Presentation
Confidence Amplification in Human--AI Moral Decision-Making
SessionWednesday Poster Session
DescriptionLarge language models (LLMs) are increasingly embedded in everyday decision-making, yet their influence on human moral judgment remains poorly understood. Most work on moral influence is
framed as categorical opinion change (e.g., switching between utilitarian and deontological choices). Here we focus on a subtler but consequential pathway: confidence amplification without decision reversals. We conducted a 2 (Time: pre vs. post; within-subjects) x 3 (AI moral framing: utilitarian, deontological, balanced; betweensubjects) mixed design with 120 subjects. We found that AI-assisted dialogue produced reliable confidence amplification with relatively few categorical reversals. A confidence-weighted opinion strength measure shifted toward the AI’s framing. Moreover, alignment between participants’ initial stance and the AI’s framing further amplified confidence, with parallel increases in linguistic certainty during dialogue. Our results suggest that amplification of confidence, rather than directional change alone, may be a central mechanism through which AI systems exert influence.
framed as categorical opinion change (e.g., switching between utilitarian and deontological choices). Here we focus on a subtler but consequential pathway: confidence amplification without decision reversals. We conducted a 2 (Time: pre vs. post; within-subjects) x 3 (AI moral framing: utilitarian, deontological, balanced; betweensubjects) mixed design with 120 subjects. We found that AI-assisted dialogue produced reliable confidence amplification with relatively few categorical reversals. A confidence-weighted opinion strength measure shifted toward the AI’s framing. Moreover, alignment between participants’ initial stance and the AI’s framing further amplified confidence, with parallel increases in linguistic certainty during dialogue. Our results suggest that amplification of confidence, rather than directional change alone, may be a central mechanism through which AI systems exert influence.
Event Type
Poster
TimeWednesday, October 21st5:30pm - 6:30pm PDT
Location