Three Pathways of Student-AI Interaction: Constraint-First Design for Higher-Order Thinking

Introduction

• Higher education has largely responded to generative AI with reactive permit-or-prohibit policies, leaving underexamined how interaction shapes students’ thinking.The paper identifies this as a more consequential question than whether or where students may use AI.Source: Introduction · p. 3
• The Three Paths framework distinguishes Passive Review, Direct Question, and Strategic Dialogue as qualitatively different modes of student-AI engagement.The paper contrasts merely submitting a prompt and transcribing the response with using AI to pressure-test one's reasoning.Source: Introduction · p. 3

Cognitive Offloading and the Stakes of AI Interaction

• Distributed cognition suggests that routinely offloading cognitive work can reduce opportunities to practice, while Bastani et al. (2024) found lower performance after AI assistance was removed.The paper frames this as a risk of dependency replacing learning.Source: Cognitive Offloading and the Stakes of… · p. 4
• Prior studies describe students deferring evaluation of AI output as “metacognitive laziness” and AI-assisted ideation producing more but less diverse ideas.The paper also notes an “illusion of understanding” when users mistake AI-processed material for their own comprehension.Source: Cognitive Offloading and the Stakes of… · pp. 4–5
• Across three sessions, Kosmyna et al. (2025) found weaker neural connectivity, lower writing ownership, and poorer recall among LLM users; in a tool-free fourth session, their neural engagement was reduced.The comparison was with students who had written independently throughout.Source: Cognitive Offloading and the Stakes of… · p. 4
• AI use frequency and over-reliance have been associated with weaker critical thinking, while repeated assistance has been associated with reduced epistemic agency.Lee et al. (2025) found that higher AI confidence was associated with less critical-thinking engagement, whereas higher self-confidence was associated with more.Source: Cognitive Offloading and the Stakes of… · pp. 4–5
• Vygotsky’s zone of proximal development frames productive AI use as scaffolding that extends what learners can do without doing the reasoning for them.The paper argues that what LLMs do for learning depends on pedagogical conditions.Source: Cognitive Offloading and the Stakes of… · p. 5
• Higher-order thinking includes analyzing, evaluating, creating, and metacognitive reflection; generative AI can perform cognitive work at these levels, making critical thinking a key competency for pedagogical design.The paper defines critical thinking as an important component, but not the entirety, of higher-order thinking.Source: Cognitive Offloading and the Stakes of… · p. 5

• AI literacy means critically understanding, evaluating, and responsibly deploying AI, including recognizing its sources and blind spots and when disciplinary judgment should outweigh its answers.It develops through repeated critical use in contexts where AI limitations matter.Source: AI Literacy and Critical Thinking as… · p. 6
• The Three Paths map onto an AI-literacy trajectory: Path 1 treats AI as authoritative, Path 2 uses it instrumentally with some evaluation, and Path 3 interrogates its reasoning while maintaining epistemic agency.These correspond to low, emerging, and high AI literacy, respectively.Source: AI Literacy and Critical Thinking as… · p. 6
• AI literacy and critical thinking are co-primary, mutually reinforcing goals; pedagogy that develops one without the other is insufficient.Source: AI Literacy and Critical Thinking as… · p. 6

Critical Pedagogy and Epistemic Agency

• Freire’s “banking model” can recur when students transcribe AI output, so the NLTB bars passive receipt and keeps students epistemic agents, with AI as a resource rather than authority.Source: AI Literacy and Critical Thinking as… · pp. 6–7

The Three Paths of Student-AI Interaction

• The typology identifies three qualitatively distinct modes of student-AI engagement, linking differences in interaction to cognitive and epistemic engagement.It distinguishes Passive Review, Direct Question, and Strategic Dialogue by student cognitive agency and the nature of the AI interaction.Source: The Three Paths of Student-AI… · p. 7
• In Path 1, students accept AI output with little cognitive investment, leaving little opportunity to exercise critical thinking.The account frames this pattern more often as a literacy deficit than a moral one: students may not understand that language models generate probabilistic text rather than verified claims.Source: The Three Paths of Student-AI… · p. 9
• Path 2 students formulate targeted questions, evaluate and integrate responses, but the exchange remains transactional and offers limited sustained reasoning evaluation or revision.Question formulation requires metacognitive awareness, making this path more cognitively demanding than Path 1.Source: The Three Paths of Student-AI… · p. 10
• In Path 3, students share and stress-test their own reasoning with AI, evaluating its responses and synthesizing intellectual movement rather than repeating its output.The path requires a position to defend, compare, and revise, engaging analysis, evaluation, and synthesis.Source: The Three Paths of Student-AI… · pp. 10–11
• Learning outcomes in AI-integrated education depend less on whether students use AI than on which interaction path they take.Source: The Three Paths of Student-AI… · p. 11

