Paper deep dive
Analysis of Types of Inquiries in Student-AI Interaction: A case study of two CS2 tasks
Matin Amoozadeh, Amin Alipour
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/19/2026, 5:37:35 AM
Summary
This study analyzes 830 interactions between CS2 students and a generative AI assistant during two programming tasks to understand inquiry types and their evolution. Using the Graesser et al. taxonomy and a few-shot learning classification approach, the researchers found that a small subset of question types dominates student inquiries, with significant shifts in questioning behavior as tasks progress. Early interactions rely heavily on assertion and verification, while later sessions show increased procedural and causal reasoning, though verification loops persist.
Entities (12)
Relation Signals (8)
Matin Amoozadeh ā affiliatedwith ā University of Houston
confidence 99% Ā· Matin Amoozadeh University of Houston Houston, USA
Amin Alipour ā affiliatedwith ā University of Houston
confidence 99% Ā· Amin Alipour University of Houston Houston, USA
Graesser et al. taxonomy ā usedforclassification ā Student Inquiries
confidence 98% Ā· We use the Graesser et al. taxonomy to classify students' inquiries into 18 types.
Few-Shot Learning ā usedforclassification ā Student Inquiries
confidence 97% Ā· We develop a few-shot learning approach to automatically classify students' interactions with AI into these categories.
Verification ā dominantin ā Early Interactions
confidence 92% Ā· students frequently rely on assertion and verification prompts during early interactions
Procedural Reasoning ā increasesin ā Later Sessions
confidence 92% Ā· later sessions exhibit greater use of procedural and causal reasoning questions.
GPT-4 ā powers ā AI Assistant
confidence 90% Ā· The AI assistant was powered by a large language model (GPT-4)
CS2 ā taughtin ā University of Houston
confidence 85% Ā· This study was conducted during the Fall 2025 semester in a second-year computer science course (CS2) at a large public university in the United States.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Background and Context: Question and inquiry are integral parts of knowledge seeking and learning. Despite their importance, students tend not to ask enough questions in the classroom. However, studies have shown that students interact extensively with generative AI systems for learning and problem solving. Objective: In this paper, we seek to better understand the types of questions that students ask AI systems, and how those questions evolve during problem solving and across tasks. Method: We use the Graesser et al. taxonomy to classify students' inquiries into 18 types. We develop a few-shot learning approach to automatically classify students' interactions with AI into these categories. We use this system to analyze 830 interactions of CS2 students across two programming tasks. Findings: Our results suggest that a small subset of question types accounts for the majority of student inquiries, and that the types of questions students ask change substantially as the task progresses.
Tags
Links
- Source: https://arxiv.org/abs/2608.17919v1
- Canonical: https://arxiv.org/abs/2608.17919v1
Trouble viewing inline? Open PDF directly ā
Full Text
35,599 characters extracted from source content.
