LLM-Driven Autonomous Vehicles Inherit Human Driver Biases in Pedestrian Yielding: Results and Implications From A New Benchmark

Research indicates that Large Language Models (LLMs) and Visual-Language Models (VLMs) used in Autonomous Vehicles (AVs) can inherit human driver biases, specifically in pedestrian yielding behavior. This raises concerns about the fairness of AV decision-making, which is crucial for public trust. New methodologies are proposed to benchmark and address these biases in AV evaluation.

Why it matters

The potential for AI systems in autonomous vehicles to perpetuate human biases represents a significant ethical and operational risk. Addressing these biases is critical for ensuring equitable outcomes, maintaining public trust, and safeguarding the long-term viability and adoption of AV technology in diverse societies.

Key insights

• LLM-driven AVs can inherit human driver biases, impacting decision-making fairness.
• Psychology studies have documented human driver biases, such as lower yielding rates to Black pedestrians in the US.
• The integration of 'common sense' models in AVs may perpetuate existing societal biases if not properly addressed.
• Public trust in AV technology is contingent upon both technical success and fair decision-making.
• Two new bias testing methodologies, 'All Else Being Equal' and 'Self-Consistency' tests, are proposed for LLMs and VLMs.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.00192