Ethics Diffusion

This repository contains the source code, data, and analysis for the research paper Seeing Bias: Demographic Representation of Professions in Image-Generation Systems which was the final product of the seminar Ethics and Epistemology of AI, which was conducted as a cooperation between the Otto-Friedrich-University of Bamberg, the Technical University Berlin, and the Berlin Ethics Lab.

Abstract

This paper investigates demographic bias in the text-to-image model Stable Diffusion 3.5, focusing on the representation of professions. By systematically varying prompt-specificity with respect to demographic attributes (gender, ethnicity, profession), we analyze the model's outputs and compare them with real-world labor statistics. Our findings show that the model overrepresents white individuals and defaults to gender stereotypes, portraying men in technical roles and women in caregiving positions. Increasing prompt specificity does not mitigate these biases and, in some cases, amplifies or hides them through "stereotype stacking". These results show that the model does not mirror reality but instead reinforces and magnifies existing societal biases, raising critical ethical concerns about representational harm, symbolic erasure, and the perpetuation of inequality through GenAI.

Read the full paper here.

🔑 Key Findings

Our analysis demonstrates that Stable Diffusion 3.5:

• 🧑‍💻 Defaults to Stereotypes: When given neutral prompts (e.g., "a photo of an engineer"), the model defaults to generating images of white men, reinforcing a narrow, stereotypical view of the profession.
• 🔥 Amplifies Existing Biases: The model doesn't just reflect real-world demographic distributions; it often exaggerates them. For example, female-majority professions were depicted with an almost exclusively female representation (95%), far exceeding real-world figures.
• 📉 Fails at Intersectionality: Prompts for intersectional identities (e.g., "Black female engineer") often resulted in lower-quality images, visual distortions, or complete erasure, suggesting the model struggles to represent identities that deviate from its trained defaults.
• 🎭 Performs "Stereotype Stacking": Increasing prompt specificity by adding demographic markers did not consistently reduce bias. Instead, it sometimes layered stereotypes, creating an illusory improvement while leaving systemic issues unaddressed.
• 🌐 Exhibits Strong "White Normativity": Across nearly all professions, the model showed a strong tendency to generate white-presenting individuals unless another ethnicity was explicitly requested.

📊 Data

You can access all results via this Google Drive.

🌳 Repository Structure

.
├── 📄 README.md            
├── 📜 LICENSE              
├── 📋 requirements.txt
├── run.py     
|
├── 📂 assets/
│   ├── prompts.csv         
│   └── imgs/     
|
├── 📂 src/
|   ├── visualize.csv 
│   └── sd_pipe.py  
|
└── 📂 analysis/
    ├── prompt-specificity/ 
    ├── self-analysis/ 
    └── survey/

⚙️ Settings

• You need a HuggingFace access token. See how you can generate it here.
• Create an .env file and save the token there. If you name it other than HF_KEY, you need to modify the code accordingly.
• The configuration for the model is set in SDConfig.
• We use the stabilityai/stable-diffusion-3.5-large-turbo model by default. Be aware that you need to get access to the models first (https://huggingface.co/stabilityai)
• Generated images and prompts are saved in the ./images directory by default.

🏃🏽‍♂️ Quickstart

Clone the repository

git clone https://github.com/VoErik/sd-3-ethics.git
cd sd-3-ethics

Create and activate a virtual environment

python -m venv venv
source venv/bin/activate  # On Windows, use `venv\Scripts\activate`

Install the required dependencies

pip install -r requirements.txt

And run it!

python run.py --task=g \
              --num_imgs=10 \
              --width=512 \
              --height=512 \
              --model="turbo-3.5" \
              --steps=50 \
              --guidance=7.0