Expand or collapse full text
Analysis of Types of Inquiries in Student-AI Interaction A case study of two CS2 tasks Matin Amoozadeh University of Houston Houston, USA Amin Alipour University of Houston Houston, USA Abstract Background and Context. Question and inquiry are integral parts of knowledge seeking and learning. Despite their importance, students tend to not ask enough questions in the classroom. How- ever, studies have shown that students interact with generative AI systems extensively in their learning and problem solving. Objec- tive. In this paper, we seek to better understand types of questions that students ask AI systems, how those questions evolve during problem solving and across tasks. Method. We use Graesser et al. taxonomy to classify studentsā inquiries into 18 types. We develop a few-shot learning approach to automatically classify studentsā interactions with AI into those categories. We use this system to analyze 830 interactions of CS2 students on two programming tasks. Findings. Our results suggest that a small subset of question types accounts for the majority of student inquiries, and the types of questions students ask change substantially as the task progresses. Implications. This paper provides insights into studentsā inquiry behaviors when interacting with AI, which can inform the design of AI-based instructional support and pedagogical interventions. CCS Concepts ā¢ Social and professional topicsāComputing education;ā¢ Computing methodologiesā Natural language processing. Keywords generative AI, programming education, student-AI interaction, help- seeking, question taxonomy, CS education ACM Reference Format: Matin Amoozadeh and Amin Alipour. 2026. Analysis of Types of Inquiries in Student-AI Interaction: A case study of two CS2 tasks. In . ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/n.n 1 Introduction Generative AI systems such as ChatGPT have reshaped program- ming education by providing students with always-available sup- port for debugging, code generation, conceptual explanation, and problem solving. Recent research shows that computing students Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. Conferenceā17, Washington, DC, USA Ā© 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-x-x-x/Y/M https://doi.org/10.1145/n.n increasingly rely on AI systems as learning resources during pro- gramming activities, often using them alongside or instead of tra- ditional help-seeking resources such as peers, teaching assistants, documentation, and online forums [7,10]. As AI tools become in- creasingly integrated into programming courses, understanding how students interact with these systems has become an important challenge for computing education research [1]. Question asking plays a central role in learning, help-seeking, and self-regulated problem solving. Prior educational research has shown that the quality and type of studentsā questions reflect under- lying cognitive processes, conceptual understanding, and metacog- nitive regulation [4]. In computing education, help-seeking be- haviors are strongly associated with studentsā ability to overcome programming difficulties and engage productively with learning resources [8]. Generative AI changes this help-seeking landscape by enabling students to engage in continuous conversational in- teractions during programming tasks. Rather than asking isolated questions to instructors or peers, students can iteratively refine prompts, request explanations, validate solutions, and delegate sub- tasks to AI assistants in real time. Although recent studies have examined studentsā perceptions of ChatGPT, AI-assisted debugging, and help-seeking preferences in programming courses, much of the existing literature focuses primarily on performance outcomes, perceptions, or overall usage frequency [2,3]. Comparatively little is known about the structure and evolution of studentsā questions during AI-supported program- ming problem solving. Existing work has not sufficiently examined how students transition between different forms of inquiry over time, how questioning behavior changes across programming tasks, or whether AI interactions reflect deeper conceptual engagement versus surface-level verification behavior. Understanding these in- teraction patterns is important because they may reveal whether students use AI as a collaborative learning partner or primarily as an answer-checking tool. To analyze these interactions, we adapted the question taxonomy proposed by Graesser et al. [4] to categorize the different types of inquiries students make during AI-supported programming tasks. The taxonomy classifies questions according to their cognitive and functional roles, including verification, procedural guidance, causal reasoning, interpretation, and evaluation. Originally developed to study question asking in tutoring and educational dialogs, the framework enables fine-grained analysis of learnersā informational intent. In