Scholarlink

<p>Advances in generative AI have enabled visual content creation through text-to-image (T2I) generation. Despite their creative potential, T2I models often replicate and amplify societal stereotypes related to gender, race, and culture. This paper introduces a theory-driven bias detection rubric and a Social Stereotype Index (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\texttt {SSI}\)</EquationSource> </InlineEquation>) to systematically evaluate bias in T2I outputs. We audited three major T2I model outputs–DALL-E-3, Midjourney−6.1, and Stability AI Core with 100 queries across <i>geocultural</i>, <i>occupational</i>, and <i>adjectival</i> categories. Results show recurring stereotypes, including gendered professions, cultural markers, and Western beauty norms. Using our rubric, we applied prompt refinement, which reduced <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\texttt {SSI}\)</EquationSource> </InlineEquation> scores by 58% (<i>geocultural</i>), 66% (<i>occupational</i>), and 53% (<i>adjectival</i>). We conducted a complementary user study, which revealed tensions—while refinement mitigates bias, it may weaken contextual alignment, and participants often viewed stereotypical imagery as more “expected.” We call for T2I systems to balance ethical debiasing with contextual relevance, supporting inclusivity without oversimplifying social realities.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Social stereotypes in AI text-to-image generation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Saharsh Barve,
  • Andy Mao,
  • Jiayue Melissa Shi,
  • Prerna Juneja,
  • Koustuv Saha

摘要

Advances in generative AI have enabled visual content creation through text-to-image (T2I) generation. Despite their creative potential, T2I models often replicate and amplify societal stereotypes related to gender, race, and culture. This paper introduces a theory-driven bias detection rubric and a Social Stereotype Index ( \(\texttt {SSI}\) ) to systematically evaluate bias in T2I outputs. We audited three major T2I model outputs–DALL-E-3, Midjourney−6.1, and Stability AI Core with 100 queries across geocultural, occupational, and adjectival categories. Results show recurring stereotypes, including gendered professions, cultural markers, and Western beauty norms. Using our rubric, we applied prompt refinement, which reduced \(\texttt {SSI}\) scores by 58% (geocultural), 66% (occupational), and 53% (adjectival). We conducted a complementary user study, which revealed tensions—while refinement mitigates bias, it may weaken contextual alignment, and participants often viewed stereotypical imagery as more “expected.” We call for T2I systems to balance ethical debiasing with contextual relevance, supporting inclusivity without oversimplifying social realities.