Analysis of Types of Inquiries in Student-AI Interaction: A case study of two CS2 tasks

Synthesis: This study analyzes the types of questions students ask generative AI systems during programming problem solving, classifying 830 interactions of CS2 students across two tasks using the Graesser et al. 18-type inquiry taxonomy via a few-shot classifier. It finds that a small subset of question types — chiefly assertion, verification, and instrumental/procedural prompts — accounts for the majority of student inquiries, and that the types of questions students ask shift substantially as a task progresses, from confusion-reporting toward more directive, procedural use. A second, descriptive finding distinguishes first-generation from continuing-generation students: continuing-generation students treated the AI as an active Problem Solving partner, while first-generation students adopted a more confirmatory, validation-oriented role and asked fewer questions overall. The work connects CS Education, Prompt Engineering, and Learning Analytics to the under-explored question of what students actually ask AI, with implications for Intelligent Tutoring design, Help-Seeking scaffolds, and the cognitive character of over-reliance on AI.

Key Findings

1. Skewed inquiry distribution: a small subset of the 18 Graesser question types accounts for most student inquiries — student-AI conversations are dominated by repeated help-seeking patterns rather than diverse questioning.
2. Task-dependent evolution: question types change substantially as a task progresses and differ across the two programming tasks; students shift from assertion/verification (confusion-reporting) toward instrumental/procedural and directive use as tasks grow more conceptually demanding.
3. First-generation difference: first-generation students asked fewer questions than continuing-generation students and leaned on verification to confirm correctness, while continuing-generation students used the AI more as a task-driving assistant — the study's primary qualitative finding.
4. Methodological contribution: a few-shot classification approach, grounded in the Graesser et al. taxonomy and cross-validated with multiple models and manual review, that makes large-scale analysis of Student-AI Interaction feasible.

Why Inquiry Type Matters

Despite the importance of questioning in learning, students rarely ask enough questions in the classroom — yet they interact extensively with generative AI for learning and problem solving. Understanding what kinds of questions students pose to AI systems is therefore central to characterizing the actual learning work happening in Student-AI Interaction. The authors leverage the established Graesser et al. taxonomy (18 question types) to move beyond coarse "how much AI do students use" measures toward a fine-grained account of the cognitive and epistemic character of those interactions. Question asking sits at the heart of Self-Regulated Learning and productive Help-Seeking: the quality and type of a learner's questions reflect underlying cognitive processes, conceptual understanding, and metacognitive regulation. In computing education specifically, help-seeking is strongly tied to students' ability to overcome programming difficulties, so the structure of their AI-directed questions is a window into whether they are engaging the Computational Thinking work or merely delegating it.

The finding that students lean on a small subset of question types, and that this shifts with task progress, has direct design implications: Intelligent Tutoring and AI-learning systems can be built to invite a wider, more productive range of inquiry, and instructors can scaffold question-asking to counter a narrow, habitual interaction style. It also informs debates about Cognitive Offloading — whether students are posing deep, well-formed questions or shallow requests that delegate the cognitive work to the model.

Methodology: Classifying Inquiries

The study was conducted in Fall 2025 in a second-year CS2 course on object-oriented programming in C++ at a large U.S. public university. Students worked in a purpose-built web-based programming environment with an integrated AI chat assistant (a large language model accessed via the OpenAI API), which logged every prompt, response, timestamp, and user identifier. Across two ~60-minute laboratory sessions — the first on foundational concepts (constructors, encapsulation, method implementation) and the second, two weeks later, on the more conceptually demanding inheritance and polymorphism — 72 unique students generated 830 prompts (432 in Session 1 across 60 students; 398 in Session 2 across 37 students; 25 participated in both).

To categorize these prompts, the authors adapted the Graesser et al. question taxonomy, which classifies questions into 18 categories by cognitive and functional role — verification, disjunctive, concept completion, feature specification, quantification, definition, example, comparison, interpretation, causal antecedent, causal consequence, goal orientation, enablement, instrumental/procedural, expectation, judgment/evaluation, assertion, and request/directive. Because students rarely asked explicit questions — many prompts were assertions of confusion or direct commands — the taxonomy proved well-suited to capturing the real shape of student-AI interaction. Classification used a few-shot learning approach: large language models (GPT-5.2 and Claude) were given the 18 category definitions plus programming-context examples and asked to assign each prompt to the single most appropriate category. Empty or non-informative prompts were removed, disagreements between models and two independent reviewers were resolved by discussion against the taxonomy, and resulting classifications fed both aggregate distribution analysis and a state-machine representation of within-session transitions.

