Ph.D. Thesis Proposal - Faria Huq
When
-
Where
NSH 3305 + Remote
Description
Faria Huq - HCII Ph.D. Thesis Proposal
Committee:
Hirokazu Shirado (Chair), Carnegie Mellon University
Aniket Kittur, Carnegie Mellon University
Maarten Sap, Carnegie Mellon University
Willam Bishop, Google DeepMind
"Designing Context-Aware AI Assistance Across Human-AI and Social Interaction"
Abstract:
AI assistants increasingly help people make decisions. Providing useful assistance in these settings requires more than responding to an explicit request: an assistant must determine what information about the user and the surrounding situation is relevant to the task. Yet this information is not always fully specified before assistance begins. Preferences can become observable only after a user reacts to an assistant’s behavior, while socially situated actions can depend on other people and on an interaction that changes over time.
In this dissertation, I study context-aware AI assistance across three complementary aims. First, I investigate how AI assistants can learn individual preferences that emerge through ongoing human-AI interaction. Using human-agent collaborative trajectories, I show that users’ interventions provide systematic signals about how they prefer to collaborate with AI and that these signals can improve adaptive assistance. Second, I examine what changes when AI-assisted actions become part of interactions with other people. Through a preregistered experiment on AI-mediated group communication, I show that incorporating relational context can produce different conversational and group-level outcomes than assistance focused only on the individual user. Third, I propose an interactive AI assistant that learns and updates both individual and relational context as interaction unfolds, using these evolving representations to provide context-sensitive assistance while preserving the user’s intended meaning.
This thesis advances a more specific view of context-aware AI assistance: context for AI assistance is not necessarily given in advance; it can emerge through interaction, and when assisted actions become socially coupled, the relevant context can extend beyond the focal user.