the context of programming education, adapting this taxonomy allows us to examine how students use AI assistants for debugging, conceptual understanding, implementation guidance, and solution validation. In this paper, we analyze 830 prompts collected from CS2 stu- dents interacting with an AI-supported programming environment arXiv:2608.17919v1 [cs.HC] 18 Aug 2026 Conferenceā17, July 2017, Washington, DC, USAMatin Amoozadeh and Amin Alipour during two laboratory sessions focused on object-oriented pro- gramming in C++. Using a few-shot learning approach grounded in the Graesser taxonomy, we examine the distribution of student question types, transitions between inquiry states, and changes in questioning behavior across sessions. Our findings reveal that students frequently rely on assertion and verification prompts dur- ing early interactions, while later sessions exhibit greater use of procedural and causal reasoning questions. We further identify recurring verification loops in which students repeatedly use AI for confirmation rather than conceptual exploration. Guided by this goal, we address the following research questions: RQ1: What types of questions do students ask when interacting with an AI during programming tasks? RQ2: How do students transition between different types of questions during their interaction with AI? RQ3: How do studentsā questioning behaviors change across programming tasks? Our results suggest that studentsā question types shift substan- tially as tasks progress, differ across conceptual areas, and reflect varying degrees of cognitive engagement. For example, students tend to use assertion-type questions to confirm understanding of introductory concepts such as access modifiers in object-oriented programming, while increasingly relying on AI to make program design decisions for more misconception-prone topics such as poly- morphism in inheritance. These findings provide insights for future AI-supported instructional systems in computing education. This paper: ā¢applies the Graesser et al. taxonomy to student-AI program- ming interactions; ā¢ develops a few-shot approach for inquiry classification; ā¢analyzes how inquiry patterns evolve across two CS2 pro- gramming tasks. 2 Related Work Help-Seeking in Computing Education. Help-seeking is an es- sential component of self-regulated learning and programming problem solving. The emergence of generative AI introduces a fun- damentally different form of help-seeking interaction because stu- dents can engage in continuous conversational exchanges with AI assistants while programming. Recent work by Hou et al. examined how generative AI influences computing studentsā help-seeking preferences and found that AI systems are increasingly becoming part of studentsā help-seeking ecosystems [5]. However, their find- ings also suggest that effective AI-supported help-seeking requires new skills related to prompt construction and interaction strategies. Other studies similarly report that students vary considerably in how they leverage AI tools during programming activities [14]. While prior work has examined studentsā perceptions of AI tools and help-seeking preferences, much of the literature focuses on surveys, interviews, or learning outcomes rather than the structure of studentsā conversational interactions with AI systems during authentic programming tasks. Question Asking and Educational Dialogue. Question asking plays a central role in learning, tutoring, and knowledge construction. Graesser and Person proposed a widely used taxonomy of question categories that characterizes questions according to their cognitive and functional roles, including verification, causal reasoning, pro- cedural inquiry, interpretation, and evaluation [4]. This framework has been extensively applied in educational dialogue research and intelligent tutoring systems to analyze how learners seek informa- tion and construct understanding during problem solving. Educational dialogue research has demonstrated that the types of questions students ask often reflect underlying cognitive pro- cesses, misconceptions, and levels of engagement [13]. Prior studies in intelligent tutoring systems have shown that analyzing inquiry behavior can provide insights into productive and unproductive learning strategies during tutoring interactions. More recent work on conversational AI in education similarly emphasizes the im- portance of pedagogically informed dialogue systems that support deeper reasoning and metacognitive engagement rather than simple answer generation [6]. Despite extensive use of question taxonomies in educational dialogue research, relatively little work has adapted these frame- works to analyze conversational interactions between students and generative AI systems in programming education contexts. Student-AI Interactions and Prompting Behavior. As conversa- tional AI systems become increasingly common in programming education, researchers have begun