Guarded adoption of generative AI in higher education: high-achieving students, successful-student identity, and epistemic agency in a single-university mixed-methods survey

Synthesis: An institution-wide mixed-methods survey at one Australian university (484 responses, 469 with valid GPA-band data) examined whether successful-student identity and self-reported achievement were associated with different patterns of generative AI engagement. Higher-achieving students reported lower active AI engagement, lower positive affect, lower perceived learning impact and lower AI-related disengagement, alongside slightly higher negative affect. Item-level analyses showed less endorsement of AI-supported autonomy and effective learning, and greater agreement that reliance on AI hinders Critical Thinking and independent problem solving. The open-ended responses described selective use for clarification, summarization and workflow support, with outputs checked and kept subordinate to the student's own judgment. The authors name this pattern guarded adoption: selective, bounded and verification-intensive use in defense of epistemic control.

Key Findings

1. High-achieving students reported lower active AI engagement, lower positive affect toward AI and lower perceived positive learning impact, with a small positive association with negative affect.
2. The strongest Spearman correlations with self-reported GPA band were perceived learning impact (rho = -0.395), positive affect (rho = -0.359) and active AI engagement (rho = -0.357) — statistically significant but only small to moderate in magnitude.
3. High-achieving students also reported lower AI-related disengagement, which rules out a simple technology-avoidance explanation: they were neither enthusiastic heavy users nor disengaged from the technology.
4. At item level, higher achievers were less likely to say AI supported their autonomy, helped them learn more effectively, or formed part of active academic engagement, and more likely to agree that reliance on AI hinders critical thinking and independent problem solving.
5. Open-ended responses showed the workaround rather than the refusal: AI used selectively for clarification, summarization and workflow support, with outputs verified and treated as subordinate to the student's own reasoning.

Guarded adoption as a defended identity

The paper's interpretive contribution is to read the pattern as identity work rather than as a technology-perception variable. For students whose sense of themselves as successful learners rests on producing work through their own effort and judgment, unbounded AI use threatens the identity the achievement depends on. That reading explains the otherwise puzzling combination of low engagement, low disengagement and mild negative affect: the students are not disengaged, they are holding a boundary. The construct also links directly to epistemic agency, since the reported behavior is verification — outputs are checked and then subordinated rather than accepted or rejected wholesale.

The institutional implication is uncomfortable for both sides of the policy debate. Campaigns promoting AI use target exactly the group already exercising discriminating judgment, while AI Detection or prohibition regimes push against students whose self-reported behavior is closest to the pedagogic ideal. The authors note the interpretive limits of their design: a single institution, self-reported GPA bands rather than transcript data, modest internal consistency on some descriptive indices, and correlations that explain only part of the variance in engagement.

What this means for practice

• Instructors. Teach verification and bounded use explicitly rather than assuming it develops on its own: the high-achieving respondents already checked outputs for accuracy and kept them subordinate to their own judgment, and that practice is what makes the difference visible and assessable.
• Instructors. Ask students directly about the boundaries they hold instead of reading low reported engagement as disengagement — higher achievers reported both lower active AI engagement and lower AI-related disengagement, which is a defended boundary rather than technology avoidance.
• Administrators. Retarget policy away from the promotion-versus-detection axis: campaigns that promote AI use reach students already exercising discriminating judgment (active AI engagement rho = −0.357 with GPA band), while detection or prohibition regimes press students whose self-reported behavior is closest to the pedagogic ideal.
• Researchers. Report item-level results rather than leaning on composite indices when internal consistency is weak, as this study did, and treat the identity explanation as a hypothesis that needs designs linking reported achievement to observed AI use.

Limitations

• Achievement is a self-reported GPA band (coded 1 = less than 3 to 9 = more than 6.5) rather than transcript data: the survey yielded 484 responses, of which only 469 carried a valid GPA band, so the key predictor holds both recall and social-desirability error.
• Every outcome measure is self-reported, and internal consistency was modest to weak on the grouped indices (active AI engagement alpha = 0.57, perceived learning impact alpha = 0.52, positive affect alpha = 0.57, AI-related disengagement alpha = 0.27, negative affect alpha = 0.21), so the authors treat those scores as descriptive summaries only.
• The associations were significant but small to moderate — rho = −0.395 for perceived learning impact, −0.359 for positive affect and −0.357 for active AI engagement — so achievement explains only part of the variation in AI engagement.
• The data come from one Australian university, recruited through central research channels over a two-month period at the beginning of 2025, so the successful-student identity mechanism remains a hypothesis for wider testing rather than an established cross-context effect.

Citation

Zagami, J. (2026). Guarded adoption of generative AI in higher education: high-achieving students, successful-student identity, and epistemic agency in a single-university mixed-methods survey. International Journal of Educational Technology in Higher Education, 23(1), 47.