Publikations-Information

Download

Michelle Dutoit, Baptiste Caramiaux
Exploring Normative Behavior in Generative AI: A Case Study of Text-to-Image Systems Using Stable Diffusion
Proceedings of the 37th Conference on l'Interaction Humain-Machine (IHM '26), October 19-23, 2026, Villeneuve-d'Ascq, France. ACM, New York, NY, USA (bib)

Generative AI systems embed and transmit norms, raising critical questions for creative practice as they increasingly shape cultural production. This paper examines how creative practitioners perceive and experience normativity when using a text-to-image (T2I) system based on Stable Diffusion. We asked artists and designers to create illustrations inspired by two media sources with different levels of specificity, and analyzed their process through a mixed method design combining topic analysis, questionnaires, and semi-structured interviews. Results show that the studied T2I system often simplified the topics involved, particularly in low-abstraction tasks, and that creative practitioners perceived normativity through stereotypical outputs, invariant patterns, and unsolicited omissions or additions of details. These tendencies generated a sense of disempowerment, as creative practitioners had to adapt their creative process and linguistic choices to the systemâs constraints. We conclude by discussing how transparency and stronger user control could mitigate these effects to make this technology more empowering for creatives and artists.