Aligned but Stereotypical? How System Prompts Shape Demographic Bias in LLM-Based Text-to-Image Models
A new research paper explores how Large Language Model (LLM)-based text conditioning in text-to-image (T2I) models can introduce demographic biases, even when demographic attributes are not specified in prompts. The study found that LLM-based systems exhibit stronger demographic skew compared to non-LLM baselines. Researchers identified system prompts as a key factor influencing these biases and proposed FairPro, a framework designed to generate fairness-aware instructions to mitigate disparities while preserving user intent. AI
IMPACT Highlights potential demographic biases in generative AI and offers a method to improve fairness in image generation.