investigating how students con- struct prompts and interact with AI assistants during programming tasks. Prior studies have examined prompting patterns, debugging conversations, and studentsā use of AI-generated explanations and solutions [3]. Research has also explored how students iteratively refine prompts and use AI systems for implementation support, debugging assistance, and code validation [9]. However, most existing work focuses primarily on performance outcomes, perceptions, or isolated prompt characteristics rather than the evolution of inquiry behavior across interactions. In partic- ular, little research has examined how students transition between different forms of questions during AI-supported programming problem solving. Existing studies rarely combine educational ques- tion taxonomies with sequential interaction analysis to characterize how inquiry strategies evolve over time. To address this gap, our work adapts the Graesser question taxonomy to AI-supported pro- gramming interactions and analyzes how CS2 students transition between different forms of inquiry while interacting with an AI assistant during programming tasks. 3 Methodology 3.1 Study Setting and Participants This study was conducted during the Fall 2025 semester in a second- year computer science course (CS2) at a large public university in the United States. The course focuses on object-oriented program- ming concepts using C++. Data were collected during scheduled laboratory sessions in which students worked on programming assignments under the supervision of teaching assistants. The study included two data collection sessions conducted dur- ing separate laboratory meetings during the Fall 2025 semester, each lasting approximately 60 minutes. Session 1 was conducted during the middle of the semester, and Session 2 was conducted two weeks later. The two programming tasks differed in both topic and com- plexity, allowing us to examine how studentsā interactions with the Analysis of Types of Inquiries in Student-AI InteractionConferenceā17, July 2017, Washington, DC, USA Table 1: Participant Characteristics by Session CharacteristicSession 1Session 2 n (%)n (%) Freshman24 (40.0)17 (45.9) Sophomore21 (35.0)12 (32.4) Junior13 (21.7)6 (16.2) Senior2 (3.3)2 (5.4) Computer Science44 (73.3)32 (86.5) Computer Engineering8 (13.3)1 (2.7) Mathematics3 (5.0)1 (2.7) Other Majors5 (8.3)3 (8.1) Male46 (76.7)26 (70.3) Female13 (21.7)10 (27.0) Prefer not to say1 (1.7)1 (2.7) First-Generation26 (43.3)20 (54.1) Continuing-Generation34 (56.7)17 (45.9) Programming Exp. (yrs) ķ= 2.25, ķķ·= 1.65 ķ= 2.59, ķķ·= 1.71 N =60 for Session 1; N =37 for Session 2. AI assistant varied across different object-oriented programming concepts. Session 1 focused on foundational object-oriented programming concepts including constructors, encapsulation, and method imple- mentation in C++. Session 2 focused on inheritance and polymor- phism using relatedPetandDogclasses. The second task involved more advanced object-oriented programming concepts and greater conceptual complexity. Because the two sessions involved different programming topics and conceptual complexity, differences in stu- dentsā questioning behavior may reflect both evolving interaction patterns with the AI assistant and differences in task demands. A total of 72 unique students participated across the two sessions. Session 1 included 60 students, while Session 2 included 37 students. Twenty-five students participated in both sessions, allowing us to observe how studentsā interactions with the AI assistant evolved across different points in the semester. During the lab sessions, students worked individually on their programming assignments using a web-based programming en- vironment with integrated AI support developed for this study. The system logged studentsā prompts and interactions with the AI assistant while they worked on their tasks. Participation occurred as part of a graded laboratory activity within the course, and all interactions took place in the regular classroom environment with teaching assistants present. The study protocol was approved by the universityās Institutional Review Board (IRB). Table 1 summarizes the demographic characteristics of the participants in each session. Most participants were freshmen or sophomores, reflecting the typical enrollment profile of CS2 courses. For this study, we developed a web-based programming environ- ment that allowed students to interact with an AI assistant through a chat interface. The system integrates the programming assign- ment description, a code editor, program execution output, and a conversational AI assistant within one single interface. The AI assistant was powered by a large language model (GPT-4) accessed through the OpenAI API. Upon logging in, participants first completed a short demo- graphic survey and were then directed to the programming en- vironment where they worked on their assignment. The interface enabled students