Chinese university students who reported more difficulty judging whether their generative AI use fit teacher-referenced expectations also reported more learning-oriented AI use, a study found. In the researchers’ main six-path model, the positive association had a standardized coefficient of 0.568, and all six specified associations were statistically significant at p < 0.001.
That result does not show that uncertainty causes students to use AI for learning. The survey was cross-sectional, and all focal measures came from same-source self-reports, so common-method inflation, socially desirable responding and generally positive response tendencies remain possible explanations for part of the pattern.
How the relationship was measured
The data came from a quantitative cross-sectional survey of students at three public universities in mainland China, collected during the 2025-2026 academic year. Researchers collected 1,020 questionnaires, removed 16 attention-check failures and analyzed a final sample of 1,004. Recruitment used purposive institutional selection and convenience-based methods. All final respondents reported prior academic or learning-related GenAI use.
Structural equation modelling, the study’s main analysis, estimates several linked relationships at once. The primary model used maximum-likelihood estimation; the researchers also ran a DWLS sensitivity analysis for the ordinal survey items. The six structural paths were specified in advance, but the study was not preregistered, and additional direct paths and robustness and split-sample checks were treated as post hoc or non-confirmatory.
A pattern that held up, with caveats
One part of the analysis tested whether five survey constructs, including PNT and LOU, could be represented as separate factors. The five-factor model showed unusually close fit on standard model checks: CFI and TLI were both 1.000, RMSEA was 0.000 and SRMR was 0.020. That close fit was reproduced in a held-out validation subsample, although the authors cautioned that it was not definitive evidence of validity and needed external replication.
The measurement picture had caveats. A two-factor version of PNT fit better statistically, but its two factors correlated 0.949, so the researchers retained PNT as a single factor. A proposed split of the related ACS measure into autonomy and competence produced an inadmissible correlation of 1.029, so ACS was retained as a combined factor.
The association survived several checks
The six-path model explained 32.2% of the variation in learning-oriented GenAI use and 25.4%, 32.2% and 23.9% in the study’s three related academic constructs. The analysis therefore did not treat the model as an exhaustive account of the relationships among the measures.
The researchers then adjusted for perceived clarity of GenAI guidance. The PNT-LOU coefficient remained positive but fell to 0.478 from 0.568; guidance clarity was also associated with LOU and ACS, and the share of LOU variation explained rose from 32.2% to 36.9%. The attenuation suggested overlap between perceived tension and guidance clarity, not a causal effect.
Other checks produced similar results. After adjustment for gender, discipline, year level and institution, the PNT-LOU coefficient was 0.570, compared with 0.568 without those covariates. Institution-specific coefficients were 0.552, 0.606 and 0.555, and the DWLS sensitivity analysis kept all six specified associations statistically significant.
What the evidence does not establish
Direction remained unresolved. When the researchers reversed the focal specification so that LOU pointed to PNT, the model had the same global fit as the hypothesized direction. The reverse model reported CFI 0.988, TLI 0.986, RMSEA 0.022 and SRMR 0.055, and the cross-sectional data could not identify whether PNT preceded LOU or LOU preceded PNT.
The six-path structure also did not capture all observed covariance. In post hoc exploratory tests, adding a direct PNT path to ACS improved fit, while adding direct paths to SRL and CT produced a smaller improvement. The added direct associations with SRL and CT were 0.092 and 0.103, respectively. The authors treated these extensions as evidence that the primary structure was incomplete, not as a new confirmed causal pathway.
Several limitations narrow the takeaway. The focal measures were same-source self-reports, leaving common-method inflation, socially desirable responding and general positive response tendencies unresolved. The student sample was recruited through convenience methods at three public universities, so the evidence describes between-student covariance within those settings rather than an established pattern across higher education.
The study therefore supports a narrower conclusion: PNT and LOU are associated in these student self-reports, but the data do not establish causation or temporal order. The authors declared that they received no financial support for the work or its publication.
Paper data and sources
Original title: Self-reported learning-oriented GenAI use under perceived teacher-referenced normative tension in Chinese higher education
Authors: Yang Z, Yang Y, Liu W
Journal/Repository: Frontiers in psychology
Status: Peer-reviewed
First online: 2026-10-19
DOI: Not available
Original paper · Full text