Part II of VII3 min read
Intent leaves physical traces all the way from cortex to air. The question is which trace you can catch, and where you have to stand to catch it.
When you decide to say a word, neurons in the cortex change their firing. Currents flow. Seconds later, blood supply shifts to feed the active tissue. Commands descend to the muscles of the tongue, jaw and larynx. Tissue moves. Air vibrates. Each stage is a place where an instrument can listen.
A BCI measures brain activity and translates it into commands for a device, giving the brain an output that does not depend on muscle.4 The term was coined in 1973.5 This is an old science with new tools.
Three rules make any result in this field easier to read. Reading is not writing: recording the brain and putting information into it are different problems, and this essay is about reading. Movement is easier than meaning: the strongest results decode attempted handwriting and speech. And the verb matters: seeing, hearing, attempting and imagining are different brain states, so always ask what the participant was doing.
A scalp electrode stands to a neuron roughly as a weather satellite stands to a raindrop. It cannot see the single event. It sees the organised behaviour of millions, blurred by the skull. An implanted electrode is a rain gauge: exact, local, and somebody has to go through bone to install it.
Satellites changed the world and rain gauges did not, for one reason. Satellites cover everyone. EEG, the recording of electrical activity at the scalp, is the satellite of this field. It is a century old, safe, light and cheap enough to build anywhere. That is why it is the instrument this essay keeps returning to.
Seven methods record the brain from different depths; two record the periphery and never touch neural tissue. Select a number to see what each does and does not measure.
EEG · scalp electrodes
fNIRS · light through the scalp
MEG · inside a shielded room
MRI · the scanner fMRI uses
ECoG · grid on the cortex (illustration)
Intracortical array · patent drawingAttempting to write. In 2021 a man paralysed from the neck down imagined writing letters by hand while implanted arrays recorded his motor cortex. The system produced 90 characters a minute at 94.1% accuracy.6
Attempting to speak. In 2023, two implant studies decoded attempted speech at 78 and 62 words a minute, with about one word in four wrong.78 A year later a third held 97.5% accuracy across eight months of daily conversation.9
Attempting to move. People with paralysis have steered robotic arms,20 felt touch written back into the brain,21 and walked again through a bridge between cortex and spinal cord.23 EEG alone has driven a robotic arm in continuous tracking.22
Perceiving. From inside a scanner, models have recovered the gist of a story a person was hearing,11 rebuilt pictures they were looking at,12 and adapted to a new person with one hour of data.13 With EEG and MEG, models can pick out which stretch of speech a listener was hearing.17 These pair a recording with a generative model. The signal supplies the meaning; the model paints the detail.
Typing. With no surgery, sentences have been decoded from magnetic recordings of people typing, at 39% word error.16 The same task on EEG is harder today: two characters in three were wrong.15 We read that as the starting line. Every curve in Part IV says EEG has the most to gain from data.
Mouthing. Away from the brain, silently mouthed words have been turned into speech from facial muscle signals,18 and a wrist sensor trained on thousands of people now works on strangers with no calibration.19
Thinking in words. In 2025, imagined sentences were decoded from four implanted participants. The same study showed how to stop a device decoding inner speech the user had not offered.25
Sixteen studies, each stated the same way: what was measured, what the person was doing, what came out, and what the result does not show.
Signal: intracortical array · Verb: attempting to reach
3-D reach and graspTwo people with long-standing tetraplegia; one drank from a bottle unaided.
Did not show: speed or accuracy near able-bodied movement; use outside supervised sessions.
Signal: intracortical arrays · Verb: attempting to write
90 characters / min94.1% raw accuracy online; >99% offline with autocorrect.
Did not show: 90 words per minute; more than one participant; multi-year stability.
Signal: intracortical recording + stimulation · Verb: attempting to grasp
20.9 s → 10.2 sMedian task time with a robotic arm, without and with evoked tactile sensation.
Did not show: rich or naturalistic sensation; writing of language or concepts.
Signal: ECoG grid · Verb: attempting to speak
Median 78 words / minMedian 25% word error rate; text, synthetic voice and avatar.
Did not show: error rates low enough for unassisted daily use; decoding of unattempted thought.
Signal: intracortical arrays · Verb: attempting to speak
62 words / min23.8% word error on a 125,000-word vocabulary; 9.1% on 50 words.
Did not show: conversational accuracy (see 2024); results beyond one participant.
Signal: intracortical arrays · Verb: attempting to speak
97.5% word accuracySustained over 8.4 months; ~32 words / min; 248+ hours of self-paced use; 30-minute first calibration.
Did not show: generalisation across patients; performance without ongoing recalibration.
Signal: intracortical arrays · Verb: imagining speech
Imagined sentences, real timeInner speech is represented in motor cortex; some unprompted inner speech was decodable; safeguards blocked it.
Did not show: decoding of arbitrary private thought; anything outside implanted motor cortex.
Signal: fMRI · Verb: listening (also imagining, watching)
Meaning, not wordsGenerated word sequences that recover the sense of heard speech.
Did not show: word-for-word transcripts; decoding without the person’s cooperation, which was required to train and to apply the decoder.
Signal: 7-tesla fMRI · Verb: looking at pictures
Retrieval and reconstructionIdentifies the viewed image among similar candidates; generates a resembling image with a diffusion model.
Did not show: imagined or remembered images; exact readout. The generator supplies detail the recording lacks.
Signal: 7-tesla fMRI · Verb: looking at pictures
~1 hour to adaptAfter pretraining across 7 other people in the same dataset.
Did not show: zero-calibration transfer; transfer across scanners, tasks or to portable sensors.
Signal: fMRI · Verb: watching video
Clips up to 6 s at 8 fpsResembling the watched footage, via a pretrained text-to-video model.
Did not show: a live feed, imagined scenes, or frame-accurate recovery.
Signal: MEG and EEG · Verb: typing memorised sentences
32% vs 67% character errorAverage for MEG vs EEG. Best MEG participants: 19%.
Did not show: free-form thought to text. Participants were physically typing; EEG got two characters in three wrong.
Signal: MEG, real time · Verb: typing natural sentences
39% word error rateAverage; improves log-linearly with data. Best participant: half of sentences with at most one wrong word.
Did not show: anything wearable (MEG needs shielding); peer review is pending; ~10 hours of recording per person.
Signal: EEG, MEG, 3T and 7T fMRI · Verb: looking at pictures
498 hours, 84 peopleThe most precise devices decode best. No plateau with more data. Hours per person matter more than more people.
Did not show: where any device’s ceiling lies; anything about language or intent.
Signal: facial surface EMG · Verb: mouthing words silently
64% → 4% word errorClosed-vocabulary condition. Open vocabulary: 88% → 68%.
Did not show: brain decoding of any kind; usable open-vocabulary accuracy; multi-speaker generalisation.
Signal: wrist surface EMG · Verb: gesturing and handwriting
20.9 words / minHandwriting, with no per-user calibration; +16% with personalisation.
Did not show: speech or brain decoding. Requires functioning hand muscles.
The laboratory has proved that the information exists. It has not yet proved that it travels.
Download the full manifesto as a PDF
Setu Glasses · Setu Cap