Rutgers University announced on November 20, 2026 a noninvasive research system called "ChatBCI-Assist," which uses brain waves measured from the scalp to select letters and phrases on a screen. In an experiment with 10 healthy participants, the system used predictions from a large language model (LLM) to reach an average of 19.7 characters per minute when participants had to reproduce a specified sentence exactly, and an average of 30.7 characters per minute when they only had to convey the same meaning. However, because the two tasks differ in what counts as a completed sentence, the gap between the figures cannot be read directly as an improvement in system performance.

The results appear in the August 2026 issue of IEEE Transactions on Biomedical Engineering as a peer-reviewed paper by Jiazhen Hong, Pradyumna Rao, Weinan Wang, and Laleh Najafizadeh. The team also proposes a new evaluation method that measures not just typing speed but how efficiently users convey the meaning they intend.

AD

How the system selects on-screen options from scalp EEG

ChatBCI-Assist is a brain-computer interface (BCI) that measures brain waves (EEG) with sensors worn on the scalp and uses the signals to operate a computer. The user wears an EEG cap and focuses attention on the letter or phrase they want from those shown on screen. Rows and columns of letters and phrases flash in sequence, and when a row or column containing the item the user is attending to lights up, a brain response known as the "P300" typically appears about 300 milliseconds later. The system matches this response against the timing of the flashes to infer which item the user meant to select.

Conventional P300 methods require users to specify letters one at a time to build a sentence, so entering long text means repeating many operations. In ChatBCI-Assist, an LLM proposes words and multi-word phrases suited to what has been entered so far and to the context, in real time. If the desired expression is among the suggestions, a single selection can enter multiple characters. If no suitable suggestion appears, the user can select the next letter and continue typing. The system also introduces an "adaptive stopping" method that adjusts the number of flash repetitions needed to decide on a selection from brain waves, aiming to make operation more efficient.

Using data collected from expressions that people with amyotrophic lateral sclerosis (ALS) use in daily life, the team further trained the existing LLM "Llama" with a technique called LoRA (Low-Rank Adaptation). The model runs on the user's own computer, so there is no need to send input to an external cloud service when generating suggestions. As an example, the university describes a user who wants to say "I want water" and selects the AI-suggested expression "Please give me water." The idea is that users can communicate their intended request without reproducing a sentence word for word.

Verbatim transcription and meaning transfer have different task conditions

According to the README of the experiment logs that the team published on GitHub, the experiment was conducted in 2025 with 10 healthy participants. All participants operated the device in real time and tried each of three conditions once. The order of conditions was randomized for each participant, and in each condition they worked on three tasks. Two involved entering sentences not used to train the LLM, and in the third, participants composed their own sentences on a randomly assigned topic. Each task had a 10-minute limit, and the log analysis treats the last record within the time limit as the end point.

The experiment compared conventional dictionary-based prediction with LLM-based prediction, and also examined the difference between "verbatim transcription," reproducing a sentence exactly, and "meaning transfer," conveying the same meaning in different wording. In the dictionary condition, candidates are narrowed down from 846 words contained in the LLM's training data, and six are shown on screen. In the LLM condition, multi-word phrases as well as single words are displayed as candidates.

Condition How candidates are generated Task success criterion Average typing speed
Dictionary-based verbatim transcription Word prediction from a dictionary Exact match with the original text Absolute value could not be confirmed
LLM-based verbatim transcription Word and phrase prediction by the LLM Exact match with the original text 19.7 characters/min
LLM-based meaning transfer Word and phrase prediction by the same LLM Conveys the same meaning (paraphrasing allowed) 30.7 characters/min

Average typing speeds are based on the paper abstract posted on the Rutgers research portal.

In both the dictionary-based and LLM-based verbatim transcription conditions, a complete match with the specified sentence was required. However, the LLM condition allowed multi-word phrases to be entered at once, and participants were also encouraged to use phrase candidates actively. The comparison between the two conditions therefore reflects not only the prediction method but also differences in how many words can be entered at a time and in instructions about operation.

Between the two conditions that used the same LLM, verbatim transcription averaged 19.7 characters per minute, while meaning transfer, which allows paraphrasing, averaged 30.7 characters per minute. According to the team's report, using AI-generated suggestions and not requiring exact reproduction of a sentence let participants convey their request with fewer operations. These results alone, however, do not show how much LLM-based candidate generation and adaptive stopping each contributed to typing speed. Nor can the published abstract and the experiment-log README confirm how speeds were tabulated in detail or how much they varied between participants.

AD

Measuring conveyed meaning and operational burden

BCI performance has traditionally been evaluated with measures such as letter-selection accuracy, characters typed per minute, and information transfer rate. But when communication through paraphrasing is allowed, the number of characters typed is not enough to assess real communication efficiency. If the same request can be conveyed with a shorter phrase, the goal can be achieved with fewer characters.

The team therefore proposed the "Semantic Information Transfer Rate (SITR)," which incorporates the semantic closeness between the entered text and the target text. According to the paper abstract, ChatBCI-Assist's SITR was 147.1 bits per minute. The average estimated information transfer rate (ITR) was reported as 105.2 bits per minute, and the overall character-level mutual information rate (MIR) as 52.9 bits per minute. All use bits per minute as the unit, but because what they evaluate and how they are calculated differ, the figures cannot be compared directly. In particular, the public abstract does not make clear the specific procedure by which SITR calculates semantic closeness between sentences.

After each condition, the team also used questionnaires such as NASA-TLX to survey the operational burden and ease of use that participants felt. According to the university's announcement, participants reported that AI assistance reduced the burden of operation and rated the meaning-transfer method as easier to use than dictionary-based verbatim transcription. Based on the paper title and DOI, no independent replication under the same conditions has been reported.

As of the November 20, 2026 announcement, evaluation of ChatBCI-Assist was limited to healthy participants, and clinical evaluation in people with ALS is still in the planning stage. Laleh Najafizadeh, the Rutgers professor leading the research, plans to work with Shan Chen, medical director of the ALS clinic at the university's Robert Wood Johnson Medical School, to examine practicality and comfort when patients actually use the system. Whether patients can select candidates that match their intentions, and whether the burden of prolonged operation can be kept low, will be key to assessing its use in daily life.