Investigation of visual bias in generative AI

Abstract

In the field of Artificial Intelligence (AI), the emergence of text‐to‐image generators, such as Stable Diffusion, Dall‐E‐3 and Midjourney has brought about new avenues for creativity. However, as with any innovation, concerns have been raised regarding the presence of bias within AI generated images, particularly those depicting individuals. This thesis explored and analysed the biases within such models by conducting a comparative analysis between the aforementioned models alongside the publicly available LAION‐400M training dataset in relation to real‐world bias. The research approach revolved around the retrieval or generation of images coinciding with the biased terms doctor and nurse. These terms were used to leverage real‐world biases throughout the bias identification process thereby exposing how each generative model deals with this innate bias and by extension discover any bias mitigation techniques along with their effectiveness in comparison to the other models. This was achieved by annotating the images using the DeepFace and FairFace feature extraction models, whose accuracy was evaluated on a human annotated subset of LAION‐400M images. Furthermore, the bias present within the images was deduced via a series of metrics these being; label count, correlation, person prominence, Shannon entropy, Simpson index and the latter twos evenness counterparts. This research highlighted the bias present within the LAION‐400M dataset along with the Stable Diffusion and Midjourney models whilst outlining the inverse bias within the Dall‐E model and the effectiveness of its bias mitigation process. The findings of this research shed light on the pervasiveness of bias in generative AI, highlighting the urgent need for proactive mitigation strategies whilst contributing to the understanding of bias and the development of fairer models and datasets.