Why Johnny Can't Think: GenAI's Impacts on Cognitive Engagement

TL;DR

The paper examines how trust in generative AI and routine usage shape students’ cognitive engagement in STEM, theorizing two pathways: trust drives routine use, which in turn reduces Reflection, Need for Understanding, and Critical Thinking; cognitive styles modulate these effects. Empirically, the authors validate a theory-driven model using PLS-SEM on 299 STEM students across five universities, showing strong negative links from Routine Usage to all engagement dimensions, and a strong Trust→Usage link that fully mediates engagement. Notably, technophilic motivations, risk tolerance, and computer self-efficacy increase AI usage and—paradoxically—associate with greater cognitive disengagement, challenging assumptions about tech-ready students. The study introduces a psychometrically validated survey instrument and discusses a cognitive debt cycle with actionable implications for curricula and AI system design to preserve and augment human thinking rather than replace it. Overall, the work highlights the need for designing educational experiences and genAI tools that encourage active cognitive labor and mindful AI use to sustain durable learning outcomes.

Abstract

Context: Many students now use generative AI in their coursework, yet its effects on intellectual development remain poorly understood. While prior work has investigated students' cognitive offloading during episodic interactions, it remains unclear whether using genAI routinely is tied to more fundamental shifts in students' thinking habits. Objective: We investigate (RQ1-How): how students' trust in and routine use of genAI affect their cognitive engagement -- specifically, reflection, need for understanding, and critical thinking in STEM coursework. Further, we investigate (RQ2-Who): which students are particularly vulnerable to these cognitive disengagement effects. Method: We drew on dual-process theory, cognitive offloading, and automation bias literature to develop a statistical model explaining how and to what extent students' trust-driven routine use of genAI affected their cognitive engagement habits in coursework, and how these effects differed across students' cognitive styles. We empirically evaluated this model using Partial Least Squares Structural Equation Modeling on survey data from 299 STEM students across five North American universities. Results: Students who trusted and routinely used genAI reported significantly lower cognitive engagement. Unexpectedly, students with higher technophilic motivations, risk tolerance, and computer self-efficacy -- traits often celebrated in STEM -- were more prone to these effects. Interestingly, prior experience with genAI or academia did not protect them from cognitively disengaging. Implications: Our findings suggest a potential cognitive debt cycle in which routine genAI use progressively weakens students' intellectual habits, potentially driving over-reliance and escalating usage. This poses critical challenges for curricula and genAI system design, requiring interventions that actively support cognitive engagement.

Figures (6)

• Figure 1: Proposed theoretical model and measurement specification. The blue region (RQ1-How) models how Trust and routine Usage affect cognitive engagement: (a) Reflection, (b) Need For Understanding, and (c) Critical Thinking. The yellow region (RQ2-Who) controls for cognitive styles (Technophilic Motivations, Risk Tolerance, Computer Self-Efficacy). Latent constructs (circles) are reflectively measured by survey items (squares) russo2021pls; higher-order constructs (double circles) are reflectively measured via their lower-order constructs (e.g., Reflection is measured via Reflective Motives & Strategies). Reverse-coded items carry the “-R” suffix (e.g., RT1-R). Directed arrows represent hypothesized relationships (H1–H4) and cognitive-style controls. Item wordings (questionnaire) are in supplemental.
• Figure 2: Cognitive engagement scores by genAI usage. Boxplots show standardized latent scores for Reflection, Need for Understanding, and Critical Thinking across quartiles of reported genAI usage (Q1 = lowest 25%; Q4 = highest 25%). This figure provides a close-up view of data distributions; subsequent figures examine patterns using median-split groups.
• Figure 3: Structural model results (RQ1-How): Standardized path coefficients (H1–H4) on arrows between constructs (circles) show how Trust and routine genAI Usage affect Reflection, Need for Understanding, and Critical Thinking, with significance ***p<.001, **p<.01, and *p<.05. Corresponding transitive effects are summarized in Tab. \ref{['tab:new_path_analysis']}. Arrows from constructs to survey items (squares), and from HOCs (double circles) to their lower-order constructs, indicate factor loadings (in blue; all $\ge$ threshold of 0.6 hair2019use). Dotted arrows and the collapsed yellow region correspond to RQ2-Who results (omitted here for clarity), which are reported in Sec. \ref{['sec-RQ2']}.
• Figure 4: Trust $\rightarrow$ usage $\rightarrow$ cognitive engagement distributions. Boxplots show standardized latent scores for Reflection, Need for Understanding, and Critical Thinking among participants grouped by genAI usage (median split) and trust in genAI (low vs. high; median split within each usage group).
• Figure 5: Structural model path associations (full model: RQs1&2). Standardized path coefficients show direct effects of Trust and cognitive styles (Technophilic Motivations, Risk Tolerance, Computer Self-Efficacy; RQ2—yellow) on Usage, and of Usage on Reflection, Need for Understanding, and Critical Thinking, with significance ***p<.001, **p<.01, and *p<.05. Paths from LOCs (circles) to items (squares), and from HOCs (double circles) to LOCs, indicate factor loadings (in blue; all $\ge$ 0.6 hair2019use). Transitive effects of cognitive styles on cognitive engagement are reported in Tab. \ref{['tab:RQ2-controls']}.