Why Trust, Not Awareness, Drives How Indian Students Engage With Generative AI

Generative artificial intelligence has moved from a specialist technology to a routine part of student academic life, and a new study from Indian higher education suggests that what really matters is not how much students know about AI’s flaws, but how useful they believe it is. Researchers at CHRIST (Deemed to be University) in Bengaluru surveyed 207 undergraduate and postgraduate students who had prior experience using generative AI for academic purposes, and used structural equation modelling to trace how epistemic awareness, perceived usefulness, trust, AI use, and learning engagement relate to one another. The findings, published in Discover Education, challenge the intuitive assumption that knowing about AI’s limitations makes students more cautious or less trusting of the technology.

The study drew on three theoretical strands: andragogical learning theory, which frames adult learners as self-directed and goal-oriented; technology acceptance research, which positions perceived usefulness as a central driver of adoption; and trust-in-automation scholarship, which treats trust as an affective-cognitive judgement about a system’s dependability. Within this framework, the researchers operationalised epistemic awareness narrowly as students’ awareness that generative AI can produce inaccurate information, lacks human-like understanding, and generates responses through pattern-based prediction rather than factual reasoning. This is distinct from broader AI literacy, which encompasses technical knowledge, prompt engineering, ethics, and operational competence.

The researchers tested six hypotheses. The first five proposed linear structural relationships: that epistemic awareness positively predicts trust, that perceived usefulness positively predicts trust, that trust positively predicts frequency of AI use, that AI use positively predicts learning engagement, and that AI use mediates the relationship between perceived usefulness and learning engagement via trust. The sixth hypothesis proposed a curvilinear over-trust effect, suggesting that extremely high trust would be associated with reduced learning engagement. Because this final hypothesis concerned a within-construct nonlinear effect rather than a linear path between latent constructs, it was evaluated through a supplementary quadratic regression conducted outside the core structural model.

The results were striking. Perceived usefulness emerged as the primary antecedent of trust, with a standardised path coefficient of 0.856 (p < 0.001). Trust, in turn, strongly predicted AI use (β = 0.860, p < 0.001), and AI use strongly predicted learning engagement (β = 0.700, p < 0.001). The bootstrapped indirect pathway from perceived usefulness through trust and AI use to learning engagement was significant (β = 0.515, p < 0.001), supporting the mediation hypothesis. The structural model explained 78.1 percent of the variance in trust, 73.9 percent in AI use, and 49.0 percent in learning engagement.

But the hypothesis that epistemic awareness would positively predict trust was not supported. The coefficient was negative in direction (β = −0.098) but statistically non-significant (p = 0.172), meaning the data provided no evidence that greater awareness of AI limitations increases or decreases trust. The researchers are careful to note that this non-significant finding should not be interpreted as proof that epistemic awareness necessarily produces cautious trust; rather, it indicates that the hypothesised positive association was simply not detectable in this sample. Alternative factors such as perceived risk, AI scepticism, prior experience, or other unmeasured characteristics may contribute to how awareness relates to trust, and future research using behavioural, longitudinal, or process-based measures could examine this more directly.

The supplementary nonlinear analysis also failed to support the over-trust hypothesis. Adding a quadratic trust term did not improve model fit relative to the linear model (ΔF = 0.34, p = 0.561), and the quadratic coefficient was not statistically significant (β = 0.020, p = 0.561). Visual inspection of the trust-engagement plot suggested a predominantly positive association rather than the inverted-U pattern the researchers had predicted. Interestingly, once AI use and perceived usefulness were entered as covariates in the supplementary model, the linear effect of trust on engagement became negative and marginally significant (β = −0.193, p = 0.050). The researchers flag this as a pattern worth future investigation, but caution that it does not support the over-trust hypothesis as originally framed, which predicted curvature rather than a negative linear effect.

The measurement model revealed important caveats. Epistemic awareness showed marginal convergent validity, with an average variance extracted of 0.453, falling just below the conventional 0.50 threshold. More significantly, the cluster of trust, perceived usefulness, and AI use failed the Fornell-Larcker discriminant validity criterion, and heterotrait-monotrait ratios between these constructs approached the conventional threshold. This means these three constructs are not fully empirically distinguishable in this sample, and the strength of the structural paths linking them may be partly inflated by shared variance among closely worded items. The researchers interpret the usefulness-trust-use pathway as evidence of a coherent instrumental-adoption process rather than proof of three fully independent psychological mechanisms.

Common method bias was assessed using Harman’s single-factor test, in which all retained indicators were loaded onto a single latent factor. This single-factor model showed poor fit (CFI = 0.718, RMSEA = 0.121), indicating that a single underlying factor could not account for the covariance among items. Multicollinearity among the predictors of trust was assessed using variance inflation factors, both returning values of 1.04, well below the conventional threshold of 5. The confirmatory factor analysis showed acceptable fit (CFI = 0.898, RMSEA = 0.075, SRMR = 0.057), while the structural model fit was more cautious (CFI = 0.880, RMSEA = 0.080), which the researchers interpret as acceptable rather than strong support for the hypothesised structure.

The study’s theoretical contribution lies in positioning generative AI use as a process of cognitive regulation rather than simply a technology-adoption behaviour. The findings suggest that engagement emerges not simply from access to AI, but from a process in which students evaluate usefulness, calibrate trust, and integrate AI into their learning behaviour. Epistemic awareness, rather than operating as a direct positive predictor of trust or adoption, may instead function as an epistemic orientation or critical awareness framework that shapes how students evaluate AI-generated information. This challenges simplistic assumptions that greater awareness of AI limitations necessarily increases acceptance or reliance, and supports emerging arguments that AI literacy should be understood as involving critical evaluation, ethical awareness, and metacognitive judgment rather than operational skill alone.

Practically, the findings suggest that educators and institutions may benefit from designing AI-supported learning activities around clearly defined academic purposes and demonstrable learning value, rather than treating AI use as an end in itself. Because the study was cross-sectional, these findings should be interpreted as associations rather than evidence that increasing perceived usefulness will necessarily cause greater trust or engagement. The researchers recommend that epistemic awareness initiatives prioritise critical awareness rather than merely operational skill, pairing instruction in prompt engineering with explicit training in critically evaluating and verifying AI-generated outputs. They also caution that the absence of a significant nonlinear effect should not be interpreted as evidence that excessive reliance on generative AI is harmless; rather, it indicates that the current sample did not exhibit a measurable decline in engagement at higher trust levels, and longitudinal, experimental, or task-based research may be better positioned to detect such effects under different educational conditions.

Subject of Research: Psychological and behavioural factors linking epistemic awareness, perceived usefulness, trust, and learning engagement in generative AI use among Indian higher education students

Article Title: Understanding learning engagement in Indian higher education through epistemic awareness, perceived usefulness, and trust in generative AI

Article References: Singh, M. I., Pradhan, S., Naveen, S., & Vinod, L. (2026). Understanding learning engagement in Indian higher education through epistemic awareness, perceived usefulness, and trust in generative AI. Discover Education, 5(1), Article 806. https://doi.org/10.1007/s44217-026-02090-9

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02090-9

Keywords: generative AI, higher education, learning engagement, epistemic awareness, trust in AI, perceived usefulness, structural equation modelling, Indian students, technology acceptance, AI literacy, over-trust hypothesis, Discover Education