BiasLens: A Software Testing Tool for Revealing Hidden Biases in Text-to-Image Models
Abstract: Bias in images generated by Text-to-Image (T2I) models remains a critical concern. These models may unintentionally reflect or amplify societal biases, reinforcing harmful stereotypes and shaping users’ perceptions. This can perpetuate prejudice and discriminatory attitudes as users are unaware of the embedded societal biases. Existing bias detection methods, primarily based on Visual Question Answering (VQA), struggle with complex inputs, particularly those involving spatial elements within images. To address this, we introduce BiasLens, a software testing tool designed to uncover potential biases in T2I-generated images. BiasLens identifies potential biases in user prompts by extracting keywords and leveraging zero-shot and few-shot prompting with a large language model. It then generates and captions images, capturing both visual elements and qualitative descriptions. By analysing adjective-noun pairs in the bias-related phrases of these captions and tracking their frequency, BiasLens provides insight into how biases manifest in generated images. We applied BiasLens to assess biases in depictions of individuals from Southeast Asian countries and Western countries. Our results indicate that BiasLens effectively highlights biases in generated images and reveals key limitations in T2I models. This approach opens new avenues for bias identification and mitigation in AI-generated content, contributing to more responsible and equitable AI systems.
Paper Type: Long
Research Area: Ethics, Bias, and Fairness
Research Area Keywords: model bias/fairness evaluation, text-to-image generation, software and tools, automatic evaluation, bias analysis
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data analysis
Languages Studied: English
Submission Number: 7869
Loading