Beyond Immediate Resolution: Generative AI as an Informal Cognitive Tutor in Novice Programming Learning

Synthesis: Aquino Vara and Encarnación Valentín examine whether generative AI tools serve as an effective informal tutor or encourage uncritical Cognitive Offloading in novice programming education. In a pilot study of 38 early-cycle Information Systems technical students at SENATI (Peru), a 20-item Likert questionnaire measured GenAI usage (functional and critical-reflective) and programming learning (conceptual understanding, Problem Solving, autonomy, confidence, global perception). Spearman rank correlation across 36 complete records found a strong, significant positive association (rs=0.802, p<0.001), with students chiefly using GenAI to untangle abstract concepts and explain compiler error messages — while indicators of independent progress scored lowest. The authors warn that the divergence between task-resolution confidence and autonomous conceptual transfer demands calibrated instructional designs that curb illusions of competence and epistemic debt.

GenAI as Informal Cognitive Tutor

The widespread adoption of generative AI tools has reshaped instructional dynamics in introductory programming education, serving as an on-demand virtual tutor when novices encounter algorithmic roadblocks. The core controversy is whether this assistance fosters durable conceptual understanding or encourages uncritical cognitive offloading. The study positions GenAI as a form of informal cognitive scaffolding that mitigates immediate syntactic frustration, while asking whether it transfers into autonomous conceptual mastery.

Findings

The pilot analyzed data from 38 students (36 complete records) using Spearman's rank correlation because composite distributions violated normality assumptions (Shapiro-Wilk p<0.001):

• A strong, statistically significant positive association between GenAI usage and perceived programming learning (rs=0.802, p<0.001).
• Students primarily used GenAI to clarify abstract programming concepts (M=4.11) and explain compiler error messages (M=4.11).
• Indicators of autonomous progress without ongoing instructor support registered the lowest values (M=3.84).

The authors conclude that while GenAI operates as effective cognitive scaffolding to reduce syntax-related frustration, the observed gap between assisted resolution and conceptual autonomy highlights the need for calibrated instructional designs that curb the illusion of competence and epistemic debt in vocational training.

What this means for practice

• Instructors. Replace answer-giving with graduated hints: configure practice environments to offer heuristics, conceptual analogies, or incomplete pseudocode rather than finished code.
• Add explanation gates, requiring students to write why the error occurred and how they intend to fix it before the tool will accept the technical query.
• Preserve unaided transfer assessments — blind code reading, manual debugging, logic diagramming — because perceived resolution (M = 4.05) outran reported autonomy without ongoing instructor support (M = 3.84).
• Curriculum designers. Sequence the work so the two highest-rated uses, clarifying abstract concepts and explaining compiler errors (both M = 4.11), are followed by independent transfer tasks that expose the gap between fluency and understanding.
• Researchers. Pair self-report scales with objective measures (blind tests, grades, repository activity), since the single-instrument design here produced Cronbach's α > 0.97 on both scales.

Limitations

• The analytic sample was 36 complete cases from a single technical institution (SENATI, Peru), filtered down from 38 voluntary, anonymous responses; the authors state this block prevents population-level generalization.
• Both constructs came from one 20-item self-report Likert instrument; no objective algorithmic performance (grades, blind programming tests, repository metrics) was collected.
• Cronbach's α reached 0.973 for GenAI use and 0.982 for programming learning, which the authors interpret as semantic redundancy and common-method bias rather than psychometric strength, so the rs = 0.802 association is likely inflated.
• The design is cross-sectional with no control group; the authors call for quasi-experimental longitudinal comparisons of unrestricted assistants against Socratic-hint tutors.

Connected Concepts

• CS Education
• Generative AI
• Cognitive Offloading
• Intelligent Tutoring
• Misconceptions about AI
• Scaffolding
• Self-Regulated Learning
• Learner Agency

Connected Articles

• The Chain Behind the Claim: Warrantability in AI-Assisted Qualitative Research — The Chain Behind the Claim
• An Emancipatory Vision for Designing (Generative) AI for Learner Flourishing — An Emancipatory Vision for Generative AI Design
• Advancing diagram-based reasoning in AI tutoring systems: a structural approach for STEM education — StructRAG: Diagram Reasoning in AI Tutoring

Citation

Aquino Vara, M. J., & Encarnación Valentín, N. (2026). Beyond Immediate Resolution: Generative AI as an Informal Cognitive Tutor in Novice Programming Learning. EdArXiv preprint.