In the spring of 2021, at Stanford University, a trial participant known in the published literature only as T5 sat in front of a monitor with two microelectrode arrays sitting in his precentral gyrus and imagined writing letters with a hand that hasn’t moved since 2007, the year a spinal cord injury left him paralyzed from the neck down. He wasn’t steering a cursor toward an onscreen keyboard, the standard BCI typing method at the time. He was imagining the actual physical act of handwriting, letter by letter, and a decoder was reading the attempt straight out of his motor cortex. The system converted that imagined handwriting into text at 90 characters per minute. It was, at the time, the fastest any paralyzed person had ever communicated through a brain-computer interface, more than double the previous record.
That number, 90, is real. It comes from Frank Willett, Krishna Shenoy, and Jaimie Henderson’s team at Stanford, published in Nature in May 2021 under the title “High-performance brain-to-text communication via handwriting.” What is not quite real, or at least not quite accurate, is the version of that number that has circulated since: ninety words per minute. Ninety characters per minute, run through the typographer’s standard conversion of five characters per word including spaces, comes out to about 18 words per minute. That is still remarkable for a man who cannot move his hands. It is not the number in most of the headlines that followed the paper, this one included, because I was asked to write to it and because it is, at this point, the number people believe. I think the gap between what was measured and what got repeated is worth sitting inside for a minute, because it tells you something about how BCI results travel and about what actually happened in that room.
Ninety characters per minute is not ninety words per minute, and the gap is the whole story.
Ninety Characters Is Still the Record That Matters
Before the handwriting study, the fastest documented intracortical typing speed belonged to the same research group, using the same participant, with a different method. In 2017, Chethan Pandarinath and colleagues, working across the Stanford and Brown BrainGate sites under Shenoy’s direction, published a point-and-click typing study in eLife. Participants steered a cursor to letters on a virtual keyboard the way you’d use a trackpad, one target acquisition at a time. That system topped out around 40 correct characters per minute, or roughly 8 words per minute. It was itself a milestone, the fastest communication a paralyzed BCI user had achieved to that point, and it took over a decade of BrainGate trial work to get there, going back to the first human implant in 2004 under Cyberkinetics Neurotechnology Systems and the 2006 Nature paper by Leigh Hochberg’s group describing a man named Matt Nagle moving a cursor with thought alone.
The handwriting study more than doubled that 2017 record by changing what the decoder was asked to read. Instead of decoding an abstract intention (move cursor northeast), it decoded a motor pattern the brain already had a rich, practiced representation for: the shape of individual letters, something T5 had been producing with a pen for fifty years before his injury. Motor cortex, it turns out, still encodes those handwriting-specific trajectories in vivid detail even when the hand producing them is gone. Willett’s team trained a recurrent neural network to recognize the neural signature of each letter as it was attempted, then fed the output through a language model that could clean up ambiguous decodes using context, the same basic principle as autocorrect on a phone keyboard. Raw accuracy on unconstrained sentences came in at 94.1 percent. With the language model correcting likely errors, accuracy climbed above 99 percent.
Handwriting Beat Pointing, and That Surprised the Field
The obvious assumption going into this line of research was that pointing would always be faster than handwriting, because pointing is simpler. You pick a target, you move toward it, you click. Handwriting requires the brain to produce a continuous, curved trajectory with variable pen pressure and stroke order, then requires the decoder to disentangle two dozen different letter shapes from a noisy population of a few hundred neurons. It seems like it should be the harder decoding problem. It turned out to be the faster one, because it front-loads more information into every motor act. A single letter carries more bits than a single point-and-click selection, and the brain, given a task it already had decades of motor memory for, produced a cleaner, more separable signal than it did for the comparatively artificial task of aiming a cursor.
He was writing letters with a hand that hasn’t moved since 2007.
That’s the part of the story that gets lost when the number gets rounded up to “90 words.” The achievement wasn’t raw speed for its own sake. It was a demonstration that decoding the brain’s existing, well-rehearsed motor vocabulary beats decoding an artificial one bolted on for the purpose of the experiment. Krishna Shenoy, who died in 2023 and who had led BCI decoding research at Stanford since the early 2000s, spent much of his career making exactly this argument: that the path to faster, more natural BCIs runs through finding movements the brain already knows how to do, not inventing new ones for it to learn.
