Cognitive offloading and metacognitive calibration in generative AI-mediated doctoral learning: a grounded theory study of accountable reconstruction.

Background: Generative artificial intelligence (GenAI) is increasingly used to explain concepts, summarize literature, suggest methods, generate arguments, and revise academic writing. These functions may reduce entry barriers to complex learning tasks while reshaping the cognitive and metacognitive work through which learners develop understanding. Doctoral research learning provides a high-complexity context for examining this issue because students must move from understanding existing knowledge to producing, justifying, and defending original claims. Objective: This study examines how GenAI-mediated cognitive offloading redistributes cognitive work in doctoral research learning, how it may support or weaken learner-controlled engagement, and how learners recalibrate and reconstruct AI-supported outputs into defensible understanding and research judgment. Methods: A grounded theory design was adopted. After four pilot interviews were used to refine the interview protocol, formal data were collected from 28 semi-structured interviews, 42 voluntarily provided AI-use artifacts, and 12 stimulated-recall discussions. Open coding, axial coding, selective coding, constant comparison, theoretical sampling, and theoretical saturation testing were used to construct a mechanism model. Interview transcripts were coded in Chinese to preserve participants' original meanings, and representative quotations were translated into English after analysis. Results: The analysis generated a five-part mechanism consisting of task-entry scaffolding, modular generation, metacognitive calibration crisis, critical reconstruction, and research accountability. Episodes reflected more learner-controlled engagement when GenAI reduced entry difficulty while preserving students' reading, verification, and reconstruction. Substitution risk appeared when AI-generated frameworks, method suggestions, thesis structures, or texts appeared coherent before learners had developed corresponding understanding. Calibration crisis appeared when recognition was mistaken for understanding, fluency for mastery, and coherence for justification. Critical reconstruction addressed this risk through source verification, self-explanation, contextual adaptation, and counterargument generation. Conclusion: Learner-controlled engagement was most evident in episodes where doctoral students transformed AI-generated outputs into owned understanding, calibrated judgment, and defensible research decisions. These episodes suggest that accountable reconstruction helps preserve learners' responsibility for verification, reasoning, and final judgment in GenAI-supported doctoral learning.