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 AI integration influences learning processes and outcomes³.

While generative AI offers the potential for personalized tutoring, instant feedback, and enhanced productivity, students do not merely consume AI-generated information passively⁴. Rather, AI-enhanced learning involves continuous interactional processes in which students formulate prompts, evaluate outputs, compare AI-generated responses with external sources, calibrate their reliance on AI systems, and integrate machine-generated content into academic tasks⁵. These interactional demands introduce new cognitive dynamics that extend beyond traditional technology usage models and may simultaneously reduce certain cognitive burdens while creating new forms of cognitive effort and overload. Central to these concerns is the question of how AI usage affects the cognitive processes underlying learning, a question that remains largely unexplored in the empirical literature. Cognitive load theory (CLT) provides a robust theoretical framework for examining these questions. CLT posits that learning is constrained by the limited capacity of working memory and that effective instructional design must manage three types of cognitive load: intrinsic load arising from content complexity, extraneous load imposed by poor instructional design, and germane load directed toward schema construction and automation⁶. In AI-enhanced learning environments, these distinctions become particularly salient⁷. Students interacting with generative AI tools must simultaneously process course content, evaluate AI-generated information for accuracy and relevance, and integrate diverse sources of knowledge processes that may significantly increase cognitive demands⁸.

Importantly, cognitive load theory conceptualizes cognitive load as a multidimensional construct consisting of intrinsic, extraneous, and germane load, each of which may affect learning differently⁹. Intrinsic load arises from the inherent complexity of learning material, extraneous load reflects unnecessary cognitive burden generated by ineffective instructional design or irrelevant processing demands, and germane load refers to productive cognitive engagement associated with schema construction and meaningful learning processes. Consequently, cognitive load should not be interpreted as uniformly beneficial or harmful. While intrinsic and germane load may support deeper understanding under appropriate conditions, excessive extraneous load can interfere with learning by consuming limited cognitive resources.

Beyond usage patterns, students' trust in AI systems emerges as a critical psychological factor shaping human-AI interaction¹⁰. Drawing upon foundational trust frameworks, trust in AI can be conceptualized along two distinct dimensions: human-like trust, encompassing perceptions of benevolence, integrity, and predictability; and functionality trust, reflecting confidence in the reliability, competence, and dependability of the technology itself¹¹. Prior research has established trust as a determinant of technology adoption and continued use¹², yet its cognitive consequences remain largely unexplored. The relationship between trust in AI and cognitive load is theoretically complex and should not be interpreted as uniformly positive or linear¹³. On one hand, appropriate trust in AI may encourage students to engage more deeply with AI-generated content, invest cognitive effort in evaluating responses, and integrate machine-generated information into learning tasks¹⁴. The relationship between AI-related factors and academic performance represents the ultimate outcome of interest for educators and institutions¹⁵. Academic performance has been operationalized in various ways across educational research, with some studies utilizing objective measures such as grade point averages¹⁶, while others rely on students' self-perceived academic efficacy¹⁷. Self-perceived measures capture students' metacognitive awareness of their learning capabilities and task performance, which research has shown to be strongly associated with actual academic achievement¹⁸. In AI-enhanced learning contexts, understanding how students perceive their own learning becomes particularly important, as these perceptions may guide subsequent engagement with AI tools and learning strategies¹⁹.

Despite the growing prevalence of generative AI in higher education, empirical research examining its cognitive and academic consequences remains scarce²⁰. Existing studies have primarily focused on adoption patterns, ethical concerns, or disciplinary applications²¹, with limited attention to the psychological mechanisms linking AI usage to learning outcomes²². Furthermore, while trust has been extensively studied in e-commerce and human-computer interaction²³, its role in educational AI contexts remains theoretically underdeveloped and empirically untested. The present study addresses these gaps by examining the associations among generative AI usage, trust in AI, cognitive load, and academic performance, with cognitive load modeled as a potential mediating pathway. Grounded in cognitive load theory²⁴, and trust literature, this study proposes and tests a mediation model examining five hypotheses: (H1) generative AI usage positively influences cognitive load; (H2) trust in AI positively influences cognitive load; (H3) cognitive load positively influences academic performance; (H4) cognitive load mediates the relationship between generative AI usage and academic performance; and (H5) cognitive load mediates the relationship between trust in AI and academic performance.

