And the sharpest ethical question is not the one outsiders fear — mind-reading — but who owns the thing inside you. A retinal-implant company has already discontinued its product and folded, leaving devices in people's bodies with nobody to support them. An implant is not a phone; you can't upgrade out of it.
Splicing a machine into the loop forces open some very old questions: