Trust-utility gap in introductory physics education: Students' adoption, domain-specific skepticism, and preferences for AI integration

Synthesis: Fouad & Bentley (2026) survey 81 introductory physics students and find a striking 50-percentage-point trust-utility gap: 91% use AI for coursework but only 41% Trust AI physics explanations — evidence of domain-calibrated skepticism, not uncritical adoption. Students spontaneously identified where AI fails in physics (visual-spatial reasoning, circuits, abstract reasoning), aligning with known benchmarks. 65% prefer optional over mandatory AI integration.

A mixed-methods survey of 81 introductory physics undergraduates reveals a 50-point trust-utility gap: 91% [95% CI: 83%–96%] use AI for physics coursework, but only 41% [95% CI: 30%–52%] Trust AI-generated physics explanations. Thematic analysis identified eight themes; the most distinctive finding was that 40% of qualitative respondents spontaneously articulated where AI fails — visual-spatial reasoning, circuit analysis, and abstract physical reasoning — aligning with known benchmarks. Students show domain-calibrated skepticism rather than uncritical adoption, and 65% prefer optional AI integration.

• 91% use AI for physics, only 41% trust it — 50-point trust-utility gap
• Students show domain-specific skepticism, not naive acceptance
• 40% spontaneously identified AI failure modes: visual-spatial reasoning, circuits, abstract physics
• Student trust calibration tracks AI competence boundaries (consistent with Kortemeyer benchmarks)
• 65% prefer optional over mandatory AI integration
• Verification gap identified: self-reported vs actual verification behavior

What this means for practice

• Learners. Verify physics AI output yourself, and concentrate that effort on diagrams, circuits, and abstract reasoning — the three areas students spontaneously identified as failures, and the ones where dimensional analysis and visual-spatial reasoning decide whether an answer is right.
• Learners. Do not read your own trust as a quality signal. Use and trust diverge sharply in this sample (91% use AI for coursework, 41% Trust its explanations), and self-reported checking likely overstates what you actually do under time pressure.
• Instructors. Teach physics-specific verification rather than assuming that tool access builds judgment: the students least equipped to catch AI errors in force diagrams or circuit analysis are the ones who most need Scaffolding.
• Instructors. Redesign out-of-class assessments and keep AI guidance optional. 92% of respondents find AI helpful for homework and 65% prefer optional over mandatory integration, so proctored in-class work plus optional support matches what students will accept.
• Administrators. Maintain peer tutoring and office hours alongside AI. Students assign AI the 24/7 availability and step-by-step algebra role and human tutors the diagram-based, whiteboard, conceptual role, so the two are complements rather than substitutes.

Limitations

• Small single-site sample. N = 81 introductory physics students at one STEM-focused technical university, an 18.12% response rate, and an academically skewed sample: 83.0% self-reported A/B grades against 69.1% institutional A/B final grades.
• Underpowered comparisons. Post hoc power was 80% for medium effects (Cramér's V ≥ 0.30) but only 45% for small effects, and the C/D (n = 13) and algebra-based (n = 13) subgroups reached roughly 25% power, so all subgroup results are reported descriptively.
• Entirely self-reported and cross-sectional. Data collection fell in weeks 10–13, which precludes end-of-semester correlations and any causal claim; the gap between reported and actual verification behavior is described by the authors as the most significant limitation of the study.
• Single-item trust measure. Trust was captured with one survey item rather than a validated multi-item scale, and the specific AI platform students used was not recorded.

Citation

Fouad, E., & Bentley, I. (2026). Trust-utility gap in introductory physics education: Students' adoption, domain-specific skepticism, and preferences for AI integration.