From Cognitive Outsourcing to Reallocation: A 3P Analysis of Student–Generative AI Engagement in Unsupervised Assessments

Synthesis: Yan, Cui, Chiu, Nakajima, and Kozima (2026) apply Biggs' presage–process–product (3P) model to one of the least-observed settings in AI education research: unsupervised, after-class essay assessments. Using in-depth qualitative interviews and 125 submitted dialogue rounds from 38 undergraduates in Japan and China, they show that student engagement with generative AI is not a single behavior but a spectrum bounded by cognitive outsourcing at one end and cognitive reallocation at the other — a GenAI-era reframing of the classic surface-versus-deep distinction. Most students adopted a learning-assistant intention (n = 31) yet enacted it through single-turn, passive question-and-answer dialogue (76.32%) and fragmented integration with independent reading and drafting (78.94%), producing an efficiency paradox: convenience gained at the cost of the intrinsic cognitive work that builds schemas. Only a minority (n = 8) worked as cognitive partners, using sustained dialogue and an integrated workflow to reallocate rather than remove effort. The authors trace the difference to presage conditions — narrow conceptions of GenAI and institutional guidance that prohibits copying but never teaches productive use — and argue that assessment design and academic GenAI literacy must target the learning process, not just policy compliance.

Key Findings

• Engagement forms a spectrum, not a type: Student use of generative AI ranged from cognitive outsourcing (delegating core cognitive work to the tool) to cognitive reallocation (shifting effort from low-level retrieval to critical evaluation and reflection), reframing Biggs' surface/deep approaches for GenAI-supported learning.
• Three usage intentions, frequently overlapping: Analyzing student intentions produced an outsourcing tool group (n = 16, quick completion, full or partial ghostwriting), a learning assistant group (n = 31, the most prevalent, focused on ideas, background knowledge, formatting and revision), and a cognitive partner group (n = 8, the rarest, discussing papers, methods and alternative perspectives).
• Passive single-turn Q&A dominated dialogue: 76.32% of students (n = 29) relied on an ask–get answer–stop pattern, typically pasting the assessment title without specifying their needs and resubmitting identical prompts when dissatisfied; sustained, iterative dialogue appeared in only 23.68% (n = 9), mainly among cognitive partners.
• Integration was mostly fragmented: 78.94% (n = 30) used GenAI either before starting (for ideas or background) or after drafting (for polishing), largely detached from independent reading and writing; only 21.06% (n = 8) alternated between independent work and GenAI throughout the task.
• An efficiency paradox marks the misaligned middle: Learning-assistant students pursued mastery-oriented goals but defaulted to surface processes, reporting overreliance, mental complacency and fast forgetting — detrimental offloading by default rather than by intention (S12: "the speed at which you forget it is also very fast").
• Outsourcing students self-identified their behavior as cheating: Group 1 students described the sequence teacher assigns → student passes the task to GenAI → GenAI generates → student checks formatting → submits; S26 stated the interaction "completely replaced my brain", reflecting abandonment of cognitive agency (Academic Integrity concerns from the student's own perspective).
• Reallocation shifted rather than reduced effort: Cognitive partners reported unchanged total effort with a changed focus — S18 moved resources from searching to quality checking, while S19 wrote short reflection notes after every GenAI session to counter the fading of instantly retrieved information (self-regulation strategy).
• Presage conditions were decisive and uniformly weak: All 38 students reported low confidence in using GenAI effectively, many viewed it as an upgraded search engine or translation tool, and students unanimously reported that instructors prohibited copying but gave almost no concrete guidance on appropriate use — an "institutional vacuum" that pushed even well-intentioned students toward outsourcing.

