GenAI Knowledge, Epistemic Orientation, and Intellectual Values Predict Undergraduate Students' Critical GenAI Use

Synthesis: A correlational study (N = 67 undergraduate psychology students, Bielefeld University) testing two protective factors against uncritical GenAI overreliance: (1) knowledge about genAI and (2) the disposition to engage in critical thinking — operationalized via Kuhn's framework as epistemic orientation (tendency away from absolutist toward evaluativist beliefs) and intellectual values (viewing intellectual engagement as worthwhile). Both factors are framed as components of AI literacy and both are trainable, motivating intervention recommendations.

Method highlights

• Critical use of genAI measured with a newly developed 10-item scale (α = 0.78) covering validation behaviors (cross-checking outputs, verifying factual accuracy, comparing with literature).
• Knowledge measured two ways — the study's key methodological contribution: self-reported (18 SNAIL items, α = 0.90) AND objective/performance-based (12 items from Köhler & Hartig's ChatGPT scale + Hornberger et al.'s AI literacy test), each objective item paired with a certainty rating to compute a misconception score (correctness × certainty; Eitel/Hefter procedure).
• Dispositions: epistemic orientation (5 items, α = 0.70), intellectual values (14 fictive-person items, α = 0.90), plus need for cognition (NFC, 4-item short scale, α = 0.54).

Key findings

H1 — GenAI knowledge predicts critical use: SUPPORTED

• Self-reported knowledge: r = 0.55, p < 0.001; objective knowledge: r = 0.31, p = 0.011 — both significant.
• The stronger self-report correlation likely reflects common-method bias (both self-report Likert scales) rather than a substantive difference — the objective test cross-validates the finding.

H2 — Critical-thinking disposition predicts critical use: SUPPORTED

• Multiple regression: F(2, 64) = 17.19, p < 0.001, R² = 0.33 — a third of the variance in critical GenAI use.
• Epistemic orientation β = 0.36, p < 0.001; intellectual values β = 0.40, p < 0.001; low multicollinearity (VIF = 1.05), Durbin-Watson 2.23.

Exploratory — misconceptions are narrow but real

• Only 9% of participants had misconception scores below zero; just 3 of 12 items showed actual Misconceptions about AI (incorrect + high confidence).
• The standout: 97% confidently believed "GenAI performs web searches" (misconception score −3.00) — though the authors note current chatbots technically trigger web searches, so this item may be superseded by evolving tool capabilities.
• Most students correctly and confidently knew GenAI "may provide content that is not based on facts" (score 2.91, highest).

Exploratory — need for cognition (NFC)

• Actual genAI use correlated only with interest (r = 0.54) and perceived usefulness (r = 0.58) — not with knowledge or dispositions.
• Perceived usefulness negatively correlated with NFC (r = −0.27, p = 0.027): high-NFC students view GenAI utility more skeptically.
• NFC positively correlated with critical use (r = 0.25, p = 0.044): the motivational drive to invest cognitive effort bridges AI literacy and actual critical behavior.

What this means for practice

• Learners. Learn what the tool actually does before you rely on it: objective knowledge about GenAI predicted critical use (r = 0.31, p = 0.011) and self-reported knowledge more strongly (r = 0.55), and institutions should target conceptual misconceptions directly rather than offering basic tool training.
• Learners. Cross-check outputs against sources and the literature as a habit, because confident error is the common failure: 97% of participants confidently believed GenAI performs web searches, and only 9% of participants had a misconception score below zero.
• Learners. Treat intellectual engagement as worth the effort: intellectual values (β = 0.40) and epistemic orientation (β = 0.36) together accounted for about a third of the variance in critical GenAI use (R² = 0.33), and each is trainable.
• Learners. Account for your own motivation and self-knowledge: need for cognition correlated with critical use (r = 0.25) and negatively with perceived usefulness (r = -0.27), so low-NFC learners may need added interactivity or feedback, and understanding the costs of Cognitive Offloading is proposed as the missing "meta-knowledge" predictor.
• Learners. Use refutation-based material and video interventions that challenge absolutist beliefs rather than assuming knowledge alone changes behavior.

Limitations

• N = 67 undergraduate psychology students at one German university (59 female, 7 male, 1 n/a; mean age 22.58), recruited voluntarily through the university's research participant platform — the authors flag selection bias and note that psychology students' training in evaluating scientific sources may raise baseline epistemic orientation and intellectual values, while the 88% female sample limits generalizability.
• Critical GenAI use was measured with a new 10-item self-report scale (α = 0.78); the authors acknowledge it carries the biases of subjective instruments and recommend performance-based measures such as the GLAT and objective behavior logs.
• The design is correlational: "transitioning from correlational to experimental or longitudinal research designs is required to establish causality."
• The sample was powered (target N = 68) only for medium-large predictive relationships, leaving little power for small or nuanced exploratory effects, and the knowledge test's items go stale as tools evolve — the study itself flags the web-search item as a limitation of this kind.

Citation

Hefter, M. H., Paaßen, B., & Berthold, K. (2026). GenAI knowledge, epistemic orientation, and intellectual values predict undergraduate students' critical GenAI use. AI Educ., 2, 27