to read the assignment instructions, write and modify code, run their program, and interact with the AI assistant through the chat interface. Students could freely ask questions and include code snippets in their prompts while receiving AI-generated guidance related to debugging, implementation strategies, and con- ceptual understanding of object-oriented programming in C++. All interactions with the system were automatically logged, in- cluding student prompts, AI responses, timestamps, user identifiers, and additional interaction metadata. These logs enable detailed anal- ysis of studentsā question-asking behavior during programming tasks. The system was used during scheduled laboratory sessions lasting approximately 60 minutes. Students were allowed unlimited interaction with the AI assistant during the session. At the end of the session, students were required to submit their programming assignment through the system; if a submission was not completed manually, the system automatically submitted the current code when the time limit expired. 3.2 Classification using Few-Shot Learning Student prompts were categorized using the question taxonomy proposed by Graesser et al. [4]. This taxonomy classifies student questions into 18 categories that reflect different cognitive functions of inquiry, including verification, explanation, causal reasoning, pro- cedural guidance, and evaluation. Originally developed to study question asking in tutoring and instructional dialog, the framework has been widely used in educational research to analyze how learn- ers seek information and construct understanding during problem solving. In the context of programming tasks, adapting this taxonomy enables fine-grained analysis of how students use AI assistants for conceptual understanding, debugging support, implementation guidance, and solution validation. For example, the taxonomy al- lows us to distinguish between prompts seeking conceptual clarifi- cation (e.g., definition or explanation questions), procedural guid- ance (e.g., how to implement a feature), debugging explanations (e.g., causal antecedent questions), and confirmation of correctness (e.g., verification or judgment questions). Such distinctions provide insight into the different inquiry strategies students employ while interacting with AI during programming activities. Table 2 sum- marizes the question categories used in our analysis along with definitions and examples adapted to programming contexts. To automatically classify student prompts into the Graesser cat- egories, we employed a few-shot learning approach using large language models accessed through the OpenAI and Anthropic APIs. Few-shot learning enables the model to perform classification by conditioning on a small set of labeled examples rather than requir- ing task-specific supervised training. For each prompt, the models were provided with the definitions of the 18 Graesser categories together with several example prompts illustrating each category in a programming context. The models were instructed to assign each Conferenceā17, July 2017, Washington, DC, USAMatin Amoozadeh and Amin Alipour Table 2: Graesser et al [4] Question Categories CategoryDefinitionExample (Programming Context) VerificationAsks whether a statement or condition is true or falseIs this constructor required? DisjunctiveAsks which of multiple alternatives is correctShould this be public or protected? Concept Completion Requests missing information needed to complete a concept What is the return type of this function? Feature SpecificationAsks about properties, attributes, or components of an entity What data members should this class have? QuantificationRequests numerical or quantitative informationHow many parameters does this function take? DefinitionAsks for the meaning of a concept or termWhat does protected mean? ExampleRequests an illustrative instance of a conceptCan you give an example of inheritance? Comparison Asks about similarities or differences between con- cepts What is the difference betweenprivateand protected? Interpretation Asks for explanation or inference about observed behavior or data What is happening in this output? Causal AntecedentAsks why an event or state occurredWhy is this line causing an error? Causal ConsequenceAsks what will happen as a result of an actionWhat happens if I remove this constructor? Goal Orientation Asks about the purpose or intention behind an actionWhat is the goal of this function? EnablementAsks what allows an action or process to occurWhat enables polymorphism here? Instrumental / ProceduralAsks how to perform an action or procedureHow do I initialize the constructor? ExpectationAsks what outcome should normally occurWhat should the output look like? Judgment / EvaluationEvaluates correctness, quality, or appropriatenessIs this implementation correct? AssertionDeclarative statement indicating belief, confusion, or knowledge I donāt understand inheritance. Request / DirectiveCommand requesting the listener to perform an ac- tion Fix this code. student prompt to the single most appropriate category; prompts that did not clearly fit any category were labeled as Other. Before classification, the dataset was preprocessed to remove empty or non-informative prompts (e.g., greetings such as āHiā). The remaining prompts were then submitted individually to both GPT-5.2 and Claude using a predefined classification template. To improve reliability, two authors independently reviewed the cate- gorized prompts and compared the classifications generated by the two language models. Disagreements between models or reviewers were manually examined and resolved through discussion based on the definitions provided in the Graesser taxonomy. This validation process helped ensure that the assigned categories aligned with the intended cognitive functions of each question type. The resulting classifications were subsequently used to analyze patterns in stu- dentsā question-asking behavior during AI-supported programming tasks. 3.3 Data Analysis Following the few-shot classification of prompts into the Graesser question categories, we conducted a quantitative descriptive analy- sis of studentsā inquiry behavior during programming tasks. In total, 830 prompts were analyzed, including 432 prompts from Session 1 and 398 prompts from Session 2. First, we examined the overall distribution of prompts across the Graesser categories to identify the most common types of questions students asked when interacting with the AI assistant. We also re- ported the distribution of question categories for all participants and separately for first-generation and continuing-generation students to provide descriptive insights into studentsā inquiry patterns. Next, we compared the distribution of question categories be- tween the two sessions to explore how questioning behaviors changed over time. Because 25 students participated in both ses- sions, we also examined how inquiry patterns evolved for these common participants across sessions. Finally, we analyzed the sequence of question types within each session using a state-machine representation of the Graesser cat- egories. This analysis allowed us to examine transitions between question types and to explore how studentsā inquiry strategies evolved during the progression of a programming task. 4 Results 4.1 Frequency of Inquiry Types Before analyzing the types of questions students asked, we first sum- marize the dataset used in this study. Table 3 provides an overview of the question dataset collected across the two laboratory sessions. In total, 830 prompts were recorded from 97 session participations. Session 1 included 60 students who generated 432 inquiries (an average of 7.20 prompts per student), while Session 2 involved 37 students who produced 398 inquiries (an average of 10.76 prompts per student). The higher average number of prompts in Session 2 suggests that students interacted more frequently with the AI as- sistant during the later session. Analysis of Types of Inquiries in Student-AI InteractionConferenceā17, July 2017, Washington, DC, USA Figure 1: Frequency of types of questions in Session 1 and Session 2. Figure 1 presents the distribution of question categories across Session 1 and Session 2, based on the Graesser et al. question tax- onomy. Assertionādefined as statements of confusion, lack of un- derstanding, or problem reports without a specific questionāwas the most frequent category in both sessions (S1 = 86, S2 = 82), re- flecting studentsā tendency to report issues or express uncertainty rather than formulate explicit questions. In contrast, Instrumen- tal/procedural prompts became the dominant category in Session 2 (S2 = 104; S1 = 79), indicating that students increasingly asked how to accomplish tasks or follow specific steps as they engaged with more complex programming challenges. Verification prompts, which seek yes/no confirmation of correct- ness, decreased from Session 1 to Session 2 (S1 = 88, S2 = 66), while Request/Directive promptsāwhere students directly asked the assis- tant to perform an actionāincreased (S1 = 41, S2 = 52). Additionally, categories such as Interpretation and Judgment showed moderate presence in both sessions but remained secondary compared to dominant categories. Higher-order categories such as Comparison, Expectational, Feature specification, and Enablement remained min- imal across both sessions, suggesting that students rarely engaged in comparative, predictive, or deeply inferential questioning. A chi-square test of independence revealed a near-significant difference between the two sessions,ķ 2 (17)=27.03,ķ= .0576, indicating a trend toward a shift in questioning behavior from confusion-reporting in Session 1 toward more procedural and task- oriented inquiry in Session 2, though this difference did not reach conventional statistical significance. Although the number of participants in the two