110 Characters Per Minute, 1.6% Errors: The BCI Typing Benchmark That Changes the Math on Brain Implants
A paralyzed man now types on a virtual QWERTY keyboard using only attempted finger movements. His speed is 61% of the average smartphone user. His accuracy is better.
22 words per minute. A man with a cervical spinal cord injury just achieved that typing speed using a brain-computer interface, a virtual QWERTY keyboard, and zero movement of his hands. He made errors on 1.6% of words. For comparison, smartphone users average 36 words per minute with error rates between 2% and 5%. He is slower. He is more accurate.
Published in Nature Neuroscience on March 16, these results come from BrainGate2, the clinical trial run by investigators at Mass General Brigham and Brown University. Two participants tested the system: one with advanced ALS, one with a cervical spinal cord injury. Both typed from their own homes, not a research lab. Calibration took as few as 30 sentences.
In a field obsessed with raw speed, 22 WPM might look modest. It is not. This is the first time an intracortical BCI has matched able-bodied typing accuracy while sustaining communication speeds fast enough for real conversation.
Your Brain Still Knows QWERTY
A simple insight drives this system: people who typed before their paralysis retain motor cortex patterns for finger movements. Neural signals for pressing 'J' with a right index finger still fire, even when that finger cannot move.
Microelectrode arrays implanted in the motor cortex capture these attempted finger movements. Each letter maps to a specific finger and one of three positions: extending up, pressing down, or curling inward. Ten fingers times three positions gives 30 distinct neural tokens covering a full alphabet plus space and punctuation. A recurrent neural network decodes signals in real time, and a 5-gram language model cleans up output.
Participant T18 had 384 electrodes implanted across both brain hemispheres. Participant T17, who has ALS, had 128 electrodes on one side. Electrode count maps almost directly to speed: T18 reached 110 characters per minute; T17 reached 47.
Speed Is Not Enough
Brain-computer interface coverage tends to fixate on a single metric: words per minute. Neuralink's Noland Arbaugh reaches roughly 40 WPM with cursor control. UCSF's speech BCI hit 78 WPM in a 2023 Nature paper. BrainGate's 22 WPM looks slow by comparison.
But speed alone is the wrong metric for communication. Here is what happens when you multiply speed by accuracy:
| System | Year | Method | Speed (WPM) | Word Error Rate | Correct WPM | Electrodes |
|---|---|---|---|---|---|---|
| BrainGate P300 | 2017 | Point-and-click cursor | ~8 | ~5% | ~7.6 | 96 |
| Stanford/BrainGate Handwriting | 2021 | Imagined handwriting | 18 | 5.9% raw | ~16.9 | 192 |
| UCSF Speech BCI | 2023 | Speech decoding (ECoG) | 78 | 25% | ~58.5 | 253 |
| Neuralink N1 (cursor) | 2024-25 | Cursor + click | ~40 | N/A | N/A | 1,024 |
| BrainGate QWERTY | 2026 | Attempted finger typing | 22 | 1.6% | ~21.6 | 384 |
UCSF's speech BCI delivers 78 words per minute at peak. But one in four words is wrong. For anyone reading that output, it is the difference between a coherent sentence and a frustrating puzzle. BrainGate's 22 WPM with 1.6% errors produces text that reads cleanly on a first pass.
A subtler point hides in these numbers. QWERTY errors cluster between neighboring keys, not between similar-sounding phonemes. Mistyping 'j' as 'k' produces gibberish no language model will auto-correct into a plausible wrong word. Error structure in speech decoding is worse: "their" and "there" are neurally adjacent, making wrong substitutions invisible to correction algorithms.
30 Sentences to Calibrate
Previous BCI typing systems demanded extensive calibration. Stanford's 2021 handwriting BCI needed substantial training data before reliably distinguishing between mentally drawn letters. Speech BCIs require mapping neural activity to thousands of phoneme combinations.
BrainGate's QWERTY system reached usable accuracy after just 30 sentences during calibration. At roughly 10 words per sentence, that is about 300 words of training data. Calibration takes minutes, not hours.
Why so fast? Finger movements are biomechanically discrete. Extending an index finger upward is neurally distinct from curling a ring finger inward in a way that handwritten 'r' and 'n' are not. Motor cortex organizes finger control in spatially separated columns, giving decoders a cleaner signal to learn from.
