[ ← INDEX ]

the empty corner in brain-to-text

Meta released Brain2Qwerty v2: non-invasive brain-to-text at 61% word accuracy, and for its best subject, half of all sentences landing within one word of correct.1 no surgery, no electrodes in the brain. it’s a real result, and genuinely hard research.

it also can’t help the people brain-to-text is for, and that isn’t a maturity problem you close with more data. it’s structural. and the reason is more interesting than the headline.

what it actually decodes

start with what’s inside the number. Brain2Qwerty decodes typing. healthy volunteers sit in an MEG scanner — a room-sized, magnetically shielded machine that picks up the faint magnetic fields your neurons throw off, not a headset you’d wear at a desk — and physically type sentences. the model reads the brain activity driving the keystrokes and reconstructs the text.

it reads the motor cortex — the brain driving the hands as they type. not imagined speech. not the words you intend. the movement of hitting keys.

which sets up the problem the headline skips. the people who’d actually reach for a brain-to-text system — someone with ALS, a locked-in patient, anyone who can’t move or speak — are the people who can’t type. the system’s input is the exact thing they no longer have. it works because the subjects can move. the ones who can move are the ones who don’t need it. (which raises the obvious question of who this is really for. I don’t know — but a system that only works for people who can already type is useless in a clinic, and Meta has spent years building neural input for its AR glasses. I have a guess.)

what 61% buys you

then there’s what the number means. “word accuracy” sounds like a typo rate — 39% of words landing a little off. that’s not what failure looks like here.

for the worst subject, the paper reports the output can be “a coherent but entirely different sentence.”1 their own example: the model decoded “had she not fallen down the stairs” when the target was “cars are not allowed on this road.” not garbled, not a near miss — a fluent, grammatical, confident sentence the person never thought.

the authors wave it off: “successful communication relies on meaning rather than strict character matching,” which is true for autocomplete, but a coherent wrong sentence is a meaning failure — the worst kind here, because nothing marks it as an error. garbled output announces that something broke; a fluent, grammatical substitution reads as exactly what the person meant, so it gets taken at face value and passed on. for someone who has no other way to speak, that means going on record saying something they never did. a device meant to give someone back their voice instead hands them a fluent impostor of it.

the machine in the middle

Brain2Qwerty isn’t one model — it’s a three-stage pipeline that ends in a language model. the first stage reads the MEG and guesses letters. the last stage is a fine-tuned Qwen3 that takes those rough letters plus a brain-derived signal and writes the final sentence, one word at a time.

that last stage does a lot of the work, and it’s where the confident-wrong sentences come from. a language model’s job is to turn a noisy, partial input into fluent text — so when the brain signal is thin, it writes something fluent anyway, filling the gaps with whatever reads well. the paper measures the cost: adding the LLM gets more whole words and more meaning right than the raw letters, but it gets more individual letters wrong (a 31% character error rate against 28% without it). it trades literal accuracy for smooth reading. (if you want the full pipeline — encoder, word-level aligner, LLM, and the AI agents that did much of the tuning — I broke it all down separately in how brain2qwerty works.)

it’s structural, and they know it

you could read all of this as early days: more data, more subjects, patient trials, and the gap closes. the paper says otherwise, in its own words.

and it goes deeper than a patient not being able to use the finished system: you couldn’t build it for them in the first place. the whole pipeline is supervised by the key presses, which for a paralyzed person are “missing, not just during inference, but also during training.” the signal you’d learn from is the one movement they can’t make. the authors say as much — patient use “remains critical to demonstrate.”

they file this under future work. it isn’t a next step. it’s the shape of the whole problem.

two axes

to see why the gap is hard rather than just unfinished, lay the field out on two axes.

how close the sensor sits to the signal. outside the skull — MEG, EEG, non-invasive, but reading through bone and tissue that smear everything — or inside it, an electrode array on or in the cortex, invasive but sitting right on the source.

