EmergencyBias：文本到图像模型在应急场景下的偏差
EmergencyBias: Bias in Text-to-Image Models under Emergency Scenarios
AI总结：
本文定义了T2I模型在应急场景下的EmergencyBias，构建评估框架发现其存在人口统计学与行为偏差，提出ActionAlign方法可减少行为差异并保留图像质量。
AI中文摘要：
文本到图像（T2I）生成中的偏差已成为多媒体内容创作与传播领域的重要问题。然而，现有研究主要关注相对静态且显性的偏差形式，如性别、种族及地域文化属性表征的差异，对不同群体在行为表现、反应方式及社会角色占据上的行为偏差关注较少。应急场景为研究此类偏差提供了极具揭示性的环境，因为这类场景要求模型不仅要描绘在场人员，还要刻画面临风险者、干预者以及责任分配方式。本文定义了EmergencyBias，即T2I模型在应急场景下的一种偏差形式，涵盖人口统计学偏差与行为偏差两类。我们构建了一套评估框架，系统研究7种主流T2I模型、6种典型应急场景及3个人口统计学维度下的EmergencyBias。实验结果揭示了三项核心发现：第一，在未指定人口统计学特征的空白提示下，T2I模型在应急场景中表现出明显的人口统计学偏差，体现为所描绘个体在性别、年龄及肤色上的分布差异；第二，在受控提示下，应急响应中的行为偏差仍与人口统计学差异存在系统性关联，其中性别维度的偏差尤为显著，且不同模型间存在实质性差异；第三，我们提出了ActionAlign这一轻量级提示嵌入校准方法，该方法在减少行为差异方面优于代表性的基于提示的基线，同时在很大程度上保留了图像质量。总体而言，本研究明确了应急场景是T2I模型偏差评估的重要场景，并为社会影响重大的情境下实现更公平的视觉生成提供了可行方向。
英文摘要：
Bias in Text-to-Image (T2I) generation has become an important problem in multimedia content creation and communication. However, existing studies have primarily focused on relatively static and explicit forms of bias, such as disparities in the representation of gender, race, and geo-cultural attributes. Less attention has been paid to behavioral bias in how different groups are portrayed acting, reacting, and occupying social roles. Emergency scenarios provide a revealing setting for studying such bias because they require models to depict not only who is present, but also who is at risk, who intervenes, and how responsibility is allocated. In this paper, we define EmergencyBias, a form of bias in T2I generation under emergency scenarios that includes both demographic bias and behavioral bias. We construct an evaluation framework to systematically study EmergencyBias across seven leading T2I models, six representative emergency scenarios, and three demographic dimensions. Our experimental results reveal three main findings. First, under blank prompts without demographic specification, T2I models exhibit clear demographic bias in emergency scenarios, reflected in the distributions of portrayed individuals across gender, age, and skin tone. Second, under controlled prompts, behavioral bias in emergency responses remains systematically associated with demographic variation, with particularly pronounced disparities along gender and substantial differences across models. Third, we introduce ActionAlign, a lightweight prompt-embedding calibration method that outperforms a representative prompt-based baseline in reducing behavioral disparities while largely preserving image quality. Overall, our work identifies emergency scenarios as an important setting for bias evaluation in T2I models and offers a practical direction toward fairer visual generation in socially consequential contexts.
