Pear Program: Leveraging AI and the Learning Sciences to Create More Fruitful Collaboration

Abstract

This dissertation examines how AI-facilitated pair programming can enhance collaborative problem-solving in online computer science education, addressing persistent barriers to equitable CS access. Through the development and evaluation of PearProgram — an AI-supported collaborative coding platform utilizing an AI facilitator — this mixed-methods study investigates how technology can support both content mastery and relational dynamics in peer learning environments among students from Stanford's Code in Place online course. The research reveals several novel phenomena: a "Gestalt mindset flip" where students with fixed mindsets experience sudden perceptual shifts toward growth mindsets during collaboration; a "shared struggle" mechanism in similarly-skilled pairs that normalizes difficulty and enhances persistence; and evidence that confirmation-response interaction patterns strongly predict collaborative success while "off-topic" conversations strengthen collaboration by building psychological safety. Key contributions include extending Barron's dual-problem space framework to AI-facilitated contexts, documenting psychological transformation mechanisms in collaborative learning, challenging assumptions about optimal skill pairing, reconceptualizing collaborative problem-solving assessment to include relational outcomes, and providing design principles for AI that enhances rather than replaces human collaboration. These findings offer a model for "multiplayer" educational AI that contrasts with dominant single-player approaches, with implications for creating more inclusive and effective online learning environments.