Generative AI Shows Bias Toward Traditional Attire Depictions

Abdulkareem Alsudais published a paper on September 30, 2026 describing how two popular text-to-image models—DALL‑E 3 and Gemini 3 Pro Preview—represent individuals from 206 nationalities across five everyday activities ¹.

The experiment involved five activity scenarios and generated 2,060 image–prompt pairs by specifying the nationality and activity for each prompt. The authors found that 28.4% of the resulting images depicted subjects wearing traditional clothing, even when such attire was impractical for the described activity. This pattern correlated significantly with geographic region; Middle East & North Africa and Sub-Saharan Africa were overrepresented compared to other areas. Similar regional and World Bank income-group associations appeared for images labeled as featuring impractical attire in two athletic contexts.

To assess whether the visual content matched the textual description, the team scored 9,270 image–prompt pairs using CLIP, ALIGN, and GPT‑4.1 mini metrics. Images showing traditional attire received higher alignment scores when prompts included country names, suggesting that model outputs are sensitive to explicit cultural cues.

Expert commentary frames these findings within broader concerns about algorithmic fairness in generative AI. Prior work—including Buolamwini & Gebru’s “Gender Shades” study (2018)—has documented disparities in facial recognition accuracy across demographic groups, and similar biases have been observed in text-to-image systems. The expert notes caution that while CLIP-based scores help quantify alignment, they themselves carry inherent biases, so results should be interpreted carefully.