Resisting epistemic loss in AI image generation

Hintze et al.'s recent study highlights the tendency of current-generation vision-language models to converge on overly generic outputs. We argue that considering AI imageries as epistemic artefact and AI-driven artistic practices as socio-cultural processes can provide better understandings around the implications of such conditions beyond merely technical aspects of image generation. Drawing from our ongoing research, we highlight the necessity of bringing the epistemic vulnerability of marginalized artists and users and their hierarchical relations with these tools into sociotechnical design conversations, and by doing so, to explore the possibility for pluriversal and just AI futures, particularly in the Global South.