Casey Harrell, who lost speech to amyotrophic lateral sclerosis in 2020, has used a BrainGate research coalition brain implant and a machine learning decoder since 2023 to keep his full time job, hitting 92% accuracy at home against 99% in the lab.
Casey Harrell goes to work. He sits at his desk, attends meetings, and replies to messages, all while a brain implant in his motor cortex translates the electrical chatter of his neurons into words on a screen at 92% accuracy. The implant is three years old. The decoder that makes it work — the machine-learning pipeline the UC Davis team calls BRAND — was developed and refined during the study period, and is now running stably in Harrell's home. And Harrell, who has amyotrophic lateral sclerosis (ALS) and lost the ability to speak in 2020, is still using the system every day to hold down a full-time job.
That combination, a brain-computer interface durable enough to live in someone's home for three years and accurate enough to support paid work, is what makes a paper published by a UC Davis team on Monday worth more than a routine research update. According to Nature Medicine, the study — conducted with the BrainGate research coalition and colleagues at Brown University and Mass General Brigham — found that Harrell, a 47-year-old man with ALS, has used the system at home for more than 3,800 hours over nearly two years, communicating more than 183,000 sentences at a self-reported accuracy rate of 92%. In a controlled lab test with a 125,000-word vocabulary, the same decoder hit 99.2% word accuracy.
The gap between those two numbers is the actual story. A 99% lab figure is a familiar headline in the intracortical BCI-for-ALS space: multiple groups have posted eye-catching accuracy numbers in trials over the past several years. What gets reported less often is what happens when a patient takes the device home, uses it for work, and keeps using it for years without a researcher in the next room. A machine-learning model trained in a quiet lab bench has to survive background noise, fatigue, electrode drift, and the slow, ongoing death of motor neurons that defines ALS. By that measure, a 92% self-reported accuracy rate held across nearly two years of daily home use is a more useful data point than a 99% number held for an afternoon.
The framing that puts "AI" at the center of this story belongs to the source, and the source is worth gently correcting. According to UC Davis, the hardware — Blackrock Neurotech microelectrode arrays — is not new. What the UC Davis group contributed is the machine-learning decoding pipeline, called BRAND (Brain-computer Interface for Rapidly Adaptive Neural Decoding), that runs on top of it: a model that reads patterns of neural firing in the ventral precentral gyrus and turns them into words, fast enough to feel like conversation, accurate enough to support employment. The advance is the decoder, not the implant, and Harrell is the subject, not the demonstration.
Co-senior author David Brandman, a UC Davis neurosurgeon and BrainGate site-principal investigator, described prior work as requiring researchers to be in a patient's home or patients to come to researchers. His team's contribution was building a system durable and stable enough for independent home use — what he called "the crossing of a threshold in BCI technology." The 99%-in-the-lab versus 92%-at-home gap is the engineering tradeoff that makes the contribution real, and that tradeoff is what the rest of the field will be watching.
Two honest limits belong in the same paragraph as the result. First, this is a single patient. The n=1 caveat is not a footnote: a decoding pipeline that works for one ALS patient over nearly two years is a proof of concept, not a treatment. Other ALS patients cannot, today, walk into a clinic and ask for a BrainGate array plus a BRAND decoder. The procedure, the device, and the daily calibration are not part of standard neurological care. Second, 92% is not 100%. Roughly one word in twelve comes out wrong on self-report, and the system relies on Harrell attempting to speak in the first place — a function that may itself degrade as ALS progresses. The fact that the decoder has held up this long is genuinely new. The fact that it is also imperfect is also new.
The thing to watch, beyond the paper itself, is durability data on more patients. The UC Davis group's contribution is not the 99% number; it is the 92% self-reported figure that has held for nearly two years in one home, on one desk, in one job. Harrell is described by the study authors as working full-time as an environmental advocate. If the same decoder can be retrained for the next patient, and the next, and the gap between lab and home narrows, the field has something to build on. If it turns out to be a story about Casey Harrell specifically, that is also a story worth reporting carefully, with his name and his work in the lead rather than a percentage.