Thinking Less, Trusting More: GenAI's Impacts on Students' Cognitive Habits

This study of 299 STEM students reveals that trusting and routinely using generative AI significantly reduces cognitive engagement habits like reflection and critical thinking, with surprisingly higher susceptibility among students possessing strong technical confidence and risk tolerance rather than those with prior experience.

The Gist

The Big Idea: The "Cognitive Debt" Trap

Imagine your brain is a muscle. Just like a leg muscle, if you stop using it, it gets weak. But there's a twist: Generative AI (like ChatGPT) is like a very convenient, super-strong exoskeleton.

When you wear an exoskeleton, you can lift heavy boxes with zero effort. It feels amazing in the moment. But if you wear it all the time and never lift a box on your own, your leg muscles eventually atrophy (shrink). You become dependent on the machine, not just for the heavy lifting, but for the very ability to stand up.

This paper argues that students are currently wearing this "AI exoskeleton" so much that their "thinking muscles" are shrinking. They aren't just getting answers; they are losing the habit of wanting to figure things out.

---

The Two Main Questions the Researchers Asked

The researchers from Oregon State University wanted to know two things:

1. How: Does trusting and using AI every day actually change how students think?
2. Who: Which students are most likely to fall into this trap? (Is it the confused ones, or the confident ones?)

The Findings: What Happened?

1. The "Lazy Brain" Effect (How)

The study found a direct link: The more students trusted AI and used it as a routine habit, the less they thought deeply.

The researchers measured three specific "thinking habits":

• Reflection: Checking your own work ("Wait, does this actually make sense?").
• Need for Understanding: The itch to know why something works, not just that it works.
• Critical Thinking: Questioning evidence and looking for flaws.

The Result: Students who used AI heavily reported doing significantly less of all three. It's as if they stopped asking "Why?" and started just accepting "Here is the answer."

The Analogy: Imagine you are learning to cook.

• Without AI: You chop the onions, smell the spices, and taste the sauce. You might burn the garlic, but you learn how heat affects flavor.
• With Routine AI: You press a button, and a perfect meal appears. You eat it, but you never learn how to chop, season, or fix a mistake. Eventually, you forget how to cook entirely.

2. The Irony of the "Tech-Savvy" (Who)

This is the most surprising part. You might think the students who are most likely to get lazy with AI are the ones who are confused or bad at technology.

The Reality: The students most likely to stop thinking were the "Tech-Natives."
These are the students who:

• Love technology (Technophilic).
• Are comfortable taking risks (Risk Tolerance).
• Are very confident in their computer skills (Self-Efficacy).

The Analogy: Think of a professional race car driver. Because they are so good at driving, they might feel confident enough to take their hands off the wheel and let the car's autopilot drive for them. A nervous, inexperienced driver would never do that. But the expert driver? They trust the machine so much they forget to drive.

The study found that these "confident" students were actually the most vulnerable to losing their critical thinking skills because they trusted the AI too much, too quickly.

3. Experience Doesn't Save You

The researchers thought, "Maybe if students have used AI for a long time, they will learn to use it wisely."
False. Whether a student was a freshman or a senior, or whether they had used AI for years or just months, the result was the same. Experience did not protect them from "cognitive disengagement."

---

The "Cognitive Debt" Cycle

The authors coin a term called "Cognitive Debt."

• Technical Debt: In software, if you write "quick and dirty" code to finish a project fast, you have to fix it later. It costs more time in the long run.
• Cognitive Debt: When students use AI to skip the hard thinking now, they are borrowing against their future intelligence. They get the grade today, but they are building up a "debt" of skills they haven't learned.

The Cycle:

1. You use AI to save time.
2. Your brain stops practicing deep thinking.
3. You get worse at thinking on your own.
4. You feel you need AI even more because you can't do it alone.
5. You use AI even more.
6. Result: A downward spiral where you become less capable, but the AI becomes more necessary.

---

What Can We Do? (The Solution)

The paper suggests we can't just tell students "Don't use AI." We have to change the environment.

For Teachers:

• Add "Friction": Make it slightly annoying to use AI. For example, ask students to write a first draft without AI, then use AI to critique it, then rewrite it again by hand.
• Make it Personal: Create assignments where the AI can't easily give the answer, like asking for personal stories or unique local perspectives.
• The "Debug" Challenge: Let students use AI to generate code or an essay, but then grade them on how well they can find the AI's mistakes and fix them.

For AI Designers:

• Don't be a "Magic Box": Instead of just giving the answer, the AI should act like a "Provocateur." It should ask, "Are you sure about that?" or "Here are three different ways to solve this; which one do you think is right and why?"
• The "Bicycle for the Mind": The goal shouldn't be to replace the rider (the student), but to make the rider faster while still requiring them to pedal.

The Bottom Line

The paper warns us that while AI is a powerful tool, using it as a routine habit is training our brains to be lazy. The smartest, most confident students are the ones most at risk of losing their ability to think critically.

To fix this, we need to design learning and AI tools that force us to keep pedaling, ensuring that we use AI to amplify our thinking, not to replace it.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in understanding the long-term cognitive impacts of Generative AI (genAI) in STEM education. While prior research has focused on episodic use (e.g., using AI to bypass a single difficult problem), this study investigates routine use driven by trust.

The core problem is the potential emergence of a "cognitive debt cycle." Just as technical debt accumulates when short-term efficiency replaces sound design, cognitive debt may accrue when students routinely offload cognitive work to AI. The authors hypothesize that this routine offloading, fueled by trust and automation bias, may erode essential intellectual habits: Reflection, Need for Understanding, and Critical Thinking. Furthermore, the study questions whether students traditionally viewed as "tech-ready" (high in self-efficacy and risk tolerance) are paradoxically more vulnerable to this cognitive disengagement.

