Towards Neuroadaptive Chatbots: A Feasibility Study

Diana E Gherman^1,2*Thorsten O Zander^1,2

• ¹Brandenburg University of Technology Cottbus-Senftenberg, Senftenberg, Germany
• ²Zander Laboratories GmbH, Cottbus, Germany

Introduction: Large-language models (LLMs) are transforming most industries today and are set to become a cornerstone of the human digital experience. While integrating explicit human feedback into the training and development of LLM-based chatbots has been integral to the progress we see nowadays, more work is needed to understand how to best align them with human values. Implicit human feedback enabled by passive brain-computer interfaces (pBCIs) could potentially help unlock the hidden nuance of users’ cognitive and affective states during interaction with chatbots. This study proposes an investigation on the feasibility of using pBCIs to decode mental states in reaction to text stimuli, to lay the groundwork for neuroadaptive chatbots. Methods: Two paradigms were created to elicit moral judgment and error-processing with text stimuli. Electroencephalography (EEG) data was recorded with 64 gel electrodes while participants completed reading tasks. Mental state classifiers were obtained in an offline manner with a windowed-means approach and linear discriminant analysis (LDA) for full-component and brain-component data. The corresponding event-related potentials (ERPs) were visually inspected. Results: Moral salience was successfully decoded at a single-trial level, with an average calibration accuracy of 78% on the basis of a data window of 600ms. Subsequent classifiers were not able to distinguish moral judgment congruence (i.e. moral agreement) and incongruence (i.e. moral disagreement). Error processing in reaction to factual inaccuracy was decoded with an average calibration accuracy of 66%. The identified ERPs for the investigated mental states partly aligned with other findings. Discussion: With this study, we demonstrate the feasibility of using pBCIs to distinguish mental states from readers’ brain data at a single-trial level. More work is needed to transition from offline to online investigations and to understand if reliable pBCI classifiers can also be obtained in less controlled language tasks and more realistic chatbot interactions. Our work marks preliminary steps for understanding and making use of neural-based implicit human feedback for LLM alignment.