Methods

• The logs came from an undergraduate research methods course using NLTB-derived protocols that gave students structured but open-ended access to an AI teaching assistant throughout the semester.The logs served as a naturalistic record for examining the typology.Source: Methods · p. 11
• Researchers randomly sampled 50 complete student turns from a corpus of 580 messages across seven assignments by 14 students for path classification.The sample was intended to capture the range of engagement across the course rather than selected examples.Source: Methods · p. 12
• A separate dataset of 57 messages from 10 students in a different assignment context was examined for content patterns but not formally coded.Source: Methods · p. 12
• Human coders achieved 68% path-level agreement (κ = .48) and agreed on Path 3 in 80% of cases (40/50); Path 2 appeared in 11 of 16 discrepancies.When a message received multiple labels, each coder’s first label determined its path classification.Source: Methods · p. 13
• Three-coder majority voting yielded P1 = 48%, P2 = 16%, and P3 = 36%, close to R1’s classifications of P1 = 46%, P2 = 18%, and P3 = 36%.R1’s final classifications were used for the primary path distributions, with the majority-vote distribution serving as a supplementary comparison.Source: Methods · pp. 13–14
• GPT-5 added calculation_check for computational-accuracy queries (28% of messages) and classified 28% as Path 3, versus 36% for R1 and 24% for R2.The additional category was absent from the human scheme, reducing direct comparability and producing low pairwise kappas between GPT-5 and the human coders.Source: Methods · pp. 13–14

Finding 1: The Three Paths Are Empirically Distinguishable

• In 50 double-coded messages, R1 classified 46% as Path 1 (Passive Review), 18% as Path 2 (Direct Question), and 36% as Path 3 (Strategic Dialogue).The messages were classifiable into three patterns corresponding to the typological paths.Source: Finding 1: The Three Paths Are… · p. 14
• Across R1, R2, and GPT-5, majority-vote shares were 48% Path 1, 16% Path 2, and 36% Path 3; GPT-5 also used a calculation_check subcategory.The distribution remained consistent with the typology’s core structure.Source: Finding 1: The Three Paths Are… · p. 14
• Paths 1 and 2 together represented 64% of the primary sample, consistent with the typology’s prediction that Strategic Dialogue is unlikely to be students’ default.Source: Finding 1: The Three Paths Are… · p. 14
• A second dataset of 57 messages from 10 students, drawn from a different AI-prompted reflection assignment, mostly showed applicative and policy-critical engagement consistent with Path 3.Source: Finding 1: The Three Paths Are… · p. 14
• Because the second dataset was not systematically coded, its contrast offers only a preliminary indication that instructional framing may influence path engagement.Source: Finding 1: The Three Paths Are… · pp. 14–15
• In a separate study of AI-assisted brainstorming in graduate education, Shah et al. (2024) found that only 13% of students critically evaluated and built upon AI recommendations, while 36% accepted aligned suggestions without interrogation.Source: Finding 1: The Three Paths Are… · p. 15
• Yang and Ma’s (2025) epistemic-relationship classification conceptually aligns with validator-versus-co-thinker engagement; these studies suggest cross-study consistency, though direct comparison requires caution.Source: Finding 1: The Three Paths Are… · p. 15

Finding 2: Instructional Framing Shifts Path Distribution

• The second dataset comprised 57 messages from 10 students completing scaffolded reflection tasks requiring reasoning with hypotheses and policy implications rather than answer-checking.Prompts also called for applicative reasoning.Source: Finding 2: Instructional Framing… · p. 15
• The primary dataset was dominated by messages seeking validation or correction, consistent with its 64% Path 1/2 distribution; the second was dominated by probing messages, Path 3 interactions by design.Source: Finding 2: Instructional Framing… · p. 15
• Because the assignments differed in content and AI prompt structure, this uncontrolled contrast does not establish causation; the pattern is consistent with NLTB design logic.In the second dataset, the design made Path 1 and Path 2 responses insufficient.Source: Finding 2: Instructional Framing… · p. 16

Finding 3: GPT-5 as a Methodological Partner

• Across the same 50 messages and codebook, GPT-5 introduced calculation_check for computational-accuracy checks on completed work, a shared pattern humans had split across two improvement-request labels.GPT-5 was asked to apply the codebook, not revise it.Source: Finding 3: GPT-5 as a Methodological… · p. 16
• Although it reduced direct comparability, the category surfaced a recurring request that may clarify the Path 1/2 boundary; future work can test whether it improves clarity and agreement.The category identifies requests to check computational accuracy in completed work.Source: Finding 3: GPT-5 as a Methodological… · p. 16
• The coding exercise illustrates that task structure can shape outputs, but GPT-5's imperfect agreement with human coders in one exercise does not establish typology generalization beyond student–AI interaction.GPT-5 identified a category neither human coder had named, but the authors caution against generalizing from this single exercise.Source: Finding 3: GPT-5 as a Methodological… · pp. 16–17

The Next Level Teaching Blueprint

• The NLTB is a replicable, three-stage instructional design system with explicit interaction protocols and a theoretical logic linking each element to Path 3 engagement.It was first implemented in this study.Source: The Next Level Teaching Blueprint · p. 17

