When Cars Have Stereotypes: Auditing Demographic Bias in Objects from Text-to-Image Models

Objects inherit demographic stereotypes from training data: When text-to-image models generate "neutral" objects without demographic context, they default to versions most similar to middle-aged and White people, revealing that these groups are implicitly over-represented as the model's baseline assumption.

Demographic cues trigger extreme stereotyping: Adding demographic information to prompts produces highly skewed results—in over one-quarter of cases (26.6%), all 20 generated images of an object shared identical attributes (such as rose gold laptops consistently appearing for women), demonstrating how strongly these models encode demographic associations.

Current debiasing approaches create new problems: While prompt-level debiasing techniques reduce differences between demographic groups, they do so by eliminating diversity within groups, essentially replacing one stereotype with another rather than solving the underlying bias problem.

Bias in objects matters beyond human depictions: This research expands the conversation beyond biased human representations to show that even "neutral" objects like cars and laptops carry demographic stereotypes, suggesting bias is more pervasive in AI systems than previously measured.

New measurement tools enable accountability: SODA provides a practical framework for auditing these biases through standardized metrics, making implicit stereotypes measurable and comparable across models—a necessary step for developers to identify and address these issues.