This study proposes a multi-layered framework to quantify gender and ethnic bias in state-of-the-art text-to-image models using geographically grounded evaluations across ten American countries. We evaluate Stable Diffusion 1.5, SDXL, Gemini, Flux, and DALL.E by creating two datasets: one organized by race–gender categories and another using only nationality prompts. Using CLIP embeddings, we extract demographic representations through three methods: (i) a KNN classifier trained solely on model-generated demographic exemplars; (ii) a distance-based comparison of country-level embedding centroids to race–gender centroids; and (iii) a semantic layer from a multimodal LLM that infers demographic attributes from each portrait. An autoencoder provides latent-space visualizations for interpretable geometric inspection of cross-country patterns. Across models, we find systematic deviations from real census distributions, including consistent male overrepresentation (except Gemini, which skews female) and strong ethnic homogenization. Latin American countries are repeatedly mapped to Indigenous archetypes, while Canada and the United States are largely rendered with White prototypes. These distortions appear in both embedding geometry and LLM-based semantic judgments, indicating that bias affects low-level visual structure and high-level interpretation. By linking synthetic portraits to demographic baselines, this study offers a scalable method to audit representational fairness in AI models and stresses the need for culturally contextualized evaluation in responsible AI development.
Franco et al. (Mon,) studied this question.