groups is rela- tively comparable, first-generation students asked fewer questions overall than continuing-generation students (Table 3). For exam- ple, in Session 1, 26 first-generation students asked 161 questions, whereas 34 continuing-generation students asked 271 questions. A similar pattern is observed in Session 2, where 20 first-generation students asked 159 questions compared to 239 questions asked by 17 continuing-generation students. In Session 1, first-generation students relied most on Verifica- tion (ā26%) and Assertion (ā19%), while continuing-generation stu- dents favored Instrumental/Procedural (ā22%) and Assertion (ā21%). Table 3: Question Dataset Summary by Session and Generation Status Session 1Session 2 Categoryķ Questions ķ Questions All Students6043237398 Continuing-Gen.3427117239 First-Gen.2616120159 Avg. Questions per Student (MeanĀ± SD) All Students7.2010.76 Continuing-Gen.7.97 (7.23)14.06 (17.78) First-Gen.6.19 (5.58)7.95 (7.76) Note. ķ = number of students. Between sessions, first-generation students showed a marked in- crease in Instrumental/Procedural questions (ā12.5% toā27%) and declines in Interpretation and Request/Directive, whereas continuing- generation students shifted toward Verification (ā8.5% toā18%) and Request/Directive questions. These patterns are descriptive; no shifts reached statistical significance after Bonferroni correction. 4.2 Inquiry Patterns by Generation Beyond aggregate diversity scores, a granular examination of indi- vidual Graesser question-type usage reveals meaningful differences in how first-generation and continuing-generation students engage with the AI chatbot. Figure 2 presents the mean difference in usage for each of the 17 question types across both sessions, with positive values indicating higher usage by continuing-generation students. Across both sessions, continuing-generation students demonstrated higher mean usage in the majority of question types (Session 1: 9 of 17; Session 2: 12 of 17), with the most pronounced advantages concentrated in procedural and directive categories. Specifically, Instrumental/procedural questions ā in which students ask the AI how to accomplish a task ā showed the largest statistically signifi- cant difference in Session 1 (Ī=+1.03,ķ= .013), and remained the second-largest advantage in Session 2 (Ī=+1.44). Similarly, Re- quest/Directive questions, in which students issue direct commands or requests to the AI, favoured continuing-generation students in both sessions (Ī=ā0.02 in Session 1;Ī=+1.97 in Session 2). By contrast, first-generation students demonstrated relatively higher usage of Verification questions ā asking whether a given answer or piece of code is correct ā and Expectational questions, which ex- plore hypothetical outcomes. This pattern suggests that continuing- generation students approach the AI as an active problem-solving partner, directing it toward task completion, while first-generation students tend to adopt a more confirmatory role, using the AI to check and validate rather than to generate or explore. This distinc- tion, replicated across both independent sessions, constitutes the primary qualitative finding of this study. The two state machine diagrams Figures 3 and 4 depict how type of studentsā questions changes in each session. In session 1, most students start with Assertion questions (prob=0.48). In session 2, students used a wider variety of question types. They mostly start with an Instrumental/Procedural question (prob=0.43). We note that the second session took place two weeks after the first session Conferenceā17, July 2017, Washington, DC, USAMatin Amoozadeh and Amin Alipour Figure 2: Mean difference in Graesser question-type usage between continuing-generation and first-generation students across both ses- sions. Positive values indicate higher usage by continuing-generation students. An asterisk (*) denotes statistical significance at ķ< .05. Figure 3: State-transition diagram of Graesser question types in Session 1 (ķ= 60 students, 432 prompts). in which the students gained more maturity in their questions- asking skills. In both sessions, students tended to remain longer within Assertion, Request/Directive, and Instrumental/Procedural question states. 