HCII PhD Thesis Defense - Kimi Wenzel
When
-
Where
NSH 4305
Description
Title: Psychological Risks of AI-Mediated Communication Breakdowns
Committee:
Geoff Kaufman, Co-Chair, CMU HCII
Laura Dabbish, Co-Chair, CMU HCII
Jeffrey Bigham, Apple, CMU HCII
Renee Shelby, Google
Michael Muller, Independent
Abstract:
Central to the human experience is the need to be understood. This is true both linguistically (i.e. having the basic semantic meaning of your words comprehended for communicative ease) as well as psychologically (i.e. feeling that you are fundamentally seen and respected). The need to be understood extends to AI conversational agents, such as voice assistants and LLM chatbots. Current conversational agent repair techniques primarily focus on semantic understanding. I argue that this focus on semantic understanding fails to index psychological understanding, and this lack of prioritization can harm vulnerable users. Conversational repair designs that account for psychological understanding have the potential to mitigate psychological and physical harms for users.
In my first set of studies, I consider AI-enabled voice assistants, notably finding that identical AI breakdowns have disparate impacts across diverse users. In a comparison of Black and white users, I find that conversational breakdowns lead to deflated group self-esteem and an inflated sense of self consciousness for Black users, but not white users. A follow-up study on multicultural users revealed a taxonomy of six distinct categories of harm that emerge from AI voice assistant breakdowns. This set of studies showcases how conversational breakdowns are not merely a semantic misunderstanding, but can demonstrate a misunderstanding and disrespect of users' fundamental personhood. This misalignment accounts for the poor wellbeing outcomes Black and multicultural users face in seemingly innocuous conversational breakdowns.
In my second line of work, I evaluate LLM refusals and breakdowns in a help-seeking context. I conduct this work across four different countries to further understand how cultural differences may impact reception of LLM refusals. Results include: exploratory data on user perceptions of AI chatbots, a dataset of 2,034 user chatlogs, and a introductory taxonomy of indirect refusals. This work has the potential to generalize to other morally-sensitive contexts, and has far-reaching implications given recent trends of LLM-use and help-seeking.
Link to dissertation: https://drive.go
Zoom link: https://cmu.zoom.us/j/98