Exposing Hidden Biases in Text-to-Image Models via Automated Prompt Search
Researchers have developed a new framework called Bias-Guided Prompt Search (BGPS) to automatically uncover hidden biases in text-to-image models. This method uses an LLM to generate prompts that, when fed into image generation models, amplify specific attributes like gender or race. Experiments on Stable Diffusion revealed previously undocumented biases, highlighting vulnerabilities in current models and offering a new evaluation tool for bias mitigation efforts. AI
IMPACT This research provides a novel method for identifying and potentially mitigating biases in generative AI, crucial for responsible AI development.