T2IBias: Uncovering Societal Bias Encoded in the Latent Space of Text-to-Image Generative Models

All major text-to-image models systematically encode gender and racial biases: The study tested five popular open-source models and found that all of them amplified societal stereotypes, with caregiving roles consistently feminized and high-status professions (CEO, politician, doctor, lawyer) overwhelmingly represented by males and White individuals.

Bias is built into the models' core architecture, not just user input: By using neutral, profession-only prompts, researchers demonstrated that biases are "systematically encoded within the pretrained latent spaces" of these models—meaning the stereotypes are embedded in how the models were trained, not simply reflecting how people write prompts.

Different models have distinct bias patterns that matter for deployment choices: The research identified model-specific biases (e.g., QWEN-Image heavily favoring East Asian outputs, Kandinsky favoring White individuals, SDXL showing broader but still biased distributions), suggesting that organizations cannot treat all T2I models as equivalent when considering responsible AI deployment.

This bias has real organizational and ethical consequences: The paper frames these findings as critical concerns for "responsible AI management" and organizational ethics, indicating that companies using these models in real-world applications risk reproducing and amplifying harmful stereotypes at scale.

Practitioners now have actionable guidance for more equitable AI use: The research provides AI project managers with specific information to "select equitable AI models and customized prompts," offering a practical pathway toward more responsible deployment of text-to-image technology.