What 22 WPM Means for 5 Million People
Approximately 5.4 million Americans live with some form of paralysis. Of those, roughly 20% report significant communication difficulties. Eye-gaze systems like Tobii Dynavox, currently dominant in assistive tech, produce 8 to 15 words per minute and cause eye fatigue that limits daily use to a few hours.
| Communication Method | Speed (WPM) | Error Rate | Requires Surgery | Daily Usability |
|---|---|---|---|---|
| Eye-gaze (Tobii Dynavox) | 8-15 | 5-10% | No | Limited by eye fatigue |
| Sip-and-puff switch | 2-5 | ~3% | No | Exhausting |
| BrainGate QWERTY iBCI | 22 | 1.6% | Yes | Used at home daily |
| Smartphone (able-bodied) | 36-38 | 2-5% | No | Unlimited |
At 22 WPM, a BCI user can sustain a real-time text conversation. Eye-gaze at 8 WPM cannot. That is the functional threshold this result crosses: from "augmented communication" to something resembling a normal text exchange.
Convergence Math
BCI typing speeds have followed a roughly exponential trajectory. Point-and-click BCIs in 2012 managed about 2 words per minute. P300 spellers hit 8 WPM by 2017. Handwriting decoding reached 18 WPM in 2021. QWERTY finger decoding now sits at 22 WPM in 2026.
If the trend holds (a substantial assumption), BCI typing would cross smartphone speed (36 WPM) around 2030 and desktop keyboard speed (50 WPM) around 2033. First author Justin Jude noted that stenographic keyboard layouts could push speeds higher without hardware changes, suggesting 22 WPM is a floor, not a ceiling.
But trend extrapolation in neurotechnology is notoriously unreliable. Progress from 2021 to 2026 came partly from adding more electrodes (192 to 384) and covering both hemispheres, a surgical escalation that cannot repeat indefinitely. Whether decoder algorithms alone can push past 30 WPM with identical hardware is an open question.
Strongest Case Against
N=2. Two participants, different conditions, different electrode counts, different hemisphere coverage. T18's 384 electrodes are 3× T17's 128. Attributing speed differences to the decoding paradigm rather than to sheer volume of recorded neural data is not possible from this sample.
Signal stability raises harder questions. Researchers found decoder accuracy degraded after several days without recalibration. Yes, 30-sentence calibration is quick. But needing to repeat it every few days means this system is not yet "set it and forget it." For a medical device patients will use for years, drift management is not a convenience problem. It is the core engineering challenge.
And comparing to speech BCIs is not entirely fair. UCSF's 25% error rate was measured on open-vocabulary continuous speech. BrainGate's 1.6% benefits from a closed-vocabulary language model and self-paced typing, where users can slow down when uncertain. Cross-paradigm comparisons require caution.
Limitations
This analysis relies on the published Nature Neuroscience paper and Brown University's press materials. Several gaps remain:
- Peak vs. sustained speed. That 22 WPM figure is T18's top speed. Sustained averages across sessions are not clearly reported. If sustained rates are 15 WPM, practical comparison to eye-gaze narrows considerably.
- Language model contribution. A 5-gram language model refines output. Raw neural-only accuracy, before correction, is not broken out in press materials. If most heavy lifting happens in post-processing, decoder performance alone may be less impressive than headlines suggest.
- Hardware dependency. Unlike Neuralink's wireless N1, BrainGate's setup involves percutaneous connectors, cables that pass through the skull. This limits mobility and introduces infection risk with long-term use.
- Electrode durability. Intracortical microelectrodes degrade over time as scar tissue forms. Performance longevity over years, not months, is unknown.
- Generalizability. T18's 384 bilateral electrodes represent one of the most extensive intracortical implants in BCI research. Whether this paradigm works with fewer electrodes remains unclear; T17 reached only 47 characters per minute with 128.
Bottom Line
For 22 years, BrainGate has been the patient tortoise of brain-computer interfaces while Neuralink and others grabbed headlines. This result justifies the approach. A typing BCI that matches able-bodied accuracy while sustaining conversational speed is not a demo. It is a minimum viable product for a communication device people with paralysis would actually use every day. Speed gap to smartphones is closing. Accuracy gap has already closed. What remains is brutal engineering: making a research device work for years without recalibration, without wires, and without a neurosurgeon on speed dial. That transition took cochlear implants 20 years. BrainGate has been at it for 22. Whether the next five resemble a cochlear trajectory or a cold fusion one depends on problems no electrode array can solve by itself.