Scaffolding human and AI instruction: Neural alignment and learning gains in online education

Social interactions are foundational to learning, yet scalable one-on-one instructor-student interaction remains challenging in online video learning. We examine whether a brief, structured pre-lecture instructor-student interaction, led by either a human or large language model (LLM)-powered AI instructor, can enhance student engagement and learning. Integrating behavioral testing with simultaneous eye-tracking and fMRI, we compared three groups in a between-subjects design (no interaction, human interaction, and AI interaction; n = 57). Both human- and AI-led pre-lecture interactions improved learning and enhanced neural alignment in critical brain regions (e.g., the default mode network) among students during learning. Neural and gaze alignments between instructor and students, as well as among students, jointly mediated learning gains, revealing a reciprocal eye-brain-behavior correspondence. While the AI instructor approximated the human instructor in learning gains, students reported lower social closeness and exhibited lower gaze alignment. These findings advance social learning theories and provide neurocognitive evidence that scalable LLM-powered AI can transform online education.