Skip to main content
Science Catalogs
Physical Sciences and Engineering

AI phrase suggestions speed up a noninvasive brain-computer speller in tests

Engineers at Rutgers University have built a noninvasive brain-computer interface, ChatBCI-Assist, that uses a language model to suggest words and whole phrases to the user. In tests with healthy participants, it roughly doubled communication speed when copying sentences and came close to tripling it when users conveyed only the meaning of a message, compared with a conventional dictionary-based predictive system. The work appears in IEEE Transactions on Biomedical Engineering.

How the speller works

The system does not decode unrestricted thoughts. The user wears a cap of sensors that records electrical activity through the scalp and looks at a screen showing a keyboard together with suggested words and phrases. Its rows and columns flash in rapid succession. When the row or column holding the item the user wants flashes, the brain produces a P300 response, a signal that typically appears about 300 milliseconds after an attended stimulus. By matching the timing of that response to the flashing pattern, the computer infers which letter, word or phrase was chosen.

P300 spellers have existed for decades, but selecting each letter in turn is slow. Writing a short request such as "I want water" can demand many separate selections.

Intent rather than exact wording

ChatBCI-Assist adds a large language model, the general type of AI behind chatbots, to propose contextually relevant words and completions. The team adapted an existing model using examples of messages that reflect the needs of people with ALS. The suggestions draw on what the user has already selected and on the ongoing conversation.

A user hoping to say "I want water" could accept a suggestion such as "Give me water". The wording differs but the meaning is close, so fewer selections are needed. The user stays in control and can accept a suggestion or carry on spelling letters.

The model runs locally on the computer rather than through a cloud service. Messages about health, physical needs or private life therefore need not travel over the internet, and the system can operate without a reliable connection.

What the tests showed

The researchers evaluated the system in two tasks. In copy spelling, participants reproduced a given sentence exactly. In semantic spelling, they had to convey the same meaning, which let the AI offer shorter or differently worded phrases. Against the dictionary-based baseline, speed was nearly twice as high in copy spelling and nearly three times as high in semantic spelling.

Participants also reported a lower workload with the AI-assisted system and rated the intent-based mode as easier to use than the baseline. The team additionally adopted evaluation measures that go beyond conventional typing accuracy, which they say broadens how intent-based BCI spellers can be assessed.

Limitations and next steps

The key caveat is that only healthy participants have used the system so far. The source gives no participant numbers, and the results cannot yet be assumed to carry over to people with advanced ALS, whose signals, fatigue and needs may differ. Clinical testing is being planned with Shan Chen, medical director of the ALS Multidisciplinary Clinic at Rutgers Robert Wood Johnson Medical School, to examine practicality and comfort.

The motivation is that many people with advanced ALS retain their cognitive abilities but lose the motor control needed to speak or type. Eye-gaze devices are a common aid, but they can become hard to use if reliable eye control is lost. The study was led by Laleh Najafizadeh, with Weinan Wang as co-lead, first author Jiazhen Hong and graduate student Pradyumna Rao.

For BCI researchers, the work illustrates a shift in design philosophy: rather than reproducing every word, assistive systems can predict what a user means and let them confirm it. Whether the speed gains and lower workload persist in patient groups, and over long everyday use, will be the central questions for the planned trials.

Source: Noninvasive brain interface uses AI suggestions to convey messages nearly three times faster (Tech Xplore, 21 November 2026).

Image: AI-generated illustration.

Working on this topic? Find a journal for your research or join the discussion on Scholars Post.