The Paradox of AI-Assisted Learning: Analyzing Self-Regulated Learning Readiness Toward the Authenticity of Coding and AI Learning Outcomes

Abstract

This study examines how self-regulated learning (SRL) readiness relates to learning outcomes when students work with AI assistance versus when AI is withheld, in the KKA (Komputer, Koding, dan Kecerdasan Buatan) subject. A second aim is to propose and test the construct of pseudo self-regulated learning (pseudo-SRL) within prompt engineering tasks for AI-based image generation using Google Gemini. This quantitative correlational study involved 33 grade XI students at a senior high school based in a pesantren. Each student completed a 29-item SRL questionnaire, an AI-assisted practical test in which students created and revised prompts to generate and edit images, and a closed-book written test on prompt structure, AI limitations, and ethics, completed without AI support, as in a mid-semester test. Descriptive statistics, Pearson correlation, paired-samples t-tests, and one-way ANOVA were used to analyze the data. Students scored markedly higher on the practical task (M = 78.8) than on the written test (M = 60.8), producing a performance-knowledge gap of t(32) = 7.63, p < .001, 95% CI [13.2, 22.8]. SRL showed only a weak, non-significant relationship with written test performance (r = .298, p = .092) and with practical scores (r = .125, p = .487); practical and written scores were themselves significantly correlated (r = .661, p < .001). The gap was descriptively narrowest among students with high SRL (12.0 points), though a one-way ANOVA showed this difference across SRL tertiles was not statistically significant, F(2,30) = 1.81, p = .181. No participants met the operational criteria for pseudo-SRL, and an exploratory clustering analysis also found no natural pseudo-SRL profile, suggesting that other factors drove the performance-knowledge gap. The paradox of AI-assisted learning also affects prompt-engineering tasks in KKA. We refer to students who fail to translate their metacognitive self-reports into an autonomous understanding of prompting principles as pseudo-SRL; this gap is masked by their proficiency with AI tools. Assessment designs that combine behavior-based self-regulation measures with both AI-assisted and AI-free tasks might be valuable in programming and AI education. The results point to the need for KKA curriculum designers to incorporate dual-assessment models that differentiate between independent mastery of concepts and performance with AI assistance, to allow AI tools to augment rather than substitute for foundational knowledge.Keywords: Generative AI, self-regulated learning, pseudo-SRL, programming education, performance-knowledge gap.