Interacting with AI by Manipulating Intents

Advanced AI models allow users to perform diverse tasks by simply expressing their high-level intents, without performing low-level operations.However, users can struggle to fully form and effectively express their intents, and inspecting and evaluating model outputs to verify whether their intents have been satisfied incurs significant cognitive load.My PhD research introduces the concept of intent manipulation, where user intents are externalized as interactive objects, allowing for direct exploration and iteration on both intents and model outputs.I explore three forms of intent manipulation: intent curation, disentangle intents into palettes users can curate their intent with; intent assembly, creating intent blocks that users can combine and experiment with; and intent framing, helping users inspect outputs through the lens of their intents.This work contributes to human-AI interaction by suggesting how interfaces can be designed to support iterative exploration and sensemaking of one's own intents and the AI models in parallel.