By Gabriele de Seta, Matti Pohjonen, Aleksi Knuutila
de Seta, G., Pohjonen, M., & Knuutila, A. (2024). Synthetic ethnography: Field devices for the qualitative study of generative models. Big Data & Society, 11(4). https://doi.org/10.1177/20539517241303126 (Original work published 2024)
The concept of synthetic ethnography emerged quite organically over time as we discussed our recent research projects and experimental approaches to ethnographic research. We are all social scientists coming from similar disciplinary backgrounds, and in the early 2020s we were all starting to do research about the social and cultural implications of machine learning, generative artificial intelligence, and synthetic media. We each use different combinations of qualitative and quantitative methods with experimental creative strategies, and we wanted to formalize a methodological approach to machine learning models, particularly the ones that power generative AI tools. The main argument of our article is that ethnographic methods can be not only useful to study the development, deployment and use of generative AI models, but that they can also actively experiment with these technologies, turning them into research tools. To support our argument, we combine two methodological traditions: digital methods, and experimental ethnography. From digital methods, we build upon Richard Rogers’ intuition that digital media can be repurposed into research tools to study their own functioning. From experimental ethnography, we draw on Tomás Sánchez Criado and Adolfo Estalella’s conceptualization of “field devices” – inventive social and material techniques used to anchor fieldwork. Synthetic ethnography combines these two approaches, arguing that qualitative analyses of generative AI models can be complemented with the repurposing of these research objects into experimental tools. After making a case for our methodological proposal, our article showcases three practical examples of synthetic ethnography in action, each based on one author’s research project. These include an ethnographic study of deepfakes revolving around first-hand experiences of synthetic media creation; an autoethnographic exploration of ethnic representation in text-to-image model training data; and the development of an interactive tool supporting ethnographic walks into the latent spaces of machine learning models.