Dissertation Title: Seeing Yourself in Latent Space: Human-AI Interaction Beyond Chat
Abstract:
The vast majority of digital content consists of unstructured text, images, and media, and it is growing faster than any person can read. This dissertation designs, builds, and evaluates embedding-based visual interfaces that render an unstructured collection as an interactive semantic map with provenance preserved. It presents Latent Lab, an end-to-end platform that requires no data-science expertise and rethinks what an information environment can be. Two embodiments make the paradigm personal: filescape, for personal file management, and Algorithmic Mirror, which maps a person's cross-platform social media history to make evident how their time is spent and what the platforms have inferred from it. Two findings anchor the evidence. Seeing personal data rendered spatially transformed adolescents' awareness of their algorithmic identities; and against a conversational agent sharing an identical AI backend, the map matched retrieval accuracy at a cost in usability, yet it kept participants' sense of the collection calibrated where chat users believed they had seen more than they demonstrably had, and when asked which tool they would trust with an important decision, participants chose the map, because visible sources let them verify answers rather than accept them. From this evidence the dissertation distills eight design principles for AI-native interfaces that preserve human agency, and argues that embedding-based visualization, long the province of data scientists, is a distinct medium for unstructured information. The title is meant in two senses, and the dissertation finds value in each: seeing for yourself, through visual control of data drawn from the real world and the ability to verify an answer against its sources, and seeing yourself, when the map reveals hidden aspects of one's own data.
Andrew Lippman
Senior Research Scientist
MIT Media Lab
Karrie Karahalios
Professor of Media Arts and Sciences
MIT Media Lab
Fernanda Viégas
Gordon McKay Professor of Computer Science Principal Scientist and Co-Founder of PAIR
Harvard / Google