Can we Debias Social Stereotypes in AI-Generated Images? Examining Text-to-Image Outputs and User Perceptions

TL;DR

The paper tackles the problem of social stereotypes in text-to-image generation by introducing a theory-driven rubric and the Social Stereotype Index ($SSI$) to quantify bias across geocultural, occupational, and adjectival prompts. It combines a computational audit of three state-of-the-art T2I models (DALL-E-3, Midjourney-6.1, Stability AI Core) on 100 prompts (totaling 1,200 images) with an LLM-enabled bias assessment and a prompt-refinement intervention that reduces $SSI$ by substantial margins. A qualitative user study with 17 participants examines perceptions, expectations, and the trade-offs between debiasing and contextual relevance, revealing that users recognize inclusivity but sometimes prefer stereotypical cues for recognizability. The work offers a scalable, model-agnostic debiasing approach and lays out practical implications for designing stereotype-aware T2I systems, along with governance considerations to balance ethical representation with real-world context.

Abstract

Recent advances in generative AI have enabled visual content creation through text-to-image (T2I) generation. However, despite their creative potential, T2I models often replicate and amplify societal stereotypes -- particularly those related to gender, race, and culture -- raising important ethical concerns. This paper proposes a theory-driven bias detection rubric and a Social Stereotype Index (SSI) to systematically evaluate social biases in T2I outputs. We audited three major T2I model outputs -- DALL-E-3, Midjourney-6.1, and Stability AI Core -- using 100 queries across three categories -- geocultural, occupational, and adjectival. Our analysis reveals that initial outputs are prone to include stereotypical visual cues, including gendered professions, cultural markers, and western beauty norms. To address this, we adopted our rubric to conduct targeted prompt refinement using LLMs, which significantly reduced bias -- SSI dropped by 61% for geocultural, 69% for occupational, and 51% for adjectival queries. We complemented our quantitative analysis through a user study examining perceptions, awareness, and preferences around AI-generated biased imagery. Our findings reveal a key tension -- although prompt refinement can mitigate stereotypes, it can limit contextual alignment. Interestingly, users often perceived stereotypical images to be more aligned with their expectations. We discuss the need to balance ethical debiasing with contextual relevance and call for T2I systems that support global diversity and inclusivity while not compromising the reflection of real-world social complexity.

Figures (6)

• Figure 1: Overview of our study design for identifying and mitigating social stereotypes in T2I output.
• Figure 2: Comparing the occurrences of stereotypical biases by rubric items. ( initial and refined T2I outputs).
• Figure 3: Examples of initial and refined image generation across three query types using the three T2I models. the first, second, and third rows are for the queries, an Iraqi person, a manager, and a beautiful person respectively.
• Figure 4: Images generated for A photo of a felon. The initial image set has $\mathtt{SSI}$ of 0.77, whereas the refined image set has $\mathtt{SSI}$ of 0.33. While the refined image set has lower $\mathtt{SSI}$, they may also seemingly deviate from the main context.
• Figure A1: A slide showing an example of the mental model elicitation section. The three columns (A, B, C) of images are generated by the three T2I models (Dall-E, Midjourney, and Stability AI). We used animations to show each of the columns one-by-one, followed by showing all the three columns together.