CM Seminar - Talk, Listen, Connect: Navigating Empathy in Human-AI Interactions

Presenter: Mahnaz Roshanaei

Description: Research consistently shows that social interactions, particularly face-to-face engagement, promote well-being, however barriers such as time, distance, and health conditions often limit access. In the last years, AI-driven chatbots have emerged as supplementary tools to facilitate social, emotional, and relational support. Yet there are still concerns about the psychological impacts of AI-mediated interactions, especially as emotional complexity increases, in particular around Empathy. In this study we investigate how empathy is evaluated in AI compared to humans and examine the role of model fine-tuning, and persona attributes.

Using the 756 human-generated narratives from 126 students, we evaluated the empathic responses from Amazon MTurkers and GPT-4o. Our findings indicate that GPT-4o rates empathy higher and with less variability than humans, but lacks the depth found in human emotional understanding. Furthermore, fine-tuning GPT-4o using human-annotated ratings and incorporating persona attributes improved GPT-4o’s alignment with human empathy judgements. Yet, AI systems still risk overstating emotional resonance and failing to authentically capture human variability. These findings highlight the need for cautious design of empathic AI, particularly in sensitive domains like mental health support.

Bio: "I’m a computational social scientist with broad interests in social behavior, digital media effects, human-AI interactions, and well-being. I am a research staff in Media and Personality Lab at the Stanford University, working with Gabriella M. Harari. I’m also Visiting Scholor at GUII lab, working with Magy El-Nasr, and Lecturer in Computational Media department at University of California, Santa Cruz. I received my PhD degree in Computer Science from University of Colorado Boulder. My research bridges my training in computer science and social science (e.g., communication, psychology) to answer substantive research questions about social behavior, media and AI effect, human-AI interaction, and well-being. I have two complementary lines of research focused on (1) Understanding and analyzing the effects of social interactions on people’s psychological experiences in both offline and online settings, and (2) Leveraging data science techniques to model and predict psychological outcomes based on digital media usage and daily behavioral data. In conducting my research, I employ a broad range of interdisciplinary methods in computational social science, including machine learning (ML), data mining, large language models (LLMs), network science, statistical and causal inference, smartphone sensing, experience sampling, multilevel modeling, and mixed methods to analyze data and uncover patterns that offer insights into the complex interplay between human behavior, technology, and psychological outcomes."

Hosted by: Professor Elin Carstensdottir

IMPORTANT: There will be a remote viewing room at UCSC Main Campus in Room E2-280.