Tool, Tutor, or Crutch?: A Grounded Theory of Cognitive Scaffolding and Offloading in AI-Assisted Programming Education

Synthesis: This Constructivism grounded-theory study (N = 24 AI-enabled + 17 contrast undergraduates in Java programming) builds a process-level model of how learners navigate the tension between "Domain Mastery" (conceptualization, explanation, evaluation) and "Tool Mastery" (procedural efficiency with AI) through two recurrent loops — Scaffolding and Offloading — interpreted through cognitive load theory and self-determination theory. It explains how performance and affect gains can co-occur with thinner germane processing and attenuated metacognitive calibration under routine offloading.

Key Findings

1. A core tension between Domain Mastery and Tool Mastery. Learners balance conceptual understanding (Domain Mastery) against procedural efficiency with AI (Tool Mastery), switching dynamically via a "Strategic Dance."
2. Two recurrent loops: Scaffolding and Offloading. The tension model centers on a Scaffolding Loop (AI support that preserves learning) and an Offloading Loop (routine delegation that reduces germane processing), with boundary conditions (time pressure, task complexity/familiarity, scaffolding design) shaping movement between them.
3. Trust-but-Can't-Verify and a Boilerplate Blindspot. Novices struggle to verify AI output; more experienced students develop a blindspot for boilerplate/generic code — two distinct evaluation challenges.
4. Attenuated metacognitive calibration. A mismatch between perceived readiness and independent capability co-occurs with sustained offloading, echoing overconfidence research.
5. Concrete instructional strategies. The model proposes dedicated 'critique-the-AI' phases, planned fading of AI assistance through offline tasks, verification journals, and contrastive prompting — shifting the debate from 'use or ban' to how and when the tool aligns with learning goals.

What this means for practice

• Instructors. Run dedicated "critique-the-AI" phases in which students evaluate AI-generated code before accepting it, because the model documents two distinct evaluation failures — novices who struggle to verify output at all, and experienced students with a boilerplate blindspot for generic code.
• Instructors. Plan fading of AI assistance through offline tasks and verification journals, so students move from the Offloading Loop into the Scaffolding Loop rather than sustaining routine delegation that thins germane processing.
• Instructors. Use contrastive prompting so students compare AI responses instead of accepting the first one, shifting the question from "use or ban" to how and when the tool aligns with learning goals.
• Designers. Design around the boundary conditions the model identifies — time pressure, task complexity and familiarity, and scaffolding design — since these shape whether students enter the scaffolding or offloading pathway; the concrete levers align with Reducing AI Misuse and AI Literacy.
• Researchers. Test the proposed levers rather than assuming them: the grounded theory proposes critique phases, planned fading, and verification journals as instructional strategies but generates them from a process model, not an experiment, and Measuring Whether LLM Tutors Teach or Solve: A Diagnostic for Educational Impact shows why the effects need direct measurement.

Limitations

• The 41 participants (24 AI-enabled, 17 contrast) were recruited by voluntary response at a single Chinese public university, and the two sections differed in instructor, time, and peer norms — an ecology-specific design not built for controlled contrasts, so the authors claim transferability through thick description rather than statistical generalization.
• The study spans one 16-week semester, so its categories are theoretically sufficient for this context rather than maximally elaborated across the multi-year span of some grounded-theory traditions.
• Dyadic interviews can invite peer influence and social desirability, and because interviews occurred near high-stakes exams, elevated anxiety is confounded with metacognitive calibration — the individual follow-ups and think-alouds that could separate them were not run.
• Concept maps give a narrow indicator of integrative structure and may not align with all forms of programming expertise.

Citation

Liu, D., Fan, G., & Pan, L. (2026). Tool, tutor, or crutch?: a grounded theory of cognitive scaffolding and offloading in AI-assisted programming education. International Journal of STEM Education, 13, 10.