Humor as a window into generative AI bias

Scientific Reports volume  15, Article number: 1326 (2025 ) Cite this article

A preregistered audit of 600 images by generative AI across 150 different prompts explores the link between humor and discrimination in consumer-facing AI solutions. When ChatGPT updates images to make them “funnier”, the prevalence of stereotyped groups changes. While stereotyped groups for politically sensitive traits (i.e., race and gender) are less likely to be represented after making an image funnier, stereotyped groups for less politically sensitive traits (i.e., older, visually impaired, and people with high body weight groups) are more likely to be represented.

Recent years have witnessed two major trends. The first is the increasing use and availability of generative artificial intelligence (AI) commercial systems (e.g., ChatGPT, Gemini). By early 2023, for example, it was estimated that OpenAI’s ChatGPT had accumulated over 100 million monthly users1. The second trend is the enhanced interoperability between generative AI models, whereby large language models (LLMs) and image generators can readily communicate to generate or modify images. For example, a user might type a prompt into ChatGPT like: “Create an image of someone reading a book.”. The LLM GPT4 processes this prompt and expands it into a much more detailed prompt: “An image of a person sitting in a cozy, well-lit room, deeply engrossed in reading a book. They are comfortably seated in an armchair with a warm blanket draped over their legs. The room is filled with soft, ambient lighting, creating a serene atmosphere…” that is used to guide the text-to-image model DALL-E3 to generate an image.

How do AI models communicate and what are the potential consequences of these interactions? In this research, we explore these questions by examining how text-to-image generators and LLMs interact to create images. Specifically, through a preregistered audit of 600 images created by a commercial generative AI, we find that using ChatGPT to update images by making them “funnier” increases the prevalence of some stereotyped groups: people with high body weight2, older, and visually impaired. Our results suggest that these observed biases are primarily attributable to the text-to-image model rather than the language model.

Leave a Comment