Key points
- The study shows text-to-image models amplify demographic and cultural stereotypes at scale, even from neutral prompts.
- The study documents consistent demographic skew in outputs from neutral occupational prompts across all tested models.
- It is a standard citation in the bias-in-generative-media literature.
Summary
Bianchi and colleagues' paper, published at ACM FAccT 2023, provides systematic empirical evidence that widely accessible text-to-image generation systems amplify demographic and cultural stereotypes at scale. Generating images from neutral occupational and social prompts, the study documents consistent over-representation of specific demographic groups and under-representation of others across models. The paper is a standard citation in the bias and representation literature and provides the kind of peer-reviewed empirical grounding that educators need when teaching equity and diversity concerns around generative AI in creative practice.
Related items
Source
Source: ACM FAccT 2023 ↗ (Research)
Cite this item
Bianchi, F., Kalluri, P., Durmus, E., Ladhak, F., Cheng, M., Nozza, D., Hashimoto, T., Jurafsky, D., Zou, J. & Caliskan, A. (2023). ‘Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes’, ACM FAccT 2023. Available at: https://arxiv.org/abs/2211.03759
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.