Documenting Patterns of Exoticism of Marginalized Populations within Text-to-Image Generators

Text-to-image AI tools depict people from Global South countries in stereotypical, culturally exoticized ways: When asked to generate images of people from India, Bangladesh, Egypt, and other Global South nations performing everyday activities like going to work or grocery shopping, the AI systems produced similar cultural attire and features regardless of the activity, whereas people from the USA, UK, and Australia were shown in activity-appropriate clothing that varied across images.

The problem extends beyond geography to marginalized populations everywhere: The research found that Indigenous populations in wealthy Western countries and marginalized groups within Global South countries experience similar exoticization patterns, suggesting this is a systemic issue affecting non-dominant populations rather than just a geographic bias.

AI fairness research has largely ignored non-Western communities: The authors note that most existing AI fairness studies have overlooked how these tools affect non-Western communities and contexts, meaning harmful biases in these systems have gone undocumented and unaddressed for marginalized populations globally.

Exoticism prioritizes cultural stereotypes over accuracy: By overamplifying cultural features and attire, these AI systems sacrifice realistic, contextual accuracy—a person from India going to an office should be depicted in office-appropriate clothing, not in stereotypical cultural dress regardless of context.

These biases can cause real-world harms: The paper documents concrete implications of how these exoticized depictions can reinforce harmful stereotypes and affect how marginalized populations are perceived and treated.