2024/10/30 by De Seta, Gabriele
#Generative artificial intelligence #latent space #machine learning models #probes #qualitative methods
paper · doi:10.6092/issn.1971-8853/19512
This essay outlines a methodological approach for the qualitative study of generative artificial intelligence models. After introducing the epistemological challenges faced by users of generative models, I argue that these black-boxed systems can be explored through indirect ways of knowing what happens inside them. Inspired by both ethnographic and digital methods, I propose the use of what I call synthetic probes: qualitative research devices designed to correlate the inputs and outputs of generative models and thus gather insights into their training data, informational representation, and capability for synthesis. I start by describing the sociotechnical context of a specific text-to-video generative model (ModelScopeT2V), and then explain how my encounter with it resulted in an extensive period of experimentation dedicated to the production of Latent China, a documentary entirely composed of synthetic video clips. Reflecting on how this experience bridges qualitative research and creative practice, I extrapolate more general observations about how a long history of research probes across disciplines can inspire the creation of methodological devices designed to allow the indirect exploration of a machine learning model’s latent space.