Do Generative AI Models Output Harm while Representing Non-Western Cultures: Evidence from A Community-Centered Approach

AI image generators perpetuate harmful stereotypes of non-Western cultures: The study found that text-to-image models produce representations of Indian culture characterized by exoticism and cultural misappropriation when given English prompts, suggesting these biases are embedded in how these widely-used systems depict non-Western communities.

Community input is essential for identifying AI harms: By conducting focus groups with diverse Indian subcultures rather than relying solely on technical metrics, researchers uncovered "novel representational harms" that might have been missed by standard AI evaluation methods, demonstrating that affected communities must be involved in assessing AI systems.

Current AI systems lack cultural sensitivity at a global scale: The paper emphasizes "the urgent need for inclusive and culturally sensitive T2I systems" and warns of "potential negative repercussions and harms" when these models are deployed globally without adequate safeguards, indicating that responsible AI development requires culturally-informed design from the start.

Existing bias mitigation approaches are incomplete: The research identifies gaps in current work by adopting a "sociotechnical perspective," suggesting that technical fixes alone are insufficient and that developers must consider broader social and cultural contexts when building AI systems used internationally.

Design guidelines must be informed by affected communities: The authors propose that future improvements to generative AI should incorporate design guidelines developed through community-centered approaches, rather than being designed primarily by technologists in Western contexts.