GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension

Researchers introduce GLEaN, a visual explainability method that transforms complex AI bias detection into understandable portrait composites, enabling non-technical audiences to grasp how text-to-image models like Stable Diffusion XL associate occupations and identities with specific demographic characteristics.

GLEaN addresses a critical gap between technical bias research and public understanding of generative AI systems. While extensive academic work documents biases in text-to-image models, most methods remain opaque to non-specialists. This research tackles that transparency problem by converting bias measurement into visually intuitive outputs—composite portraits that show what a model 'imagines' for prompts like 'a doctor' versus 'a felon.' The approach is elegant: generate thousands of images from identity prompts, filter and align faces using facial landmarks, then blend them into median-pixel composites that represent the model's central tendency. User studies with 291 participants confirm these portraits communicate bias insights as effectively as statistical tables while requiring significantly less cognitive load. The method's model-agnostic design—requiring only generated outputs, not internal model access—makes it applicable to proprietary systems like GPT-4V or Claude's multimodal capabilities. This scalability matters considerably. As generative AI becomes embedded in content creation pipelines, design tools, and public-facing applications, the ability to audit and explain biases transparently becomes a regulatory and reputational necessity. GLEaN's public release signals growing momentum in making AI accountability accessible rather than gatekept behind technical expertise. For AI developers and platforms, this represents both opportunity and pressure: opportunity to differentiate through bias transparency, pressure to demonstrate fairness in systems that increasingly influence visual media consumption. The work suggests future auditing standards may favor audience-comprehensible evidence over technical metrics alone.

• →GLEaN converts abstract bias measurements into visual composites that non-technical audiences can intuitively understand in seconds
• →User studies confirm portrait-based bias explanations match statistical tables in effectiveness while reducing cognitive burden
• →The method works on any black-box generative AI system without needing access to model internals, enabling broad auditability
• →Testing on Stable Diffusion XL revealed new associations between predicted emotion and skin tone beyond documented occupational biases
• →Publicly available implementation enables scalable bias auditing across diverse text-to-image models and future generative AI systems