A brain implant let a man with ALS communicate independently at home for nearly two years, decoding his attempted speech into text and synthesized speech at 56 words per minute, according to a 2026 study in Nature Medicine — though the results come from just one participant, who needed researcher help for months before using it on his own.
What does the implant actually do?
The device is an intracortical brain-computer interface: four microelectrode arrays, carrying 256 electrodes total, implanted in 2023 in the speech motor cortex of Casey Harrell, a 47-year-old man with ALS who has lost the ability to speak clearly and has weakness in his arms and legs. When Harrell attempts to talk, the arrays pick up the electrical activity of neurons that would normally drive his tongue, lips and vocal cords, and software decodes the intended words onto a screen word by word. At the end of a sentence, he can optionally have it read aloud through a text-to-speech voice trained to sound like his own voice before ALS took it. The same implant also let him control a computer cursor and select on-screen buttons using eye tracking, so he could operate his own computer.
The research is led by a team at the University of California, Davis, working with collaborators at Brown University and the Mass General Brigham Neuroscience Institute, as part of the ongoing BrainGate2 clinical trial. The full findings were published in Nature Medicine on June 15, 2026. The point of this study was not to test whether the decoding worked — an earlier study by the same team had already shown that — but whether it could keep working, day after day, largely without a research team standing by.
How well did it work, and how do researchers know?
Over about 22 months, Harrell used the system in his home for more than 3,800 hours, according to the study and to UC Davis Health, the study's home institution. He communicated more than 183,000 sentences, totaling close to 2 million words, across in-person conversation, video calls, email, text messaging and work. In controlled tests where he was asked to say specific words shown on a screen, drawing on a 125,000-word vocabulary, the system decoded attempted speech with more than 99% word accuracy. Across his everyday, unscripted use, Harrell rated 92% of his sentences as decoded at least "mostly correctly."
Independence built up gradually rather than starting on day one. For roughly the first nine months after implantation, using the system required a research assistant to oversee setup two to four times a week, and Harrell used it about 3.7 hours a day on average. Once regulators allowed his care partners to set up the device without a scientist present, daily use rose to about 9.5 hours a day, and he could initiate sessions independently, whenever he wanted.
"For years, BCIs have been proof-of-concept devices that lived in highly controlled research labs," said David Brandman, one of the study's senior researchers, in the UC Davis release. "This work shows that we may have crossed a threshold, by empowering a person with paralysis to speak on his own terms."
How does this build on the team's earlier results?
The same UC Davis group first reported that its decoder could turn Harrell's attempted speech into text with roughly 97% word accuracy in an August 2024 study in The New England Journal of Medicine, as recounted in Scientific American's coverage of that earlier work — but at the time, Harrell could only use the system when a member of the research team was present to set it up. What the 2026 Nature Medicine paper adds is evidence that a refined version of the same approach can run at slightly higher accuracy while surviving nearly two years of ordinary household use, largely without researchers present.
That progression is the kind of lab-to-market arc that matters for judging any assistive neurotechnology: a decoder that performs well in a research session is a different claim from one that keeps performing after two years in someone's living room, through fatigue, hardware wear and the messiness of real conversation. Sergey Stavisky, another senior author, said in the UC Davis release that the earlier work showed "97% accurate word decoding," but Harrell "could only use the neuroprosthesis when someone from our research team was there to set it up," while the new system is "even more accurate," works faster, and "has been working very well for almost two years."
What can this study not show?
The researchers list their own limitations directly in the paper. This is a single-participant study, and "the generalizability of these results to other individuals, electrode implant sites, intracortical electrode types or neurological conditions is not yet known." The hardware is also far from consumer-ready: it relies on percutaneous wired connections running through the skull, needs daily setup by trained care partners, and its bulky, multi-computer rack limited portability enough that Harrell could only use it inside his home. The researchers note they did not systematically measure user fatigue or long-term wear on the implant, though they say the system's sustained accuracy over time suggests it held up over the study's timescale.
Several of the paper's authors also disclosed financial ties to the neurotechnology industry: one is an inventor on intellectual property licensed to Blackrock Neurotech and Neuralink and has advised or consulted for several BCI companies; others hold patents related to speech BCIs licensed to industry; and Mass General Brigham, a collaborating institution, has received gift agreements from several device makers, including Neuralink, Synchron and Paradromics, that support a BCI research consortium one author works with. The study reports funding from the NIH Office of the Director, the U.S. Department of Veterans Affairs, the U.S. Department of Defense, the ALS Association, the A.P. Giannini Foundation, the Burroughs Wellcome Fund, the Searle Scholars Program and the Achievement Rewards for College Scientists Foundation.
Why this matters beyond one participant
Brain-computer interfaces for speech have mostly been demonstrated inside research sessions, with engineers on hand to fix problems as they came up. The significance here is narrower than "a cure" or "a breakthrough device" — it is evidence that one such system kept working with steadily less outside help over nearly two years, which is closer to the reliability an assistive device would need before it could plausibly move from a clinical trial toward something a person could depend on day to day. Stavisky noted that the trial is also producing what he called, per the UC Davis release, "the largest individual brain recording dataset with single neuron resolution," which researchers are using separately to study how the brain produces speech.
The system remains available only through enrollment in an ongoing clinical trial — the BrainGate2 study is still recruiting participants — rather than as an approved medical device, and questions about eligibility for that kind of trial belong with a treating clinician, not a news article.