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

The integration of generative artificial intelligence into higher education has shifted from a novel technological experiment into an institutional governance challenge. While university students increasingly incorporate tools such as large language models into academic writing, brainstorming, and information retrieval, educators and administrators continue to negotiate the boundaries of acceptable use. A recent empirical study conducted across three public universities in mainland China examines how students navigate this pedagogical gray area. Specifically, the research investigates whether perceived teacher-referenced GenAI normative tension—defined as students’ reported difficulty in locating their own AI practices within stable instructor standards—correlates with learning-oriented GenAI use, rather than leading to disengagement or academic withdrawal.

Background Context and Research Framework

The rapid proliferation of generative artificial intelligence tools has exposed a significant regulatory vacuum across global academia. Prior research has predominantly utilized technology acceptance models to evaluate adoption rates, user intentions, and academic integrity concerns. However, these traditional frameworks leave a critical question unanswered: how do students orient their behavior when institutional and instructor expectations remain fragmented, inconsistent, or altogether unarticulated?

In Chinese higher education, where instructional authority traditionally plays a vital role in shaping academic conduct, students face a distinct balancing act. They experience rising academic pressures alongside mixed signals from faculty members regarding what constitutes permissible assistance versus unacceptable task substitution. To address this phenomenon, researchers conceptualized Perceived Normative Tension as a student-reported metric capturing perceived disagreement, alignment uncertainty, unclear expectations, and cross-teacher inconsistencies. Rather than assuming that normative ambiguity automatically results in student alienation or academic dishonesty, the study hypothesized that students who continue utilizing AI under unsettled expectations consciously pivot toward cognitively active, learning-oriented practices to maintain academic defensibility.

Methodology and Sample Demographics

The cross-sectional study gathered data during the 2025–2026 academic year utilizing purposive institutional selection and convenience-based student recruitment across three public universities in mainland China. Out of 1,020 initially administered questionnaires through the online platform Wenjuanxing, 1,004 valid responses were retained after screening out failed attention-check items.

The final sample demographic distribution reflected a balanced participation rate: 491 male students (48.9%) and 513 female students (51.1%). In terms of academic specialization, 361 respondents (36.0%) were enrolled in science, technology, engineering, and mathematics (STEM) disciplines, while 643 respondents (64.0%) pursued humanities and social sciences. The cohort comprised 490 sophomores (48.8%) and 514 juniors (51.2%), distributed across Institution A (n = 331), Institution B (n = 306), and Institution C (n = 367). Additionally, an unmatched complementary survey of 148 faculty members from the same institutions provided contextual background regarding institutional risk perceptions and ethical boundary strictness.

Covariance-based structural equation modeling executed via R software evaluated a five-factor measurement model alongside a theory-constrained six-path associational model. Robustness checks included ordinal-estimator sensitivity analyses using diagonally weighted least squares, guidance-clarity adjustments, demographic covariate controls, and split-sample internal replication.

Key Empirical Findings

The evaluation of the measurement model demonstrated close approximate fit indices, confirming that learning-oriented GenAI use, normative tension, combined autonomy- and competence-related academic experience, self-regulated learning, and critical-evaluation tendencies remained empirically distinguishable.

The structural model yielded several notable statistical outcomes:

• Perceived teacher-referenced normative tension demonstrated a positive and statistically significant association with self-reported learning-oriented GenAI use ($beta = 0.568, p < 0.001$), accounting for 32.2% of the variance in learning-oriented use.
• Learning-oriented GenAI use showed positive associations with combined autonomy- and competence-related academic experience ($beta = 0.504, p < 0.001$), self-regulated learning ($beta = 0.361, p < 0.001$), and critical-evaluation tendencies ($beta = 0.331, p < 0.001$).
• Combined autonomy- and competence-related experience further correlated positively with self-regulated learning ($beta = 0.292, p < 0.001$) and critical-evaluation tendencies ($beta = 0.230, p < 0.001$).

Crucially, robustness checks controlling for GenAI guidance clarity confirmed that the focal association between normative tension and learning-oriented use remained positive and significant ($beta = 0.478, p < 0.001$). However, because the study relied on a cross-sectional design and same-source self-reports, researchers noted that temporal precedence could not be established. An observationally equivalent reverse specification produced identical global fit statistics, indicating that the data cannot definitively prove whether normative tension drives learning-oriented usage or whether reflective users are simply more attuned to instructor-level normative inconsistencies.

Implications for Academic Governance and Pedagogy

The empirical insights highlight vital considerations for university administrators and classroom instructors navigating the artificial intelligence transition. The coexistence of normative tension and learning-oriented self-reports underscores that broad institutional policies alone are insufficient. When overarching ethical guidelines are not translated into explicit, course-specific expectations, students are left to independently reconcile their AI usage with ambiguous instructor standards.

Faculty feedback captured during the research revealed that instructors hold significantly stricter ethical boundaries regarding artificial intelligence than students do, pointing toward a persistent communication gap. Experts suggest that universities should move beyond prohibitive policies toward clear, granular frameworks. Academic departments are advised to define permitted, restricted, and prohibited GenAI functions for specific assignments—differentiating between acceptable concept clarification or brainstorming and unacceptable answer substitution.

Furthermore, educators are encouraged to implement process-oriented assessments that require students to document prompts, critically evaluate generated outputs, and articulate how independent reasoning was integrated into final submissions. Such pedagogical transparency can minimize avoidable normative ambiguity while reinforcing academic integrity.

Broader Impact and Future Outlook

As generative artificial intelligence continues to transform global higher education, this study provides a foundational lens for understanding the psychological and behavioral dynamics of student AI adoption under regulatory uncertainty. While the findings validate that students facing normative tension often lean toward cognitively engaged practices, the authors caution against interpreting normative ambiguity as an advantageous pedagogical tool.

Future academic inquiry must move beyond cross-sectional self-reports by incorporating longitudinal tracking, multi-source faculty-student matched data, and objective behavioral traces. Establishing whether these patterns hold across diverse international education systems will be essential as universities worldwide strive to harmonize technological innovation with rigorous academic standards.