AI-generated faces influence gender stereotypes and racial homogenization.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Social Psychology

Background:

• Text-to-image generative AI models like Stable Diffusion are widely used.
• The presence and extent of racial and gender stereotypes in these models are not fully understood.
• AI-generated imagery may perpetuate societal biases.

Purpose of the Study:

• To document racial and gender biases in Stable Diffusion.
• To investigate the degree of racial homogenization in AI image generation.
• To propose and evaluate debiasing solutions and their impact on human biases.

Main Methods:

• Analysis of Stable Diffusion outputs across six races, two genders, 32 professions, and eight attributes.
• Examination of racial similarity in generated images.
• Development and testing of debiasing techniques allowing user-specified demographic distributions.
• A preregistered survey experiment assessing the effect of inclusive vs. non-inclusive AI faces on human biases.

Main Results:

• Significant racial and gender stereotypes were documented in Stable Diffusion.
• Substantial racial homogenization was observed, with specific groups depicted stereotypically (e.g., Middle Eastern men).
• Proposed debiasing solutions demonstrated potential for user control over demographic distributions.
• Exposure to inclusive AI faces reduced human biases, while non-inclusive faces increased them, irrespective of AI labeling.

Conclusions:

• Text-to-image AI models like Stable Diffusion exhibit significant biases and stereotypes.
• Racial homogenization is a key issue in current AI image generation.
• Debiasing strategies and the use of inclusive AI-generated content can mitigate AI-driven and human biases.
• Addressing biases in AI is crucial for responsible technology development and deployment.