Speech Has Already Lapped Handwriting
Two years after the handwriting paper, two separate speech decoding studies made 90 characters per minute look almost quaint. In 2023, Edward Chang’s group at UCSF published a speech neuroprosthesis for a woman named Ann, paralyzed by a brainstem stroke, that decoded attempted speech into text and synthesized speech at 78 words per minute, using electrodes placed over speech-motor cortex rather than the hand area. The same year, Willett’s Stanford group, now working with Erin Kunz and with Henderson still leading the surgical side, published a speech decoding study with a participant named Pat Bennett, who has ALS. That system reached 62 words per minute with a vocabulary above 9,000 words, translating attempted speech, not attempted handwriting, into text.
Speech has an inherent throughput advantage over handwriting or typing, because spoken language is simply denser per unit time than written motor output. But it also depends on having enough intact speech-motor cortex left to read from, which is not guaranteed in every condition that causes paralysis, particularly brainstem injury. Handwriting and cursor-based systems remain the fallback, and in some cases the primary option, for patients whose paralysis originates lower in the motor pathway. The field is not converging on one modality. It’s building several, and matching them to where in the nervous system the injury actually sits.
| Year | Team | Decoded output | Speed |
|---|---|---|---|
| 2006 | Hochberg / Cyberkinetics, BrainGate1 | Cursor movement | Proof of concept, not benchmarked for text |
| 2017 | Pandarinath / Shenoy, Stanford & Brown BrainGate | Point-and-click typing | ~40 characters/min (~8 words/min) |
| 2021 | Willett / Shenoy / Henderson, Stanford | Imagined handwriting | 90 characters/min (~18 words/min) |
| 2023 | Metzger / Chang, UCSF | Attempted speech | 78 words/min |
| 2023 | Willett / Kunz / Henderson, Stanford | Attempted speech | 62 words/min, 9,000+ word vocabulary |
None of This Runs Without a Connector Bolted to a Skull
Every number in that table came out of a hardwired system. T5’s arrays connect through a percutaneous pedestal, a port fixed to the skull that a cable plugs into, tethering him to a rack of amplifiers every time the system runs. The decoder has to be recalibrated within a session because the tiny, day-to-day drift in how a few hundred neurons fire changes the mapping the software relies on. None of this exists as a product a patient takes home and uses unsupervised. BrainGate has run as an investigational device exemption study since 2004, across sites at Stanford, Brown, Massachusetts General Hospital, and the Providence VA, and it remains a research trial, not an approved therapy.
That gap between demonstration and deployment is where the rest of the industry is trying to plant a flag. Synchron, led by Tom Oxley, threads its Stentrode electrode array up through the jugular vein into a blood vessel against the motor cortex, avoiding open brain surgery entirely, and has been running feasibility trials in ALS and paralysis patients since 2022. Neuralink implanted its first human participant, Noland Arbaugh, in January 2024, demonstrating cursor control and gameplay, though without the kind of peer-reviewed, head-to-head speed benchmark that the Stanford and UCSF papers put through formal trial protocols. Blackrock Neurotech, which manufactures the Utah arrays used in most of the academic work described here, is building toward a wireless, fully implanted version that would eliminate the pedestal altogether. The race now is less about beating 90 characters per minute and more about who can deliver a comparable number without a cable coming out of someone’s head.
The Point Was Never the Cursor
I keep coming back to what changed between 2017 and 2021, because it’s a cleaner story than the headline number. The field spent its first decade treating the brain as a joystick, something you decode movement out of and then map onto a task, cursor, keyboard, robot arm. The handwriting study, and the speech studies that followed it, treat the brain as something closer to its actual function: a system that already knows how to produce language, if you can just find the neurons still carrying that instruction and read it directly. The cursor was never the point. Language was always the point, and it took fifteen years of BCI research to build a decoder good enough to go get it.
Credit: Bhautik Patel on Unsplash