whether the action is carried out or only attempted. an executed action means muscles actually firing, which produces the strongest, cleanest neural signal there is. an attempted action is the intention alone, nothing carried out — a far fainter trace.

put those together and you get four corners. Brain2Qwerty sits in the easy one: non-invasive sensor, executed action. the weakest sensor, but aimed at the strongest possible signal — real typing. that pairing is how you hit 61% without opening a skull.

every system that actually works in a patient sits in the opposite corner: invasive sensor, attempted action. an electrode in cortical tissue, reading an attempt with nothing executed. Stanford’s decoder read a paralyzed man’s attempts to handwrite, letter by letter, off his motor cortex at 90 characters a minute.2 a 2024 implant let a man with ALS hold a live conversation at around 32 words a minute by decoding his attempts to speak.3 both needed an electrode array placed surgically in the brain. the electrode is what buys the signal-to-noise to read an intention that faint. the invasiveness isn’t a wart on those systems — it’s the thing that makes them work.

the empty corner

now look at the corner nobody’s in: non-invasive sensor, attempted action. read what someone is trying to do, from outside the skull, no surgery. that’s the corner that would help the people who need it. it’s empty.

it’s empty because getting there means moving on both axes at once, and both moves cost signal. drop the electrode and you go blind — back to reading through bone. drop executed movement and you halve what’s left — an attempt is fainter than an action. you’d be asking the weakest sensor to read the faintest signal, at the same time. and the cortex you’re reading has often reorganized over years of paralysis, so the map you’d decode from has drifted from the one in the textbook.

the obvious fix is to learn your way across: if you can decode executed movement, why not train a translator from the attempted signal to the executed one, then hand off to the decoder that already works? because training that translator needs the same intent recorded both ways — imagined and performed — and the people you’re building it for can only ever give you the imagined half. you’d fit it on healthy movers and hope it transfers to a cortex that’s spent years reorganizing around the injury, with no way to check, since checking needs the executed signal you don’t have. and even then, an attempt carries less than an execution — there’s no feedback from a limb that never moved — so the translator can’t recover what was never in the signal. it can only make it up. none of which makes it a dead end — this is a real research direction, plausibly the one non-invasive decoding has to take, and maybe where brain2qwerty goes next. i’m not saying it can’t be done. only that it isn’t free: you’d be synthesizing a signal that was never recorded.

here’s where the confabulation comes back. Brain2Qwerty already needs the language model to reach 61% even here, on the cleanest signal it will ever get. take signal away on both axes and the encoder has less to hand over, so the LLM fills more of the gap — and you already know what fills the gap. the assistive corner isn’t only harder to reach. the closer you get, the more the system drifts toward telling the user what they probably meant instead of what they meant.

that’s the trap in one line: the move that would make this useful — dropping executed movement — is the same move that guts the signal it runs on.

the number worth watching

none of this makes the work bad. the accuracy is scaling, non-invasive is the right long bet, and this is patient, unglamorous, actually-hard research.

but the result and the headline point in different directions. 61% is what the easy corner buys. the number worth watching isn’t that one — it’s whatever first crawls out of the empty corner, however bad, because that’s the corner that would matter. the 61% is just the part that fit in a headline.

Footnotes

  1. Zhang, Lévy, Rommel, et al. (Meta AI), “Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings” — the Brain2Qwerty v2 preprint, 29 June 2026. facebookresearch.github.io/brain2qwerty, code at github.com/facebookresearch/brain2qwerty. the v1 result appeared in Nature Neuroscience: nature.com/articles/s41593-026-02303-2. ↩ ↩2

  2. Willett et al., “High-performance brain-to-text communication via handwriting,” Nature, 2021 — nature.com/articles/s41586-021-03506-2. ↩

  3. Card et al., “An accurate and rapidly calibrating speech neuroprosthesis,” New England Journal of Medicine, 2024 — nejm.org/doi/full/10.1056/NEJMoa2314132. ↩