Self-reported learning-oriented GenAI use under perceived teacher-referenced normative tension in Chinese higher education

Synthesis: A survey of 1,004 students at three Chinese public universities asks whether the awkwardness of not knowing where one's Generative AI use sits relative to teacher expectations — perceived teacher-referenced normative tension — goes together with self-reported learning-oriented use. It does, and positively: tension and learning-oriented use coexist rather than sit at opposite ends of one scale. The finding resists the common reading that unsettled norms simply suppress productive use. But the design is cross-sectional and same-source, so the authors are explicit that the association does not establish that tension improves behavior; a reverse direction and shared unmeasured traits fit the data equally well.

Key Findings

1. Perceived teacher-referenced normative tension (PNT) was positively associated with self-reported learning-oriented GenAI use (LOU) at β = 0.568, p < 0.001, with the model explaining 32.2% of the variance in LOU.
2. The association held after adjustment for GenAI guidance clarity (β = 0.478, p < 0.001), for demographic and institutional covariates (β = 0.570, p < 0.001), and under an ordinal estimator.
3. A five-factor measurement representation showed close fit and was reproduced in a held-out subsample, supporting the separation of LOU from related academic experiences and learning tendencies.
4. An observationally equivalent reverse specification produced the same global fit, so the data cannot identify temporal direction — the paper states this plainly rather than implying a causal path.
5. Exploratory extensions showed the six-path specification did not exhaust the covariance structure, particularly for one academic-experience factor, so the model is a constrained summary rather than a complete account.
6. Data came from 1,004 students at three public universities, collected by cross-sectional survey and analyzed with covariance-based structural equation modeling plus a battery of sensitivity analyses.

How the study was built

The sample is 1,004 students at three public universities, surveyed cross-sectionally. The construct at the center is deliberately narrow: perceived teacher-referenced normative tension, the felt difficulty of locating one's own GenAI practices against what teachers seem to expect when acceptable-use rules are unsettled. The outcome is self-reported learning-oriented GenAI use, kept separate from adjacent academic experiences and learning tendencies so it can be tested as its own factor.

The analysis is a covariance-based structural equation model with a theory-constrained six-path structure, preceded by a five-factor measurement model. The authors then run an unusually thorough robustness battery: internal split-sample measurement checks, alternative measurement models, covariate-adjusted models, an ordinal-estimator sensitivity analysis, institution-level sensitivity analyses, a reverse focal specification, and exploratory structural extensions.

Why the tension finding is not the obvious one

The intuitive expectation is that ambiguity about what teachers permit suppresses learning-oriented use — students avoid the tool when they cannot tell whether it is allowed. The paper finds the opposite association. Students who report more normative tension also report more learning-oriented use, and the relationship survives adjustment for how clear the GenAI guidance is.

The authors' reading is that tension and productive use can coexist, because locating one's practice relative to an unsettled norm is itself an active, interpretive act rather than a barrier. That is a useful reframing for instructors: a student wrestling with "is this allowed, and for what?" may be more engaged than one who has settled the question either way.

What the cross-sectional design cannot settle

The paper is careful here, and the caution deserves to travel with the finding. Because the design is cross-sectional and same-source, a reverse direction — learning-oriented use producing more felt tension, as students engage more and notice the ambiguity more — fits the data just as well; the reverse specification produced the same global fit. Shared unmeasured characteristics, common-method inflation, social desirability and legitimacy-oriented self-presentation are all listed as live alternatives.

The practical consequence is that the association should not be reported as "ambiguity drives learning." It is a measurement contribution with a directional question left open: a validated way to separate learning-oriented use from neighboring constructs, and a documented association that later longitudinal or experimental work can test for direction.

What this means for practice

• Instructors. Do not assume that unclear AI rules simply reduce use. This sample associates unsettled norms with more learning-oriented use, so the constructive move is to make the reasoning behind a policy visible rather than to issue a bare prohibition and expect compliance.
• Course and program designers. Separate the question "did students use GenAI?" from "did they orient it toward learning?". The study's five-factor structure treats learning-oriented use as its own thing, and a course evaluation that collapses the two will miss the distinction the paper is built on.
• Administrators. Treat ambiguity as something to manage openly rather than to eliminate by decree. If tension and learning-oriented use can rise together, a clarity campaign that removes all discretion may not produce better use.
• Researchers. The directional question is open and cheap to attack: a longitudinal design or a two-wave panel would test whether tension precedes use, use precedes tension, or both are driven by a shared trait.

Limitations

• The design is cross-sectional and single-source, so temporal direction is unidentified; the authors report that a reverse specification fits equally well and name common-method inflation and social desirability as live threats.
• Both focal constructs are self-reported, and the outcome is self-reported learning-oriented use, not observed learning behavior or a learning gain.
• The sample is 1,004 students at three public universities in one national system, so institutional and cultural generalization is untested.
• Exploratory extensions found the constrained six-path model did not fully reproduce the covariance structure, so the reported model is a deliberate simplification rather than a complete account.

Citation

Yang, Z., Yang, Y., & Liu, W. (2026). Self-reported learning-oriented GenAI use under perceived teacher-referenced normative tension in Chinese higher education. Frontiers in Psychology, 17.