Unmaking AI: A Framework for Critical Investigation

Abstract While generative AI image models are both powerful and problematic, public understanding of them is limited. In this essay, we provide a framework we call Unmaking AI for investigating and evaluating text-to-image models. The framework consists of three lenses: unmaking the ecosystem, which analyzes the values, structures, and incentives surrounding the model's production; unmaking the data, which analyzes the images ad text the model draws on, with their attendant particularities and biases; and unmaking the output, which analyzes the model's generative results, revealing its logics through prompting, reflection, and iteration. We apply this framework to the AI image generator Stable Diffusion, providing a case study of the framework in practice. By supporting the work of critically investigating generative AI image models, “Unmaking AI” paves the way for more socially and politically attuned analyses of their impacts in the world.