Never say never. exploring the effects of knowledge availability on agent persuasiveness in controlled physiotherapy motivation dialogues

Abstract

Generative Social Agents (GSAs) have the capability to influence their human users through persuasive communication. On the one hand, they might motivate users to pursue positive personal goals, such as following a healthier lifestyle. On the other hand, GSAs are linked to negative outcomes like manipulation and deception. These emerge as a consequence of the fact that we only have limited control over probabilistic agent outputs. However, at the same time, GSAs manifest communicative patterns based on available knowledge. Therefore, their communication behavior can be shaped by regulating their access to such knowledge. Following this approach, we explored persuasive messages from GSAs in the context of human-robot physiotherapy motivation. We did this by comparing ChatGPT-generated responses to predefined inputs from a hypothetical patient in physiotherapy. In Study 1, we qualitatively analyzed 14 ChatGPT-generated dialogue scripts with varying knowledge configurations. In Study 2, third-party observers ( N = 27) rated a selection of these scenarios in terms of the agent's expressiveness, assertiveness, and persuasiveness. Our findings indicated that LLM-based GSAs can adopt assertive and expressive personality traits, thereby significantly enhancing perceived persuasiveness. Moreover, persuasiveness improved when information about the patient's age and past profession was available, mediated by perceived agent assertiveness and expressiveness. Context-related knowledge, e.g., regarding benefits associated with physiotherapy did not significantly impact agent persuasiveness. This might be due to the fact that the LLM we used already included such information from pre-training. Overall, the present research highlights the importance of studying autonomous GSA behavior from an empirical perspective. Particularly, future research should focus on the information that is required in order to enable and assure coherent and responsible communication with generative AI systems.