Study Design & Method

• Design: Qualitative study using Biggs' (1993) presage–process–product model as an overarching analytical framework, operationalized to capture attitudes and perceived guidance (presage), usage intentions, dialogue patterns and integration with other learning activities (process), and self-perceived outcomes (product).
• Participants: 38 undergraduate students (19 from Japan, 19 from China; 24 male, 14 female), first- to fourth-year, drawn from engineering, science, literature, education, law, medicine, foreign languages and economics; recruited voluntarily through personal introductions and student networks, with no supervisory or evaluative relationship to the researchers.
• Context: Unsupervised, argumentative essay-type assessments — authentic tasks intended to cultivate self-directed inquiry, scholarly source use, synthesis and self-regulation.
• Data collection: Semi-structured interviews of 30–60 minutes in which students opened their own GenAI chat histories and explained each step (screen-assisted elicitation), plus submitted dialogue logs totaling 125 interaction rounds; the study was approved by institutional Ethics committees and all participants gave informed consent.
• Analysis: Hybrid thematic analysis following Braun and Clarke, combining the 3P categories with bottom-up open coding; two authors independently coded all transcripts (Cohen's κ = 0.91 for top-level categories), and remaining co-authors audited the codebook, coded samples and thematic map.
• Limitations: A qualitative sample of 38 undergraduates from two East Asian contexts is intended for theory building rather than statistical generalization, and the authors call for large-scale quantitative testing and validation in more diverse cultural and pedagogical settings.

Implications for AI in Education

• Make expert GenAI dialogue visible: Educators should demonstrate what sustained, iterative human–generative AI interaction looks like and why it produces different cognitive outcomes from passive Q&A, building an epistemic relationship with the tool rather than only technical prompting skill.
• Replace prohibitions with task-specific guidance: Instructors should articulate which cognitive tasks students must retain ownership of and which forms of GenAI assistance are appropriate for each Assessment, since blanket plagiarism bans leave students to default to low-effort, fragmented use.
• Design assessments for process, not just product: Requiring intermediate artifacts such as reflections on how GenAI was used alongside other learning activities makes the workflow visible and incentivises deliberate engagement; making reasoning an explicit object of evaluation shifts students away from producing polished text.
• Target the efficiency paradox directly: The largest at-risk group holds mastery goals but lacks the metacognitive strategies to avoid detrimental offloading by default — pedagogical support should focus on monitoring, reflection and dialogue competence for this group.
• Reconsider what counts as learning: When GenAI removes the searching burden, cognitive effort is reallocated rather than reduced, raising questions about whether surface output quality can evidence genuine learning and pushing assessment toward judgment and process evidence.
• Treat GenAI as a potential co-regulator: Under sustained reflective dialogue, GenAI can take on more-knowledgeable-other-like functions — prompting, explaining and offering alternatives — but only when learners have enough GenAI literacy to structure the interaction (Human AI Collaboration).

Connected Concepts

• Generative AI
• Cognitive Offloading
• Metacognition
• Assessment
• Student Engagement
• Human AI Collaboration
• Self-Regulated Learning
• Authentic Assessment

Connected Articles

• Cognitive Offloading in Student–AI Collaboration: A Longitudinal Analysis of Prompting Strategies — Prompting and cognitive offloading in GenAI-supported learning
• Artificial intelligence, cognitive offloading and implications for education — Detrimental versus beneficial cognitive offloading framework
• How university students work on assessment tasks with generative AI: matters of judgement — How students work on assessment tasks with GenAI: matters of judgment
• Chat as Learning: Student-AI Conversations as Discipline-Associated Cognitive Engagement Patterns — Analyzing student–GenAI conversations for cognitive engagement
• Is using artificial intelligence tools for academic work cheating? Student perceptions, ethics, and the impact — Student perceptions of AI-assisted academic work as cheating

Citation

Yan, W., Cui, Y., Chiu, T. K. F., Nakajima, T., & Kozima, H. (2026). From cognitive outsourcing to reallocation: A 3P analysis of student–generative AI engagement in unsupervised assessments. Australasian Journal of Educational Technology, 42(3), 41–61.