5 Discussion and Concluding Remarks First-Generation Students and AI-Supported Help-Seeking. We ob- served that first-generation students asked fewer questions than continuing-generation students, consistent with prior studies show- ing that continuing-generation students are generally more likely to seek help and engage in questioning behaviors [11,12]. Although Figure 4: State-transition diagram of Graesser question types in Session 2 (ķ= 37 students, 398 prompts). conversational AI may reduce some interpersonal barriers asso- ciated with classroom help-seeking, differences in inquiry behav- ior persisted in our study. In particular, during the second task involving more complex object-oriented programming concepts, first-generation students asked substantially fewer questions. These findings suggest that AI-supported tutoring systems may benefit from scaffolding mechanisms that encourage students to formulate questions and engage in more active help-seeking behaviors. Dynamics of Inquiry in AI-Supported Programming. A central theme emerging from our findings is that studentsā questioning behaviors evolved as programming tasks became more concep- tually demanding. In Session 1, students relied more heavily on assertion and verification prompts, often seeking confirmation of understanding for foundational concepts. Such behavior is charac- teristic of novice help-seeking patterns in programming contexts, where students frequently seek reassurance while navigating unfa- miliar concepts [8]. In contrast, Session 2 showed increased use of instrumental/procedural and request/directive prompts, suggesting a shift toward more directive and procedural forms of AI usage. This difference may reflect increasing familiarity with the conversational AI system, differences in task complexity, or evolving strategies for interacting with AI during programming activities. Importantly, the second assignment focused on inheritance and polymorphism, concepts commonly associated with higher cognitive complexity and persistent misconceptions in introductory object-oriented pro- gramming [2]. Additionally, the observation that a small subset of question types accounted for the majority of interactions suggests that student-AI conversations are often dominated by repeated help-seeking patterns. These findings highlight the importance of designing AI-supported programming environments that en- courage reflective inquiry and productive help-seeking rather than repeated confirmation-seeking or excessive procedural delegation. However, because the two sessions differed in both programming topics and conceptual complexity, these findings should be inter- preted cautiously. Analysis of Types of Inquiries in Student-AI InteractionConferenceā17, July 2017, Washington, DC, USA References [1]Matin Amoozadeh, Daye Nam, Daniel Prol, Ali Alfageeh, James Prather, Michael Hilton, Sruti Srinivasa Ragavan, and Amin Alipour. 2024. Student-ai interaction: A case study of cs1 students. In Proceedings of the 24th Koli Calling International Conference on Computing Education Research. 1ā13. [2]Brett A Becker et al.2024. Generative AI in Computing Education: Opportunities and Challenges. In Proceedings of the 55th ACM Technical Symposium on Computer Science Education. ACM. [3]James Finnie-Ansley et al.2024. Prompting Patterns in AI-Assisted Programming Education. arXiv preprint arXiv:2407.20792 (2024). [4]Arthur C Graesser and Natalie K Person. 1994. Question asking during tutoring. American educational research journal 31, 1 (1994), 104ā137. [5]Irene Hou, Sophia Mettille, Zhuo Li, Owen Man, Cynthia Zastudil, and Stephen MacNeil. 2024. The Effects of Generative AI on Computing Studentsā Help- Seeking Preferences. In Proceedings of the 26th Australasian Computing Education Conference. Association for Computing Machinery, 39ā48. doi:10.1145/3636243. 3636248 [6]Xiaoming Hu et al.2025. Generative AI and Conversational Tutoring Systems in Education: Opportunities and Challenges. Educational Technology Research and Development (2025). [7]Majeed Kazemitabaar et al.2024. CodeAid: Evaluating AI Copilots for Program- ming Education. arXiv preprint arXiv:2401.02262 (2024). [8]Sami Marwan et al.2019. Productive Help-Seeking in Computer Science Educa- tion: A Review. In Proceedings of the ACM Conference on International Computing Education Research. ACM. [9] Duy Phung et al.2023. Generative AI for Programming Education: Benchmarking ChatGPT, GPT-4, and Human Tutors. arXiv preprint arXiv:2306.17156 (2023). [10]James Prather et al.2024. Studentsā Use of Generative AI Tools in Programming Courses. Computer Science Education (2024). [11]Nicole M. Stephens, Stephanie A. Fryberg, Hazel Rose Markus, Camille S. John- son, and Rebecca Covarrubias. 2012. Unseen Disadvantage: How American Universitiesā Focus on Independence Undermines the Academic Performance of First-Generation College Students. Journal of Personality and Social Psychology 102, 6 (2012), 1178ā1197. doi:10.1037/a0027143 [12] Nicole M. Stephens, Sarah S. M. Townsend, Hazel Rose Markus, and L. Tay- lor Phillips. 2012. A Cultural Mismatch: Independent Cultural Norms Pro- duce Greater Increases in Cortisol and More Negative Emotions among First- Generation College Students. Journal of Experimental Social Psychology 48, 6 (2012), 1389ā1393. doi:10.1016/j.jesp.2012.06.005 [13] Kurt VanLehn. 2011. The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist 46, 4 (2011), 197ā221. doi:10.1080/00461520.2011.611369 [14]Cynthia Zastudil et al.2023. Generative AI in Computing Education: Perspectives and Experiences of Students and Instructors. Proceedings of the ACM Conference on International Computing Education Research (2023).