Socratic, Artifact-Gated Generative AI Scaffolding in a 6E Inquiry Unit: A Clustered Quasi-Experiment in Elementary Students’ Computational Thinking

Abstract

Open-ended computational thinking (CT) inquiry challenges one teacher to support learners progressing at different rates. This clustered quasi-experiment evaluated TALPer, a generative-AI tutor providing student-initiated, artifact-contingent Socratic prompts during a six-lesson 6E unit. Ten intact Grade 3 classes in Taiwan (N = 229) completed the same curriculum, tasks, teacher, platform, and instructional time. Conditions differed in a bundled scaffolding configuration involving source, delivery modality, initiation, privacy, and artifact requirements. A two-level model controlling for pretest CT estimated an adjusted posttest difference favoring TALPer ( B = 4.05, SE = 1.58, Wald 95% CI [0.94, 7.15], p = .011; sensitivity p = .034; Hedges’ g = 0.52). Class-clustered analyses found no Holm-adjusted difference on any Intrinsic Motivation Inventory subscale; Perceived Choice had the largest positive estimate ( B = 0.27, cluster-adjusted SE = 0.12, 95% CI [−0.02, 0.55], p = .061, pHolm = .306; Hedges’ g = 0.38). The Condition × Pretest interaction was nonsignificant ( B = −0.109, SE = 0.125, 95% CI [-0.354, 0.137], p = .385). Descriptive coding of 101 matched responses showed different distributions of expected and planned AI-use orientations; no inferential transition test was conducted. Findings provide preliminary evidence concerning the bundled intervention but do not isolate component contributions.