Paper short abstract:

This paper proposes a new form of digital anthropological research called "latent space ethnography", which explores the latents spaces of AI systems to better understand the emerging entanglements between datasets, AI algorithms and the social, political, and cultural dynamics embedded in them.

Paper long abstract:

One area of artificial intelligence that has gathered interest in recent years are generative models, which have been widely used to produce new forms of representations including deepfakes, AI art, and other kinds of synthetic media. Models in this broad family include generative adversarial networks (GANs) and diffusion models that generate images based on text prompts. Generative models work by identifying “latent spaces”, representing the implicit patterns in the large datasets these models are trained on, and based on which new representations can be generated.

This paper proposes latent space ethnography as a way to reconfigure anthropological approaches to AI. The anthropological study of AI has produced significant new insights into the workings of ‘black box’ AI systems, including critical analyses of predictive policing, facial recognition and algorithmic bias. These critical approaches, however, have generally approached these new powerful systems from the outside - that is, as objects of anthropological inquiry. This paper describes the opportunities that such new generative models offer when considered not only as objects of study but also as tools affording new research methods. Through latent space ethnography, we produce and engage with generative models, probing their latent space in various ways to understand the entanglements of data, AI and social/political dynamics.

In our case studies, we ask these models to envision key historical events to explore the way in which these events have been visually reconstructed in the “distributed representations” of new generative AI models.