Dissertation Title: Re-Imagining AI for the Social Ecology of Learning

Abstract:

Artificial intelligence, via large language models, has rapidly entered formal and informal education at scale. The prevailing design paradigm for learning with these tools optimizes for personalization, efficiency, and measurable outcomes, effectively treating learning as a transaction between a student and a system. This dissertation argues that such a framing misses something fundamental: learning is social, and social presence shapes not just what information reaches a learner but how they learn -- their curiosity, effort, and interest. As learners increasingly turn to AI for support, we risk losing the benefits of peer interaction and, more deeply, the conditions that make participation in a learning community possible in the first place.

This research addresses three questions. First: what does human social presence contribute to learning, beyond the technical capabilities of AI? A comparative study of peer and AI-assisted learning reveals a collaboration-quality divide: high-quality peer interactions generate curiosity and engagement that AI cannot replicate, while low-quality peer interactions underperform it. This points to interaction quality rather than modality as a key variable. A randomized controlled trial isolating feedback source attribution demonstrates that perceived human presence increases behavioral engagement even when feedback content is identical, implicating social presence as a motivational mechanism separate from information quality.

Second: why do learners move toward or away from human support? A longitudinal co-design study with high school students, positioning them as ethnographers and designers of their own learning ecology, shows social risk aversion driving many support-seeking decisions: learners often avoid peer help not because AI is better, but because human interaction requires vulnerability that AI does not. In the study's design phase, students who identified their learning challenges as social-emotional primarily designed scaffolding technologies, while those who framed their challenges as cognitive designed replacement technologies.This suggests that how learners understand their own needs shapes the role they want technology to play.

Third: how can AI support participation in the social ecology of learning? Drawing on the empirical and co-design findings, this dissertation develops a set of theoretically grounded design principles for AI systems that make peer presence and process visible, opening pathways from individual AI use toward community. Two design probes, InquiryBits and PeerScope, demonstrate these principles and offer a concrete picture of what it means to design AI as a bridge to participation.

This work reframes the central question of educational AI design: not how AI can optimize individual learning, but how AI can be designed to support participation in the social ecology that makes learning transformative. Contributions include empirical evidence on the mechanisms of social presence in learning, ecological insight into support-seeking behavior in authentic classroom settings, a framework of design principles grounded in the learning sciences, and prototype systems demonstrating those principles in practice.

Patricia Maes
Professor of Media Technology
MIT Media Lab

Mitch Resnick
LEGO Papert Professor of Learning Research
MIT Media Lab

Sidney d'Mello
Director: NSF iSAT (National AI Institute for Student-AI Teaming), Professor in ICS & CS @ U Colorado Boulder
NSF iSAT, University of Colorado Boulder