Patterns of Student Cognitive Offloading to AI in Higher Education

Generative AI is increasingly woven into students’ everyday academic practices, yet the extent of cognitive offloading, and at what cognitive levels remain insufficiently understood. This study aims to characterize the cognitive nature of naturalistic student - AI interactions and to quantify the extent and types of offloading involved. We analysed 3,047 ChatGPT messages from 46 undergraduates (Study 1) gathered over one semester through unrestricted naturalistic use of GPT for general exam preparation, and 1,140 messages from a 16-student subsample providing learning outcome information (Study 2). In Study 1 an LLM-based coding approach was used to classify messages into question type and Bloom’s taxonomy level. Procedural requests were the most frequent (23 %), whereas higher-order “Analyse” queries dominated the taxonomy distribution (27 %), suggesting that students routinely outsource complex thinking. In Study 2 conversations were manually coded for degree of cognitive offloading (none, light, heavy), task and for Bloom's level. Results showed that “Heavy offloading” clustered in “Create” dialogues (e.g., drafting arguments or code) and appeared more often among students who receive high grades; lower-tier students favored light or no offloading. Because Study 2 is small and single-site, findings are descriptive and not causal; we do not assess learning gains or integrity, but gain valuable insights into the nature of cognitive offloading. Contributions include: (i) an operational, reproducible rubric for offloading that decouples it from Bloom levels, (ii) the joint distribution of Bloom × offloading in authentic student–AI interactions, and (iii) prospective research for confirmatory studies.