Engineers at Rutgers University (USA) have presented ChatBCI-Assist, a noninvasive brain-computer interface that uses artificial intelligence (AI) to translate brain activity into words and phrases. The development allows communication almost three times faster than conventional systems in which text must be typed letter by letter. The results were published in the journal IEEE Transactions on Biomedical Engineering.
The noninvasive interface requires no surgery: the user wears a cap with sensors that pick up electrical signals from the scalp. The system does not read free thoughts; it determines which specific element on the screen the person has consciously focused on. A keyboard and word and phrase suggestions are displayed on the monitor; rows and columns flash rapidly. When a row or column containing the desired choice flashes, the brain produces an electrical response, the so-called P300. This signal appears about 300 milliseconds (thousandths of a second) after a person notices a stimulus important to them. The sensors record the response, and the computer uses the timing of the flash to determine which letter, word or phrase the user has selected.
Classic P300-based spellers have existed for decades but work slowly: to write "I want water," one has to select each letter. ChatBCI-Assist speeds up the process through a large language model, the same type of AI used in chatbots. The model was further trained on examples of messages reflecting the needs of people with amyotrophic lateral sclerosis (ALS), a progressive disease in which the nerve cells that control muscles die. Based on the user's choice and the context of the conversation, the AI suggests ready-made formulations. For example, instead of typing "I want water" letter by letter, one can select the suggestion "Give me water": the meaning is the same, but less effort is required.
"We are calling for a change of perspective," says the head of the study, Laleh Najafizadeh, a professor at the Rutgers School of Engineering. "Instead of reproducing every word, assistive systems can predict intent and help express it more efficiently." The user remains in control: they decide whether to accept a suggestion or continue typing text. The language model runs locally on the computer rather than in the cloud, so personal messages about health and private life are not transmitted over the internet; the system can also work without a stable connection.
In experiments with healthy participants, ChatBCI-Assist was tested in two modes. In the first ("copying"), the task was to reproduce a sentence exactly; the speed was almost twice as high as with a traditional system using dictionary-based prediction. In the second ("semantic typing"), the task was to convey the same meaning while allowing the AI to suggest shorter or different formulations; the speed increased almost threefold compared with the baseline. Participants also noted that the system reduces workload and that the AI mode is subjectively easier.
So far, the trials have been conducted only with healthy volunteers. Clinical tests with ALS patients are in the planning stage: Najafizadeh is collaborating with Dr. Shan Chen of the Rutgers University medical clinic. The scientists emphasize that for people with advanced ALS who retain their intellect but have lost muscle control, such systems could become a way to express their thoughts. Traditional eye-tracking devices become difficult to use over time if the eyes stop obeying. The new development is a step toward faster and more accessible communication, but a finished product is still a long way off.