Future of Educational Science: Generative AI Use, Trust, and Cognitive Load as Predictors of Self-Perceived Academic Performance in Higher Education

Summary

This study examines how generative AI use and trust in AI affect academic performance among higher education students through cognitive load. Using partial least squares structural equation modeling (PLS-SEM) on data from 390 students reveals that both factors significantly increase cognitive load, which positively predicts performance and mediates their effects.

Abstract

This study investigates the complex relationships between generative AI usage, trust in AI, cognitive load, and academic performance among higher education students. Grounded in cognitive load theory and trust literature, the research examines how students' engagement with generative AI tools and their trust in these systems influence academic outcomes, with cognitive load as the mediating mechanism. A quantitative cross-sectional research design was employed with a stratified random sample of 390 higher education students. Data were collected using validated scales measuring generative AI usage trust in AI (human-like and functionality dimensions), cognitive load (intrinsic load, extraneous load, and self-perceived learning), and academic performance. Partial least squares structural equation modeling (PLS-SEM) with bootstrapping procedures (5,000 resamples) was used to test the hypothesized direct and mediating relationships. The results revealed that generative AI usage (β = 0.34, p < 0.001) and trust in AI (β = 0.28, p < 0.001) were significantly positively associated with cognitive load. Cognitive load was also positively associated with academic performance (β = 0.52, p < 0.001). Furthermore, cognitive load significantly mediates the relationships between generative AI usage and academic performance (β = 0.18, p < 0.001) and between trust in AI and academic performance (β = 0.15, p < 0.001). The model explained 45% of the variance in academic performance, and PLS Predict confirmed high predictive power. The findings extend cognitive load theory to AI-enhanced learning contexts and suggest that cognitive load may function as an important explanatory pathway linking AI-related factors to academic outcomes. Practically, the study underscores the importance of fostering AI literacy, designing learning environments that optimize cognitive load, and prioritizing functional reliability in AI tool development. This study provides empirical evidence that cognitive load serves as a significant mediating mechanism through which generative AI usage and trust in AI translate into academic performance.

Introduction

The rapid advancement of generative artificial intelligence (GenAI) has precipitated a paradigm shift in higher education, fundamentally transforming how students’ access, process, and construct knowledge¹. Tools such as ChatGPT, Gemini, and Claude have become increasingly embedded in academic environments, with recent estimates suggesting that over 30% of university students regularly employ these technologies for coursework-related activities². This technological revolution presents both unprecedented opportunities and significant challenges for educators, administrators, and researchers seeking to understand how....

Protocol

The table of materials summarizes the instruments, software, and scales used in the study, including adapted and validated scales for generative AI use, trust in AI, cognitive load, self-perceived academic performance, sampling and survey procedures, and data analysis tools (SPSS and SmartPLS). Moreover, ethical approval was not applicable for this study as it did not involve any experimental manipulation, medical procedures, or interventions that could pose physical or psychological risks to participants. The research used a non-invasive, self-administered online questionnaire to collect data on students' perceptions, experiences, and behaviors r....

Results

Descriptive statistics

Descriptive statistics were computed to summarize the central tendencies and dispersion of the key variables in this study, including generative AI usage, human-like trust, functionality trust, intrinsic load, extraneous load, self-perceived learning, and academic performance. The descriptive analysis was conducted using SPSS version 26, and the results are presented in Table 1 below.

Discussion

The first major finding of this study revealed that generative AI usage is positively associated with cognitive load (β = 0.34, p < 0.001), a result that is consistent with Hypothesis 1. This finding suggests that students who report more frequent use of generative AI tools such as ChatGPT also report higher cognitive load during their learning activities. This positive association between generative AI usage and cognitive load may be better understood from a human‑AI interaction perspective rathe.......

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Blindfolding Procedure	SmartPLS 4 (built-in)	N/A	Computes Q² for predictive relevance of the structural model.
Bootstrapping Procedure (5,000 resamples)	SmartPLS 4 (built-in)	N/A	Used to assess significance of path coefficients (β) and indirect effects.
Cognitive Load Scale – Extraneous Load (3-item)	Hadie & Yusoff (2021)	N/A (validated scale)	Cognitive burden from instructional design and information presentation.
Cognitive Load Scale – Intrinsic Load (3-item)	Hadie & Yusoff (2021)	N/A (validated scale)	Perceived complexity/difficulty of learning materials/tasks.
Cognitive Load Scale – Self-Perceived Learning (4-item)	Hadie & Yusoff (2021)	N/A (validated scale)	Proxy for germane cognitive processing; captures mental effort for understanding and schema construction.
Fornell-Larcker Criterion	SmartPLS 4 (built-in)	N/A	Assesses discriminant validity of constructs.
Generative AI Usage Scale (8-item)	Abbas et al. (2023)	N/A (adapted scale)	Measures frequency, purpose, and perceived efficacy of Generative AI (ChatGPT) use for academic tasks. 5-point Likert scale.
HTMT Ratio of Correlations	SmartPLS 4 (built-in)	N/A	Heterotrait-Monotrait ratio; additional discriminant validity check.
Informed Consent Form	Study team (institutional guidelines)	N/A	Detailed description of study purpose, voluntary participation, anonymity, confidentiality, and right to withdraw. Implied consent via survey completion.
Online Questionnaire Platform	Not specified (generic)	N/A	Structured, self-administered web-based questionnaire; all items on 5-point Likert scale (1 = Strongly Disagree to 5 = Strongly Agree).
Sampling Frame	University Registrar’s Office	N/A (institutional record)	Complete list of currently enrolled students across all faculties used for stratified random sampling.
Self-Perceived Academic Performance Scale (4-item)	Adapted from educational research precedent	N/A (adapted scale)	Subjective assessment of academic efficacy and achievement (e.g., confidence in coursework, efficient task management).
SmartPLS 4 Software	SmartPLS GmbH	Version 4	Variance-based structural equation modeling software; used for PLS-SEM analysis, bootstrapping, blindfolding.
SPSS version 26	IBM	N/A	Used for preliminary data screening, descriptive statistics, and any pre-PLS-SEM data management (if applicable).
Stratified Random Sampling Procedure	Study team (manual)	N/A	Strata based on academic discipline (4 categories) and year of study (5 levels: 1st-4th year undergrad + postgrad). Random number generator used within strata.
Trust in AI Scale – Functionality Trust (5-item)	Adapted from technology trust literature	N/A (adapted scale)	Measures perceived reliability, competence, and dependability of AI for academic tasks.
Trust in AI Scale – Human-like Trust (6-item)	Adapted from interpersonal trust (Rempel et al.) and technology trust literature	N/A (adapted scale)	Assesses benevolence, integrity, and predictability attributed to AI.