FORMULATION OF THE INTENTIONALLY GAPPED PROBLEMS FOR AI-LITERATE LEARNING IN HIGHER EDUCATION

Abstract

This case study proposes and evaluates a pedagogy for integrating generative AI into software engineering students' work by turning known model limitations–hallucination, confident completion, and gap-repair–into structured learning opportunities. Building on evidence that large language models tend to produce fluent but unverifiable answers, the method reframes assessment as a sequence of activities that require higher-order thinking, process transparency, and oral defense. An instructor begins with a «seed» programming task and either diagnoses its ambiguities or deliberately introduces controlled gaps to create a hands-on problem that is underspecified by design. Students then pursue one of two routes: consult an LLM to obtain an initial solution or independently analyze the prompt to surface missing constraints. In both cases, they must request clarifications, revise the specification, and substantiate design choices through tests and justifications. Assessment weights the quality of clarification requests, identification and explanation of ambiguities, correctness and coherence of the final solution, and conceptual mastery demonstrated in a brief viva that probes reasoning rather than code mechanics. An example–the classic factorial table in C#–illustrates how underspecification (range, data type limits, error policy, output format, and purity) predictably elicits plausible yet defective LLM outputs (e.g., sentinel values that conflate error modes, mixed I/O and computation, unreliable overflow checks). Contrasting this seed with a fully specified “corrected” version shows that explicit constraints improve uniform grading but reduce opportunities to critique AI assumptions. By contrast, an intentionally gapped version most effectively compels students to interrogate model completions, formulate testable requirements, and defend choices orally. The approach thereby promotes responsible AI use (verification over deference), strengthens academic integrity (process artifacts and viva-based authorship evidence), and targets upper levels of Bloom’s taxonomy (analysis, evalu- ation, and creation). The paper concludes that gapped, critique-centered tasks, coupled with process-oriented submissions and oral assessment, offer a scalable, human-centered pathway for AI-literate computing education, and motivates empirical studies comparing learning and integrity outcomes against conventional, fully specified assignments.