Exposing Hidden Biases in Text-to-Image Models via Automated Prompt Search

Text-to-image (TTI) diffusion models have achieved remarkable visual quality, yet they have been repeatedly shown to exhibit social biases across sensitive attributes such as gender, race and age. To mitigate these biases, existing approaches frequently depend on curated prompt datasets - either manually constructed or generated with large language models (LLMs) - as part of their training and/or evaluation procedures. Beside the curation cost, this also risks overlooking unanticipated, less obvious prompts that trigger biased generation, even in models that have undergone debiasing. In this work, we introduce Bias-Guided Prompt Search (BGPS), a framework that automatically generates prompts that aim to maximize the presence of biases in the resulting images. BGPS comprises two components: (1) an LLM instructed to produce attribute-neutral prompts and (2) attribute classifiers acting on the TTI's internal representations that steer the decoding process of the LLM toward regions of the prompt space that amplify the image attributes of interest. We conduct extensive experiments on Stable Diffusion 1.5 and a state-of-the-art debiased model and discover an array of subtle and previously undocumented biases that severely deteriorate fairness metrics. Crucially, the discovered prompts are interpretable, i.e they may be entered by a typical user, quantitatively improving the perplexity metric compared to a prominent hard prompt optimization counterpart. Our findings uncover TTI vulnerabilities, while BGPS expands the bias search space and can act as a new evaluation tool for bias mitigation. Despite significant advances in text-to-image generation, diffusion models (DMs) (Ho et al., 2020; Rombach et al., 2022) perpetuate and amplify social biases, such as gender, race/ethnicity, culture and age (Seshadri et al., 2024; Bianchi et al., 2023), that prove remarkably persistent across various models like Stable Diffusion (Luccioni et al., 2023), DALL E (Cho et al., 2023) and Midjourney. These patterns reveal how descriptive modifiers and contextual cues encode biases throughout the prompt space - regions that current debiasing techniques, despite reporting success on curated datasets, leave entirely unexplored. Manual or LLM-assisted prompt curation yields realistic test cases but explores only a limited fraction of the prompt space. On the other end, gradient-based prompt optimization discovers high-bias regions but produces unreadable text, e.g. "nurse kerala matplotlib tbody" (see section 4.3), unsuitable for practical auditing or understanding bias mechanisms.