Cognitive Distribution in Student–GenAI Interaction: A Process Framework from Self-Directed ChatGPT Use in Research Tasks

Abstract

Many studies assess how generative AI (GenAI) affects student learning through benchmarks such as tests, scores, and self-reports. This outcome-focused approach casts GenAI as either a shortcut to replace thinking or a tutor to support learning, overlooking how interaction reshapes students' cognitive activity as they navigate the jagged frontier of AI capability. To examine this process, we conducted a qualitative-dominant mixed-methods case study of self-directed student–ChatGPT interaction in a university transdisciplinary research course. Drawing on human–computer interaction and cognitive philosophy, we developed the Cognitive Distribution Framework, which identifies five interaction types: extension, outsourcing, alignment, transformation, and decoupling; and four roles: cyborg, centaur, ghost, and hermit. Applying it to 53 ChatGPT logs and questionnaires from 112 students, we found diverse distributions of cognitive work. Students most often extended their thinking through active engagement, but also outsourced bounded tasks, aligned the system with their goals, reported shifts in their thinking, and at times limited AI input. They perceived ChatGPT as significantly more cognitively activating than cognitively reducing. Student–GenAI interaction is better understood as a shifting process of cognitive distribution than as fixed dependence or assistance. We call for process-focused research and pedagogies that recognize how students reorganize cognitive work with AI.