Detecting Gender Bias in LLM Image Generation
An open framework for detecting and documenting gender bias in LLM image generation — automated, continuous, and fully public.
The Challenge
AI image models are now embedded in advertising, editorial, and communication workflows — but their tendency to default to gendered stereotypes is rarely measured or disclosed. We wanted to build the infrastructure to test this systematically: same prompts, multiple models, published results.
The Approach
We designed and built ARIA (Automated Representational and Inequality Auditing for LLMs) — an open framework that sends gender-neutral prompts to multiple leading image generation models, analyses the outputs, and publishes bias scores publicly. The testing pipeline runs continuously, and all findings are available as live data.
The Results
ARIA is now live at bcagency.io/aria with real-time results from 5+ leading image models. The data reveals consistent male-skewed defaults across professional prompts, and the open methodology allows any researcher or journalist to reproduce the findings independently.
