LLMs and VLMs Reproduce Human Biases in Pedestrian Yielding

Summary

A new arXiv study finds that LLMs and VLMs used for autonomous-vehicle decisions can vary pedestrian-yielding behavior according to gender, ethnicity, religion, disability, age, skin tone, and socioeconomic status. The authors introduce “All Else Being Equal” and “Self-Consistency” tests and call for bias evaluation and stronger safeguards.