T2I Models Show Bias Toward Traditional Attire

The paper by Abdulkareem Alsudais was published on September 30, 2026 and examines how popular text-to-image models DALL‑E 3 and Gemini 3 Pro Preview depict people from different nationalities when prompted to generate everyday activities. ¹

Using prompts that specified nationalities across five activity scenarios—shopping, cooking, commuting, exercising, and socializing—the authors generated 2,060 images of individuals engaging in those activities for 206 distinct nationalities. ¹

Aggregating results across all activities and both models revealed that 28.4% of the images showed subjects wearing traditional attire, including cases where such clothing would be impractical for the depicted activity. This pattern was statistically associated with geographic regions; Middle East & North Africa and Sub‑Saharan Africa were disproportionately affected, as were World Bank income groups. Similar region‑ and income‑linked patterns appeared for images labeled as depicting impractical attire in two athletics-related activities. ¹

To evaluate image–text alignment, the study scored 9,270 image-prompt pairs using CLIP, ALIGN, and GPT‑4.1 mini. Images featuring traditional attire received significantly higher alignment scores when the prompt included a country name compared to generic nationality terms. ¹

Earlier research on algorithmic fairness, including Buolamwini & Gebru’s “Gender Shades” study, highlighted disparities in facial recognition accuracy across demographic groups. Experts caution that while CLIP scores are useful, they can reflect underlying biases—neutral prompts may still elicit stereotypical responses from generative AI models. The work underscores the need for ongoing scrutiny of how T2I systems represent culture.