Aligned but Stereotypical? How System Prompts Shape Demographic Bias in LLM-Based Text-to-Image Models

Text-to-image systems that use large language models to interpret prompts can introduce demographic bias even when no demographic attributes are mentioned, according to a study evaluating eight recent models. The research, posted to the arXiv preprint repository, finds that LLM-based text-to-image systems consistently exhibit stronger demographic skew than systems that do not rely on large language models [1][2]. Large language models are neural networks trained on vast text corpora for tasks such as language generation and are foundational to modern chatbots and generative AI tools [3][5]. The study authors constructed a benchmark covering diverse prompt settings to systematically investigate the behavior across varying levels of prompt ambiguity and complexity [2]. A key mechanism identified in the paper is the role of system prompts — instructions unique to LLM-based text-to-image systems that guide how a user’s prompt is interpreted and expanded. The analysis shows that these system prompts strongly influence text embeddings, which in turn leads to biased image generations [2]. The finding adds to a growing body of work on algorithmic bias, a central concern within the broader field of AI ethics that also encompasses fairness, accountability, and transparency [4]. To address the problem, the researchers propose FairPro, a training-free debiasing framework that adaptively generates fairness-aware instructions while preserving user intent. Experiments demonstrate that FairPro substantially reduces demographic disparities while maintaining prompt fidelity [2]. The paper, titled “Aligned but Stereotypical? How System Prompts Shape Demographic Bias in LLM-Based Text-to-Image Models,” was submitted by NaHyeon Park and revised on June 12, 2026 [1]. arXiv, where the study appears, is an open-access repository of electronic preprints that has hosted scientific papers since 1991 and now receives roughly 24,000 submissions per month [9]. The repository is not peer-reviewed, though it remains a primary distribution channel for research in computer science, physics, and related fields [9]. The study was shared through arXivLabs, a framework that allows community collaborators to develop and share experimental tools and features on the platform [7][8].