Date of Award

1-2026

Document Type

Dissertation

Publisher

Santa Clara : Santa Clara University, 2026

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science and Engineering

First Advisor

Margareta Ackerman

Abstract

Generative AI models, like Midjourney, Stable Diffusion, and GPT have been shown to reinforce harmful biases, often perpetuating outdated and discriminatory stereotypes. This thesis examines a largely overlooked bias in AI systems: Brilliance Bias. By age six, many children begin to internalize the damaging notion that intellectual brilliance is a male trait — a belief that persists into adulthood.

This thesis consists of three studies that assess brilliance bias in generative AI. The first explores the behavior of popular text-to-image models when given intelligence-based prompts such as “brilliant” and “genius.” In particular, we seek to explore whether the models produce images of people who are male, female or non-binary when asked to depict a highly intelligent individual. Our findings reveal a significant presence of images of men compared to women and non-binary individuals, particularly in Midjourney and Dall-E, demonstrating brilliance bias in these popular models.

The second analysis examines how intellectually gifted women and men are portrayed by pioneering large language models. For this study, we generate stories with identical prompts except for the gender of the main characters. We then run an adjective, verb and lexicon analysis on the stories to determine differences. Our results show that men are written about as having greater achievements and more leadership than women, demonstrating brilliance bias in pioneering GPT models.

The final study explores how a text-to-image model responds to multimodal occupational prompts. For example, when a model such as Midjourney is given an image of a women paired with the word ’Professor’, how is the image modified? We analyze which occupations lead to alterations in a person’s gender. Our findings show that Midjourney exhibits  a consistent pattern of modifying women’s facial features to appear more masculine in roles associated with high intellectual capabilities.

In summary, the findings of this thesis demonstrate that AI models exhibit brilliance bias, reinforcing the misguided notion that exceptional intelligence is inherently male. We conclude by outlining directions for future research to further explore and address this bias in AI.

Download

Share

COinS