Findings: How Inquiry Evolves

Assertion dominated both sessions. Assertions — statements of confusion, lack of understanding, or problem reports without a specific question — were the most frequent category in both sessions (S1 = 86, S2 = 82), reflecting students' tendency to report issues or express uncertainty rather than formulate explicit questions. Instrumental/procedural prompts became the dominant category in Session 2 (S2 = 104; S1 = 79) as tasks grew more complex, while verification (yes/no correctness confirmation) decreased from S1 = 88 to S2 = 66 and request/directive prompts increased (S1 = 41, S2 = 52). Higher-order categories such as comparison, expectational, feature specification, and enablement stayed minimal in both sessions, indicating that students rarely engaged in comparative, predictive, or deeply inferential questioning. A chi-square test showed a near-significant shift toward procedural, task-oriented inquiry in Session 2 (𝜒²(17) = 27.03, 𝑝 = .0576), though it did not reach conventional significance.

State-machine dynamics. In Session 1, students most often began with an assertion (prob ≈ 0.48), whereas in Session 2 they most often started with an instrumental/procedural question (prob ≈ 0.43). In both sessions students lingered within assertion, request/directive, and instrumental/procedural states — evidence that a narrow repertoire of recurring help-seeking patterns dominates the conversational flow.

Generation differences. First-generation students asked fewer questions overall than continuing-generation students (e.g., in Session 1, 26 first-generation students asked 161 questions vs. 34 continuing-generation students asking 271). Across both sessions continuing-generation students showed higher mean usage in most question types, with the largest advantages in procedural and directive categories (instrumental/procedural Δ = +1.03, 𝑝 = .013 in Session 1, remaining second-largest in Session 2; request/directive Δ = +1.97 in Session 2). By contrast, first-generation students relied more on verification and expectational questions. The authors interpret this as continuing-generation students approaching the AI as an active problem-solving partner while first-generation students used it to check and validate rather than to generate or explore — a distinction replicated across both independent sessions. Although conversational AI may reduce some interpersonal barriers to classroom help-seeking, these inquiry differences persisted, suggesting AI-supported systems may need scaffolds that encourage first-generation students to formulate questions and engage in more active help-seeking.

Connections to AI in Education

This is a measurement and characterization study in the CS Education domain. It complements research on Prompt Engineering by studying real student prompts as they naturally occur in CS2 problem solving, and it feeds Learning Analytics pipelines that aim to infer learning behavior from AI interaction logs. It extends prior Help-Seeking work in computing education by moving beyond surveys, perceptions, and usage frequency to the sequential structure of authentic conversational interactions, and it responds to calls for conversational tutoring systems that support deeper reasoning rather than simple answer generation. For instructors, the taxonomy offers a lens for noticing when students' questioning patterns are narrow or shallow and for designing interventions that broaden inquiry and address equity in AI-supported higher education.

What this means for practice

• Instructors. Scaffold question-asking rather than assuming it: model and prompt a wider range of inquiry, since assertion, verification, and instrumental prompts accounted for most of the 830 prompts students sent.
• Instructors. Proactively prompt first-generation students to formulate questions and engage as active problem-solvers instead of answer-checkers, since they asked fewer questions and leaned on verification.
• Instructors. Time scaffolds to the conceptual difficulty of the material — for example inheritance and polymorphism — because inquiry shifted from assertion and verification toward procedural and directive use as tasks grew harder.
• Designers. Build AI-supported programming environments whose interfaces and prompt scaffolds invite reflective inquiry and discourage repeated confirmation-seeking or excessive procedural delegation.
• Researchers. Classify prompts against a cognitive taxonomy such as Graesser et al.'s instead of counting usage, giving Learning Analytics a richer signal for detecting shallow over-reliance from interaction logs.

Limitations

• Single CS2 object-oriented C++ course at one large U.S. public university in Fall 2025: 72 unique students produced 830 prompts across two ~60-minute lab sessions, with only 25 students participating in both.
• The two sessions differed in both programming topic and conceptual complexity, so the shift in inquiry types conflates topic, difficulty, and growing familiarity with the tool; the authors state these findings "should be interpreted cautiously."
• The Session 1 to Session 2 shift was only near-significant (χ²(17) = 27.03, p = .0576), not conventionally significant.
• The generation difference is descriptive and rests on uneven group sizes (Session 1: 26 first-generation students asking 161 questions vs. 34 continuing-generation students asking 271), and classification was model-assigned with disagreements resolved by two reviewers.

Citation

Amoozadeh, M., & Alipour, A. (2026). Analysis of Types of Inquiries in Student-AI Interaction: A case study of two CS2 tasks. [cs.HC, cs.AI].