Users prefer generative AI to reflect real over idealized gender proportions

Area of Science:

• Artificial Intelligence
• Computer Vision
• Human-Computer Interaction

Background:

• Generative artificial intelligence (GAI) systems face challenges in representing gender in images.
• Existing evaluation standards for gender representation include proportional, uniform, and substantive equality.
• User perspectives on appropriate gender representation in AI-generated images are under-explored.

Purpose of the Study:

• To investigate user evaluations of gender representation in AI-generated images across different social categories.
• To determine which standards of gender representation align with user expectations.

Main Methods:

• An experiment was conducted with 719 US participants.
• Participants evaluated gender representations in various social categories generated by AI systems.
• Statistical analysis was used to assess user preferences.

Main Results:

• User evaluations showed little preference for uniform equality or substantive equality standards.
• Participants favored AI-generated gender representations that mirrored their perception of current statistical reality.
• A strong preference for statistical alignment in gender representation was observed.

Conclusions:

• User expectations for gender representation in GAI are primarily driven by perceived current reality.
• Current AI development may need to prioritize statistical alignment over other equality metrics to meet user expectations.
• Further research should explore the nuances of user perception in AI-driven content generation.