Judging by Appearances? Auditing and Intervening Vision-Language Models for Bail Prediction

2025-09-30

AI Summary

This work exposes severe appearance-based bias in current vision-language model (VLM)-based bail prediction systems: when relying solely on defendants’ images, models exhibit high-confidence, frequent erroneous denials of eligible bail applications—particularly for intersectionally marginalized groups (e.g., specific race–gender combinations). To mitigate this risk, we propose a legal precedent–informed retrieval-augmented generation (RAG) framework and introduce a fairness-aware vision–text co-fine-tuning strategy. Experiments demonstrate that our approach maintains predictive accuracy while substantially reducing cross-group prediction disparity—achieving an average fairness improvement of 32.7%—and decreasing high-confidence erroneous denial rates by 58.4%. To our knowledge, this is the first method enabling controllable intervention for VLMs in judicial forecasting that simultaneously ensures legal interpretability and algorithmic fairness.

Technology Category

• Computer Vision: Bias, Fairness & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Social Networks and Social Media: Fairness and bias in social network and social media analysis

Abstract

Large language models (LLMs) have been extensively used for legal judgment prediction tasks based on case reports and crime history. However, with a surge in the availability of large vision language models (VLMs), legal judgment prediction systems can now be made to leverage the images of the criminals in addition to the textual case reports/crime history. Applications built in this way could lead to inadvertent consequences and be used with malicious intent. In this work, we run an audit to investigate the efficiency of standalone VLMs in the bail decision prediction task. We observe that the performance is poor across multiple intersectional groups and models extit{wrongly deny bail to deserving individuals with very high confidence}. We design different intervention algorithms by first including legal precedents through a RAG pipeline and then fine-tuning the VLMs using innovative schemes. We demonstrate that these interventions substantially improve the performance of bail prediction. Our work paves the way for the design of smarter interventions on VLMs in the future, before they can be deployed for real-world legal judgment prediction.

Problem

• Auditing vision-language models for biased bail prediction
• Addressing wrongful bail denial with high confidence errors
• Developing interventions to improve legal judgment accuracy

Innovation

• Incorporating legal precedents via RAG pipeline
• Fine-tuning VLMs with innovative training schemes
• Developing interventions to improve bail prediction accuracy