A Proof-of-Concept Prompt Study of User-Controlled Disclosure in LLM Generation for Text-Based BCI’s for Emotional Communication

Background/Objective: Emotional expression remains a significant challenge in text-based brain-computer interfaces (BCIs), which often struggle to convey the nuances of interpersonal communication. This study proposes a seemingly new pipeline that integrates affective BCIs (aBCIs), large language models (LLMs), and user-controlled communication settings. The study was originally intended to prove how the pipeline could improve emotional expression while preserving user autonomy. Instead, other implications were found. This does not include real BCIs, but is instead only a prompt study, examining if the concept works in theory.

Methods: A prompt-based simulation was developed to model a communication system in which hypothetical valence and arousal information, user-selected tone, and disclosure level were provided to a local LLM. All aBCI and P300-related inputs were manually simulated. A comparison was conducted between a baseline model with information regarding emotional state and communication tone, and an enhanced model containing the same information plus a user-controlled disclosure parameter of 0%, 50%, or 100%. Generated sentences were evaluated using XLM-RoBERTa sentiment classification scores, with mean, standard deviation, and Hedges’ g used to characterize the resulting score distributions.

Results: The addition of the disclosure slider produced changes and variance in the sentiment classification scores. The amount and direction of these changes differed across emotional states and communication tones. Formal conditions generally produced more concentrated classifier scores, while humorous conditions showed greater variability in several emotional categories. These findings demonstrate that adding a user-controlled disclosure parameter can change the characteristics of the LLM-generated messages under otherwise same prompting conditions.

Conclusion: This proof-of-concept study demonstrates the potential for user-controlled disclosure to provide an additional level to control emotional expression in LLM-BCI communication, that it can measurably influence generated sentences, and to propose a pipeline to possibly express emotions consistent with what BCI users want.