2. Methodology

The study employs a quantitative, survey-based approach grounded in Dual-Process Theory, Cognitive Offloading, and Automation Bias.

• Participants: $N=299$ STEM students from five North American universities.
• Data Collection: A structured survey administered via Qualtrics over three months (April–June 2025).
• Theoretical Model: A structural model was developed to test two Research Questions (RQs):
  • RQ1 (How): How do trust and routine genAI usage affect cognitive engagement?
  • RQ2 (Who): Which cognitive styles make students more prone to disengagement?
• Key Constructs:
  • Independent Variables: Trust in AI, Routine GenAI Usage.
  • Dependent Variables (Cognitive Engagement):
    • Reflection: Metacognitive monitoring (Motives + Strategies).
    • Need for Understanding: Intrinsic drive for conceptual coherence.
    • Critical Thinking: Analytic evaluation (Openness + Skepticism).
  • Control Variables (Cognitive Styles): Technophilic Motivations, Risk Tolerance, Computer Self-Efficacy.
• Analysis Technique: Partial Least Squares Structural Equation Modeling (PLS-SEM) using SmartPLS 4. This method was chosen for its ability to handle complex models with latent variables and non-normal data distributions.
• Validation: The survey instrument was psychometrically validated using:
  • Convergent validity (Factor loadings $\ge$ 0.60, AVE $\ge$ 0.50).
  • Internal consistency (Cronbach's $\alpha$ and Composite Reliability $\ge$ 0.70).
  • Discriminant validity (HTMT ratios < 0.90).
  • Common Method Bias checks (Harman's single-factor test and VIF analysis).

3. Key Contributions

1. Theoretical Model: A statistically grounded model explaining the mechanism by which trust-driven routine genAI use degrades cognitive engagement habits in STEM coursework.
2. Empirical Evidence of "Cognitive Debt": Quantitative proof that routine offloading to AI correlates with a significant reduction in deep cognitive processing (reflection, understanding, critical thinking).
3. Counter-Intuitive Demographic Findings: Identification that students with traits typically celebrated in STEM (high self-efficacy, risk tolerance, and technophilic motivation) are more susceptible to cognitive disengagement, not less.
4. Validated Instrument: A psychometrically validated survey tool for measuring cognitive engagement and AI interaction constructs in educational settings.

4. Key Results

The PLS-SEM analysis yielded the following significant findings:

RQ1: The Impact of Trust and Usage (How)

• Trust $\rightarrow$ Usage: Trust is a strong predictor of routine genAI usage ($\beta = 0.64, p < .001$).
• Usage $\rightarrow$ Cognitive Engagement: Routine usage has a significant negative impact on all three dimensions of cognitive engagement:
  • Reflection: Large negative effect ($\beta = -0.66, f^2 = 0.51$).
  • Critical Thinking: Moderate negative effect ($\beta = -0.41, f^2 = 0.20$).
  • Need for Understanding: Small but significant negative effect ($\beta = -0.21, f^2 = 0.12$).
• Mediation: Trust affects cognitive engagement entirely through its effect on routine usage (Full Mediation, VAF > 0.80). High trust leads to high usage, which in turn leads to cognitive disengagement.

RQ2: The Role of Cognitive Styles (Who)

• Predictors of Usage: Students with higher Technophilic Motivations, Risk Tolerance, and Computer Self-Efficacy reported significantly higher routine genAI usage.
• The Paradox: Contrary to the assumption that these traits protect learners, they were associated with lower cognitive engagement when mediated through high AI usage.
  • High self-efficacy and risk tolerance students showed significantly reduced Reflection, Need for Understanding, and Critical Thinking compared to peers with lower scores in these traits.
• Robustness: Multi-Group Analysis (MGA) confirmed that these relationships hold regardless of academic level (lower vs. upper division) or prior genAI experience. Experience does not act as a protective factor.

Model Adequacy

• The model explains 54% of the variance in Usage and 65% of the variance in Reflection.
• The model fit (SRMR = 0.067) and predictive relevance ($Q^2 > 0$) indicate a robust and well-specified model.

5. Significance and Implications

The study challenges the narrative that AI literacy or technical proficiency alone will mitigate the risks of AI in education.

• The "Cognitive Debt Cycle": The authors propose a self-reinforcing cycle where routine AI use weakens the very cognitive muscles (reflection, critical thinking) required to use AI effectively. As these skills atrophy, students become increasingly dependent on AI, accelerating the cycle.
• Educational Implications:
  • Awareness is insufficient: Prior experience and academic seniority do not inoculate students against disengagement.
  • Pedagogical Redesign: Educators must design "friction" into learning. This includes alternating AI-assisted and independent work, requiring students to critique/debug AI outputs, and using curiosity-driven puzzles that expose AI errors.
• Design Implications:
  • Complementarity over Substitution: AI systems should be designed as "bicycles for the mind" (augmenting human thought) rather than replacements.
  • Intentional Friction: Systems should be designed to prompt questioning, require justification of choices, and present multiple perspectives to force System 2 (deliberate) thinking.
  • Adaptive Interfaces: Systems should adapt to cognitive styles, potentially nudging high-risk-tolerance or high-self-efficacy users to engage more deeply before accepting AI suggestions.

Conclusion: The paper concludes that without deliberate intervention in curriculum design and AI system architecture, the "efficiency" gained from genAI may come at the cost of a generation of STEM professionals who are adept at querying intelligence but have lost the habit of exercising their own.