Inclusive Prompt Engineering: A Methodology for Hacking Biased AI Image Generation

In this research paper, we offer a methodology, which we call inclusive prompt engineering, to describe the process of examining and modifying the relationship between written prompts and visual outputs of human subjects using AI. Through what we call "toxic positivity," DALL-E attempts to avoid negative stereotypes by defaulting to the "least offensive" images. However, as we will show, this bias towards positivity further reinforces stereotypical images by defaulting to what the LLM sees as "positive": young, beautiful, skinny, and white. Through a series of prompt engineering practices, we show how users can bypass this toxic positivity to create images that better demonstrate the diversity of humanity, including a range of ages, body sizes, races, and genders. We utilize the following process to mitigate bias: Provide a prompt using basic demographic descriptors related to gender, race, and age. Because AI generators function by expanding on the details of a user-provided prompt to generate the image, we examine the prompt AI used to understand how it arrived at the resulting image before adjusting our next prompt. Using the prompt structure laid out by DALL-E, we alter sections that misrepresent the subject. Through continued refining, we arrive at an image more aligned with our intended inclusive result.