Research Summary: Newer Is Not Fairer: Gender Stereotyping in Text-to-Image AI Across Model Generations

A recent study evaluating gender representation in text-to-image generative AI models found a significant bias towards male subjects across various occupations, including those historically female-coded. This bias persists and, in some cases, worsens across different generations of Stable Diffusion models, indicating that newer models are not inherently fairer in gender representation.

Why it matters

The pervasive gender bias in text-to-image AI models has significant implications for how these technologies shape public perception and professional imagery. Organizations deploying or integrating such AI must address these biases to maintain equity, avoid misrepresentation, and ensure their outputs align with diverse societal values and expectations.

Key insights

• Text-to-image AI models exhibit a substantial gender bias, with 76.4% of generated subjects identified as male across 8,000 images.
• Even for historically female-coded occupations, 57.6% of generated images depicted male subjects, suggesting a systemic misrepresentation.
• The bias is present across four generations of Stable Diffusion models (SD 1.5, SD 2.1, SDXL, SD 3 Medium), demonstrating that model evolution does not automatically lead to improved fairness in gender representation.
• The research used 20 occupations and 5 prompt templates, generating 100 images per occupation-model cell for a robust evaluation.
• The classification of gender in generated images was performed using DeepFace, ensuring a consistent methodology.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.18007