Mechanism

• Stage 1 requires students to engage the problem before AI use and document their reasoning, hypotheses, and uncertainties, giving them material for genuine Path 3 dialogue.Without this work, Stage 2 may default to Path 1 or 2; the documentation also gives instructors visibility into unmediated student thinking.Source: Mechanism · p. 18
• Stage 2 structures Path 3 dialogue: students share their Stage 1 reasoning with AI, ask it to challenge their thinking, and probe its assumptions and alternatives.Students may not ask AI for answers, summaries, or drafts.Source: Mechanism · p. 19
• Within Stage 2’s structural constraints, students steer the reasoning while AI expands the problem space; disciplinary implementation can vary, but Paths 1 and 2 are not the goal..Source: Mechanism · pp. 17–19
• Stage 3 returns students to their Stage 1 documentation to compare original and AI-augmented thinking, explain changes, and construct an integrated position.Source: Mechanism · p. 19
• Students evaluate the exchange and revise their position, producing an account of intellectual movement rather than an AI summary, with evaluation and synthesis built into the assignment.Source: Mechanism · p. 19

Scaling: From Pilot to Campus-Wide Initiative

• After the pilot, a $22,100-funded campus initiative across two phases involves six faculty in three colleges and targets 12 to 15 faculty and approximately 350 students over three years.Its cross-disciplinary adoption suggests transferability, but formal evaluation is beyond the paper’s scope.Source: Scaling: From Pilot to Campus-Wide… · p. 19

Discussion

• Two coders distinguished the three paths with moderate agreement, including 80% agreement on Path 3, supporting the typology’s claim that engagement mode shapes students’ thinking.The framework describes interaction modes that can shift with design, rather than fixed student types.Source: Discussion · p. 20
• The primary sample was 64% Path 1/2, whereas the reflection-structured assignment dataset was predominantly Path 3, suggesting instructional design may influence students’ interaction paths.The authors present this contrast as suggestive, not conclusive.Source: Discussion · p. 20
• Across seven course assignments, Path 3 engagement varied with assignment design and in-class scaffolding, indicating that interaction modes can shift when design changes.Figure 2 presents the trajectory of path distributions across all seven assignments.Source: Discussion · p. 20
• Paths 1 and 2 accounted for 64% of coded interactions, the most directly actionable result; without deliberate scaffolding, students default to interaction modes offering the least cognitive resistance.Source: Discussion · p. 20
• The NLTB responds to insufficient calls to “use AI critically” by requiring students to bring their own reasoning, expose it to challenge, and account for what changed.Without those requirements, a polished answer or quick verification can remain sufficient.Source: Discussion · p. 21
• A GPT-5 finding raises the possibility that engagement mode influences AI contributions in other cognitive tasks, but this interpretation rests on one coding exercise and needs further investigation.Source: Discussion · p. 21
• The discussion frames institutional AI policy as instructional: prohibition forecloses Path 3 benefits, while permissive use without scaffolding can produce Path 1/2 users at scale.The NLTB is presented as an architecture for making Path 3 the structural default by designing engagement conditions rather than restricting AI.Source: Discussion · p. 21

Limitations and Future Directions

• The study cannot answer whether interaction-data cognitive displacement translates into measurable learning differences; same-course pre-post outcomes are reported separately and not presented here.The unresolved learning question is described as the study’s most pressing unanswered question.Source: Limitations and Future Directions · pp. 21–22
• The coded sample comprised 50 messages from 14 students in one institution and course; path-level κ = .48, with the Path 1/2 boundary the most contested.A second dataset (n = 57) from the same course but a different assignment was not formally coded and serves only as descriptive context.Source: Limitations and Future Directions · pp. 21–22
• The immediate priority is a formal comparison of scaffolded and unscaffolded conditions to test the design claim at the center of NLTB.The paper identifies this comparison as its immediate research priority.Source: Limitations and Future Directions · p. 22
• Future work should expand coding across institutions, compare NLTB-designed with conventionally taught courses, refine the Path 1/2 boundary, and examine varied higher-education populations.The populations named include first-generation, international, non-Anglophone, and non-US students.Source: Limitations and Future Directions · p. 22

Conclusion

• The Three Paths framework gives differences in students’ cognitive work with AI a language, and the NLTB translates that language into an instructional sequence.The conclusion argues that AI use alone is too coarse a category to guide pedagogy.Source: Conclusion · p. 22
• Pilot data indicate that the three interaction modes can be identified, although moderate inter-rater agreement and a small sample mean the coding scheme needs refinement.The conclusion characterizes the findings as preliminary rather than large-scale evidence.Source: Conclusion · p. 22
• The author plans to continue testing and refining the framework and hopes others will do likewise in their own settings.The conclusion says the framework needs testing across larger samples, disciplines, and settings.Source: Conclusion · pp. 22–23