BIBO

Abstract

The BIBO dataset. This dataset provides the companion image corpus for the BIBO study. BIBO study stands for Bias In, Bias Out: A Human-Validated Audit of Default-European Race-to-Role Assignment for High-Status Roles across Four Text-to-Image Models, as detailed on https://hazratali.github.io/bibo/ and https://bibo.stir.ac.uk/. In summary, the main BIBO directory has 1,056 images organized into four sub folders, corresponding to the four audited models, namely, OpenAI GPT Image 1.5, Google Imagen 4, xAI Grok Imagine, and Stability AI SD 3.5 Large. TheBIBO role-swap follow-up folder has the images for the follow-up study. In the main BIBO study every prompt describes the high-status role first. The follow-up describes the low-status role first, with nothing else in the prompt changed, to test whether the European-presenting person is favoured for the high-status role or simply for being the person described first. In the Follow-up directory, there are a total of 314 images for the follow-up study, organized into two sub-folders, corresponding to two audited models, namely, OpenAI GPT Image 1.5 and xAI Grok Imagine. Disclaimer: These are AI-generated images of fictional people, not photographs of real people. The race labels are prompt and labeling categories, not claims about anyone's identity. The README file provides a detailed folder structure of the data directory.