Presenters
Sarah Carter
Kind of session / presentation

Preserving Autonomy: The “Dos” and “Don’ts” of Mental Health Chatbot Personalization

Large language models utilized for basic talk therapy, often referred to as mental health chatbots, are frequently personalized based on user interactions or other input. While personalization could improve the patient’s experience, it could also pose a risk to their autonomy through, for example, the inappropriate use of personalized nudges.

This presentation aims to define the limits of autonomy-preserving personalization in mental health chatbots. To do this, I will use the four-dimensional theory of autonomy (4DT) proposed by Suzy Killmister as a lens to concretely define the myriad of means that autonomy can be promoted and challenged through personalization. 4DT draws on the breadth of autonomy literature to create a comprehensive framework of autonomy into dimensions self-definition (personal identity), self-realization (intention and action), self-unification (consistency), and self-constitution (agency). In addition, the systematic nature of 4DT makes it a helpful tool when trying to design for autonomy in technology.

I will firstly consider self-definition by interrogating the effects of chatbot-based personalization on personal identity formation, awareness, and internal consistency. Next, for self-realization, I will explore personalization-based nudges that encourage certain actions. Similarly, for self-unification, I will analyze the consistency of such actions with patient identity and personal values. Lastly, for self-constitution, I will consider how personalization could cause system overdependence or foster a (renewed) sense of agency. Throughout these dimensions, I will also consider the ramifications of shifting personal values and the challenges presented by patient value tensions.

I then will conclude by proposing initial design recommendations based on these insights, including designing personalization based on patient personal values to promote reflection and value-consistent actions. Using a 4DT approach, this work contributes added insight into how personalization in chatbots could be used in an autonomy-preserving and appropriate manner to support patient well-being, self-determination, and care.