Nearly Two Million Words Through a Brain-Computer Interface at Home

A brain-computer interface has crossed an important threshold: from an impressive demonstration into a tool woven through ordinary days.

A 45-year-old man with amyotrophic lateral sclerosis, or ALS, used an implanted brain-computer interface at home for 3,801 hours over 19 months. He produced 183,060 sentences totaling 1,960,163 words. He used it to talk with family and friends, send messages and email, browse the web, join video calls and sustain full-time employment.

The milestone is not simply that software decoded neural activity. It is that the system stayed useful when the researchers went home.

Research references: [1]

From a research session to daily life

The participant had severe dysarthria and paralysis. Four arrays, each with 64 microelectrodes, had been implanted in the dominant ventral precentral gyrus, part of the speech motor cortex. The arrays connected through the skin to a cart of computers in his home.

Three decoders ran from the same neural signals. A transformer-based model translated attempted speech into text. Two recurrent neural-network decoders translated imagined hand movement into cursor motion and clicks. Eye tracking let him switch controls and use interface buttons.

After training, care partners could set up the system in about 20 minutes. Once it was running, it could operate for as long as 19 hours without further care-partner intervention. Researchers were not present during independent sessions. After independent use was enabled, he used the system on 364 of 397 days.

That pattern matters. Assistive technology is only empowering when it fits into life reliably enough to become a preference, not a special event.

Research references: [1]

Performance that held up at home

During natural use, the participant averaged 56 words per minute. He rated 92 percent of sentences as at least mostly correct. In a structured copy task, the speech decoder reached 99.2 percent word accuracy with a 125,000-word vocabulary.

Those numbers describe different things. The 99.2 percent result came from prompted testing, while the 92 percent rating reflects his own judgment across real-world sentences. Keeping them separate makes the achievement easier to understand: the system was highly accurate under formal measurement and useful enough in everyday communication to support nearly two million words.

Daily calibration also became lighter. The participant could choose between zero and 20 prompted sentences at the start of a day. For cursor control, the newer recurrent model reduced the time to initial control from 2.0 minutes to 0.8 minutes, while average daily calibration fell from 16.9 minutes to 9.2 minutes.

The work was conducted within the BrainGate2 clinical trial, registered as NCT00912041. The research team has also released code and data used to reproduce the paper’s figures and statistics.

Research references: [1] [2] [3]

AI as a carrier of intention

The most human part of this story is not the model architecture. It is what the architecture carried: jokes, work, questions, decisions and the small exchanges that make up a relationship.

That is a useful frame for human-centered AI. The system did not choose what mattered or speak on the participant’s behalf. It translated intended speech and movement into actions he could direct. Correction tools, privacy controls and user-selected calibration kept him in the loop.

The result suggests a design principle beyond brain-computer interfaces: advanced systems create the most agency when they reduce the distance between intention and action while leaving the person in command.

A breakthrough, with a clear boundary

This was one participant. The system was invasive, connected through the skin, required a trained care partner for daily setup and occupied a bulky multicomputer cart. It was not portable, and the findings cannot establish how well the approach will generalize to other people, implant sites, electrode types or neurological conditions.

Those limits do not erase the breakthrough. They locate it. The study does not show a consumer-ready device or a treatment for ALS. It shows that an intracortical interface can support sustained, independent communication and computer control at home—nearly every day, across 19 months.

For a field often measured in carefully staged minutes, 3,801 hours is a different kind of evidence. It is evidence that a bridge between intention and action can remain standing long enough for a person to build life across it.

*Disclosure: This article is independent editorial analysis of a peer-reviewed study, its registered clinical trial and the authors’ public research repository. Next for Humanity has no reported commercial relationship with the study team or device makers. Several study authors disclosed patents, licenses, consulting or advisory relationships in neurotechnology; the full declarations appear in the paper.*

Research references: [1]