Students Use AI as Answer Machine or Verifier, Rarely as Thinking Partner

Generative artificial intelligence can now solve an entire quantitative STEM problem in seconds, complete with steps, explanations, and interpretations delivered in fluent natural language. That capability has ignited a fierce debate among educators: when a chatbot can produce a polished worked solution on demand, what exactly are students doing cognitively when they turn to it? A new study published in the International Journal of STEM Education offers one of the most detailed answers yet, and its central finding is striking. Given the same task, students in the same classroom used generative AI in profoundly different ways, and almost none of them used it the way many technologists promise—as a genuine partner in reasoning.

Researchers Aparajita Jaiswal and Gaurav Nanda of Purdue University studied 38 undergraduate students enrolled in a junior-level operations management course during a unit on linear programming in manufacturing contexts. The design of the assignment was deliberate. Students first solved a linear programming problem manually using the graphical corner-point method, which requires defining an objective function, identifying constraints, representing the feasible region, and finding the corner point that optimizes the objective. Only afterward did they solve the same problem using ChatGPT. The students then submitted written reflections describing how their manual reasoning compared with the AI’s approach, whether the AI helped them identify errors or confirm their logic, and what they learned about concepts such as optimal solutions, shadow prices, and sensitivity analysis. Where possible, the researchers also collected the students’ actual ChatGPT interaction records—chat transcripts, share links, screenshots, and uploaded images—to check whether what students said matched what they did.

To make sense of this variation, the team applied the ICAP framework, a well-established model from the learning sciences developed by Michelene Chi and Ruth Wylie. ICAP sorts cognitive engagement into four hierarchical modes: Passive, Active, Constructive, and Interactive. Passive engagement means simply receiving information; Active engagement means manipulating what is given, such as checking or confirming work; Constructive engagement means generating new ideas beyond the material, such as asking for explanations or drawing inferences; and Interactive engagement means dialogic co-construction, in which both parties make generative, contingent contributions that build on each other. The framework’s core claim is that deeper modes predict stronger learning—a claim the study put directly to the test in the context of human-AI interaction.

The results revealed three clearly documented patterns. Passive engagement appeared when students submitted the problem and read the AI’s output without using the interaction to compare, verify, reinterpret, or extend anything. One student wrote plainly, “With ChatGPT, I just copy pasted the question/problem and then it showed me the steps and got the correct solution.” Another described the process as “a simple copy and paste and it spit out the answer with work shown.” Interestingly, the passive cases were not uniform. Some students provided minimal direction; others uploaded course materials or submitted images of the problem so the AI would follow the class method. What united them was the absence of any intentional follow-through: noticing that the AI skipped the graph or omitted a feasible point, for instance, did not translate into probing questions or deliberate verification.

Active engagement looked different. Here, students positioned ChatGPT as a confirmation or reassurance tool, using it to check work they had already completed by hand. “AI confirmed my answer and the logic used seems to be the same,” one student reported. Another noted, “The AI approached the problem the same way I did and helped me confirm that I was following the steps correctly without error.” A third summed it up bluntly: “All it did to help was verify my answers and method.” In these cases, students were doing more than receiving—they were comparing and validating—but the interaction remained centered on correctness rather than on building new understanding. One illustrative case showed a student who asked ChatGPT to solve the problem with the corner-point method, received a mathematically accurate solution identifying the optimum at the point (0,3) with an objective value of 45, and then requested a sensitivity analysis. Although the second prompt extended the content, the exchange stayed confirmatory in structure, which is why the researchers classified it as Active rather than Constructive.

Constructive engagement emerged when students sought explanations and interpretation, particularly around the concepts they found hardest: sensitivity analysis, shadow prices, and binding versus non-binding constraints. “I didn’t really understand how sensitivity analysis and shadow prices worked until it explained them to me,” one student wrote. Another said the AI “pointed to details I would miss out on, such as identifying binding constraints and shadow prices,” and a third reflected that “sensitivity analysis made more sense once the AI broke it down in normal language.” In these cases, students used the chatbot not to obtain an answer but to understand what the answer meant—how a shadow price of 3.75 on a binding constraint translates into the value of relaxing that constraint, or how allowable ranges describe how far coefficients can shift before the optimal solution changes. The researchers were careful to note, however, that the study did not assess whether these AI-generated explanations actually improved conceptual understanding on later tests or transfer problems.

The most theoretically significant finding was what the researchers did not find. No student reflection and no available interaction record met the operational criteria for Interactive engagement, the highest ICAP mode requiring iterative, contingent co-construction in which each party’s contributions build on and challenge the other. Even students who used multi-turn exchanges tended to remain in a request-response pattern, soliciting more information without refining, challenging, or jointly developing ideas with the system. Longer conversations, the study emphasizes, were not automatically deeper ones: some multi-turn interactions stayed verification-oriented from start to finish. The authors are cautious here—the finding represents an absence of documented Interactive engagement, not proof that no unrecorded interactive exchange occurred, since many submitted records were incomplete. Still, the pattern is telling.

Why didn’t genuine dialogue with AI emerge? The instructional context offers clues. The task was convergent and well structured, with a known method and a single correct answer—conditions that naturally cue answer-seeking rather than exploration. Students received no explicit guidance on how to prompt the system dialogically, no models of iterative questioning, and no assessment incentives rewarding sensemaking; the assignment was graded for completion. Under those conditions, there was little reason for students to adopt an interactive stance. The authors argue that this absence should be read not as a failure of students or technology but as a design signal: interactive cognition with AI must be deliberately invited, scaffolded, and normed, and it will not appear simply because a conversational interface exists.

The broader implication is that “AI use” is not a meaningful analytic category in itself. Two students can produce identical, correct solutions—one by copying a prompt, the other by interrogating an explanation of shadow prices—and the cognitive consequences are worlds apart. The authors suggest that prompting should be understood as a cognitive and epistemic practice rather than a mere technical skill: a single copy-and-paste submission reflects a fundamentally different orientation toward learning than a sequence of targeted why, what if, and how would this change questions. For educators, the takeaway is concrete. If assessments reward only final answers, students will use AI passively or for verification. If tasks require explaining why a method works, comparing alternative approaches, or reasoning about how changed assumptions alter results, AI can be repositioned as a resource for reasoning rather than a shortcut around it. The educational value of generative AI, the study concludes, lies not in the tool itself but in the forms of thinking that instruction invites students to enact with it—and right now, in at least one typical quantitative course, those invitations mostly stop short of genuine dialogue.

Subject of Research: Cognitive engagement patterns of undergraduate students using generative AI for technical problem solving

Article Title: More than getting the answer: How students engage with generative AI in technical problem solving

Article References: Jaiswal, A., & Nanda, G. (2026). More than getting the answer: How students engage with generative AI in technical problem solving. International Journal of STEM Education, 13(1), Article 58. https://doi.org/10.1186/s40594-026-00646-7

Image Credits: AI Generated

DOI: 10.1186/s40594-026-00646-7

Keywords: generative AI, STEM education, cognitive engagement, ICAP framework, linear programming, ChatGPT, problem solving, sensitivity analysis, shadow prices, instructional design, higher education, than