[Arbeitstitel] Analyzing Bias in AI-Generated Imagery: A Comparative Study of Adobe Firefly/DALL-E and Stable Diffusion

Within a few years up to 90% of online content might be AI generated [1]. Biases and stereotypes do not only exist in the “real” world, but are even stronger in data generated by AI models [2]. As the models “improve”, they often use the new data on the Internet, which includes more and more images generated by these models. Therefore the biases could grow stronger and stronger. To face this problem, in this MA thesis we look deeper into this matter and compare two of the most used text to image generators: Adobe Firefly/ DALL-E and Stable Diffusion.

Goal of this work is to calculate the biases of the tools with focus on skin color and gender. Therefore several images will be generated and analysed. The dependent variable is “Job”: As biases and disadvantages in jobs and payment are an everlasting topic, images of people with different professions (some high paying and some low paying) should be generated and their skin tone and gender should be analysed. Additionally the aspect “Family” shall be taken in concern: Images of Families/ Couples should be generated and analysed.

1. Research about biases in text to image generators.
2. Testing the APIs and Prompt Engineering for getting the best results.
3. Generate the images to different topics/jobs with the two tools.
4. Annotation of skin color with the Monk Skin Tone Scale [3] and Gender. (This will probably will have only the three types: female/ male/ ambiguous, as everything else is hard to judge from only images of fictional people.)
5. Analysis and Calculation of Statistics

• Python
• Analytical competence

Bianchi, F., Kalluri, P., Durmus, E., Ladhak, F., Cheng, M., Nozza, D., … Caliskan, A. (2023, Juni). Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale. In 2023 ACM Conference on Fairness, Accountability, and Transparency (S. 1493–1504). Zugriff am 2024-06-06 auf http://arxiv.org/ abs/2211.03759 (arXiv:2211.03759 [cs]) doi: 10.1145/3593013.3594095

Cho, J., Zala, A. & Bansal, M. (2023, August). DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generation Models. arXiv. Zugriff am 2024-06-06 auf http://arxiv.org/abs/2202.04053 (arXiv:2202.04053 [cs])

Fraser, K. C., Kiritchenko, S. & Nejadgholi, I. (2023, Februar). A Friendly Face: Do Text-to-Image Systems Rely on Stereotypes when the Input is Under-Specified? ar- Xiv. Zugriff am 2024-04-26 auf http://arxiv.org/abs/2302.07159 (ar- Xiv:2302.07159 [cs])

Jha, A., Prabhakaran, V., Denton, R., Laszlo, S., Dave, S., Qadri, R., … Dev, S. (2024, Februar). ViSAGe: A Global-Scale Analysis of Visual Stereotypes in Text-to-Image Generation. arXiv. Zugriff am 2024-03-25 auf http://arxiv.org/abs/2401 .06310 (arXiv:2401.06310 [cs]) doi: 10.48550/arXiv.2401.06310

Koch, A., Imhoff, R., Dotsch, R., Unkelbach, C. & Alves, H. (2016, Mai). The ABC of Stereotypes About Groups: Agency/Socioeconomic Success, Conservative- Progressive Beliefs, and Communion. Journal of personality and social psychology, 110, 675–709. doi: 10.1037/pspa0000046

Sandoval-Martin, T. & Martínez-Sanzo, E. (2024, Mai). Perpetuation of Gender Bias in Visual Representation of Professions in the Generative AI Tools DALL·E and Bing Image Creator. Social Sciences, 13 (5), 250. Zugriff am 2024-06-12 auf https://www.mdpi.com/2076-0760/13/5/250 (Number: 5 Publisher: Multidisciplinary Digital Publishing Institute) doi: 10.3390/socsci13050250