Algorithmic Bias in Adjective Representation

==============================================

Course Project

Title: Unveiling Algorithmic Bias: A Comparative Study of Age, Gender, and Ethnic Bias in Adjectives using DALL-E 3 and Flux
Team Members: Matan Mayerowicz & Mohamed Sayed Ahmd Course Name: Algorithmic Transparency Seminar Academic Term: Winter 2025 Institution: University of Haifa

Table of Contents

---

1. Project Overview
2. Methodology
3. Tools and Technologies
4. Dataset
5. Results and Findings
6. Conclusion
7. Future Work
8. How to Replicate
9. References

Project Overview

---

This project investigates algorithmic bias in the generation of adjectives across age, gender, and ethnicities, leveraging two text-to-image synthesis tools: DALL-E 3 (a state-of-the-art commercial model) and Flux (an open-source image generation from text tool). Our analysis is driven by a curated set of promotional prompts found in myPromo.csv, aiming to highlight potential biases in AI-generated content.

Methodology

---

1. Prompt Engineering: Systematically modified myPromo.csv prompts to represent diverse age, gender, and ethnic groups.
2. Text-to-Image Synthesis:
   • DALL-E 3: Utilized for generating images and analyzing adjectives in text descriptions.
   • Flux (Open-Source): Employed in parallel to DALL-E 3, allowing for a comparative analysis of bias in open-source vs. commercial models.
3. Bias Analysis: Quantitative and qualitative assessment of generated adjectives for biases, comparing distributions across demographic categories and models.

Tools and Technologies

---

• DALL-E 3: Commercial text-to-image synthesis model.
• Flux: Open-Source image generation from text tool, facilitating comparative bias analysis.
• myPromo.csv: Custom dataset of promotional prompts, varied by demographic attributes.
• Analysis Framework: [Specify any additional frameworks or libraries used for bias analysis, e.g., Python, Pandas, Matplotlib]

Dataset

---

• Source: myPromo.csv, containing [Insert number] base prompts, each modified to represent [Insert number] demographic variations.
• Scope: Comprehensive coverage of age, gender, and ethnic diversity in promotional contexts.

Results and Findings

---

The findings revealed that both models displayed significant biases, with DALL-E 3 showing a more balanced gender representation, but both models predominantly generated images of young Caucasian individuals. This study highlights the need for more diverse training datasets and transparent bias auditing to mitigate these biases in future model development.

Conclusion

---

This comparative study sheds light on the presence and nuances of algorithmic bias in adjective generation across commercial (DALL-E 3) and open-source (Flux) text-to-image models. Our findings emphasize the importance of model selection, dataset diversity, and ongoing bias mitigation strategies in AI development.

Future Work

---

• Model Customization: Explore fine-tuning Flux for reduced bias, leveraging its open-source nature.
• Expanded Demographic Analysis: Incorporate additional demographic attributes into the study.

How to Replicate

---

1. Clone Repository: git clone https://github.com/matanmay/Algoritmic-Bias-Course-Project.git
2. Install Dependencies: pip install -r requirements.txt
3. Access DALL-E 3: Follow official guidelines for API/key setup.
4. Utilize Flux: Clone from [Flux's open-source repository URL], integrate as per our documented methodology.
5. Run Analysis Pipeline: python main.py --prompts myPromo.csv

References

---

References [1] Rohan Naik and Besmira Nushi. “Social biases through the text-to-image generation lens”. In: arXiv (2023). url: https://arxiv.org/abs/2304.06034.

[2] Emily M. Bender et al. “On the dangers of stochastic parrots: Can language models be too big? ” In: Proceedings of the 2021 ACM Conference on Fairness, Accountabil- ity, and Transparency (FAccT). 2021, pp. 610–623. url: https://s10251.pcdn. co/pdf/2021-bender-parrots.pdf.

[3] Timnit Gebru et al. “Datasheets for datasets”. In: Communications of the ACM 64.12 (2020), pp. 86–92. doi: 10.1145/3458723.

[4] Inioluwa Deborah Raji et al. “Closing the AI accountability gap: Defining an end- to-end framework for internal algorithmic auditing”. In: Conference on Fairness, Accountability, and Transparency (FAT). 2020. url: https : / / arxiv . org / pdf / 2001.00973.

[5] Abeba Birhane and Vinay Uday Prabhu. “Large image datasets: A Pyrrhic win for computer vision?” In: Proceedings of the 2021 Conference on Fairness, Accountabil- ity, and Transparency. 2021, pp. 272–286. url: https://ieeexplore.ieee.org/ document/9423393.

[6] Wei Wang et al. “New job, new gender? Measuring the social bias in image generation models”. In: Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT). 2024. url: https://arxiv.org/abs/2401.00763.