Presentation
Assessing AI Social Aptitude Using Crowd Sourced Cognitive Labels
DescriptionIn the pursuit of effective human-AI cooperation, understanding and aligning with human social cognition is critical. This paper investigates the capacity of algorithms to imitate human social behaviors, particularly in cooperative scenarios such as automated cars interacting with pedestrians. While existing studies have developed prediction models for pedestrian intentions, the subjective nature of cognitive labels and temporal variations pose challenges to accurately imitate human socialization. We introduce the Human-Like index (HL-i) as a novel metric to assess algorithmic social aptitude during interaction processes. Through empirical experiments and analyses, we address three research questions regarding AI social performance and human discernment of AI-generated outputs. Findings reveal significant variability in algorithmic mimicry of human behavior across scenarios, with HL-i serving as a crucial indicator of alignment with human cognitive patterns. The study underscores the complexity of human social cognition and proposes HL-i as a valuable tool for enhancing algorithmic social aptitude.
Event Type
Lecture
TimeWednesday, September 11th11:35am - 11:55am MST
LocationFlagstaff
Human AI Robot Teaming (AI)