The integration of generative artificial intelligence (GenAI) into higher education represents one of the most significant technological shifts in contemporary pedagogical practice²⁵. Tools such as ChatGPT, developed by OpenAI, have demonstrated remarkable capabilities in generating human-like text, solving complex problems, and providing personalized educational support²⁶. Recent empirical investigations have documented widespread adoption among university students. Students utilize GenAI for diverse academic purposes, including brainstorming, content summarization, proofreading, and concept explanation²⁷. This proliferation has prompted scholars to examine both the affordances and challenges associated with GenAI in educational contexts. Proponents argue that GenAI tools can enhance learning by providing immediate feedback, scaffolding complex tasks, and offering multiple perspectives on disciplinary content²⁸. The students using ChatGPT for programming tasks showed improved code quality and reduced completion time, suggesting potential cognitive offloading benefits. Similarly, GenAI can personalize learning experiences by adapting content to individual student needs and learning paces²⁹. However, critics have raised concerns about academic integrity, the potential for plagiarism, and the erosion of critical thinking skills when students over-rely on AI-generated content^30,31. These competing perspectives underscore the need for rigorous empirical investigation into how GenAI usage affects fundamental learning processes.

Cognitive load theory (CLT) provides a comprehensive framework for understanding the cognitive demands imposed by learning tasks and instructional designs. CLT is grounded in the established architecture of human cognition, which comprises an effectively unlimited long-term memory and a severely limited working memory in both capacity and duration³². According to CLT, learning occurs when information is successfully processed in working memory and subsequently stored as schemas in long-term memory. The theory distinguishes among three types of cognitive load: intrinsic load, which is determined by the inherent complexity of the learning material and the learner's existing expertise; extraneous load, which is imposed by poorly designed instructional materials and activities that do not directly contribute to learning; and germane load, which refers to the cognitive resources devoted to schema construction and automation³³. Research has consistently demonstrated that effective instructional design must manage these three load types to optimize learning outcomes³⁴. A high extraneous load diverts cognitive resources away from learning processes, impairing performance, whereas appropriate intrinsic and germane loads facilitate deep understanding and knowledge transfer³⁵. a comprehensive instrument for measuring these load dimensions in educational settings, enabling empirical investigation of cognitive load across diverse learning contexts. More recently, scholars have extended CLT to technology-enhanced learning environments, examining how digital tools influence cognitive processing³⁶. The relationship between generative AI usage and cognitive load is theoretically complex and empirically underexplored³⁷. On one hand, GenAI tools may reduce extraneous load by streamlining information search, synthesizing diverse sources, and presenting information in accessible formats³⁸. Students no longer need to navigate multiple databases or interpret complex texts independently, potentially freeing cognitive resources for deeper processing³⁹.

Trust is a fundamental psychological mechanism that shapes human-technology interaction. In the context of information systems, trust is multidimensional, encompassing beliefs about the competence, benevolence, and integrity of technological systems⁴⁰. Applied to AI, trust can be distinguished along two conceptually distinct dimensions: human-like trust, involving attributions of anthropomorphic qualities such as benevolence and predictability; and functionality trust, reflecting confidence in the reliable performance of technical capabilities⁴¹. This distinction is particularly relevant in educational contexts, where students must decide how much to rely on AI-generated content for academic tasks. According to the elaboration likelihood model, individuals process information through either the central or peripheral route depending on motivation and ability. When trust is high, students may engage in central route processing, scrutinizing and elaborating upon AI-generated content, thereby increasing cognitive load directed toward learning⁴². Academic performance constitutes the ultimate criterion variable in educational research, operationalized through both objective measures such as grade point averages and subjective measures capturing students' self-perceived academic efficacy⁴³. Self-perceived measures offer valuable insights into students' metacognitive awareness of their learning capabilities and have demonstrated strong associations with objective performance indicators. A self-perception to assess students perceived academic efficacy, including items related to learning efficiency and task completion⁴⁴. This approach is particularly appropriate in AI-enhanced contexts, where students' confidence in their ability to learn with and through AI tools may influence actual learning outcomes.

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 regarding the use of generative AI in academic settings. Since the study involved no direct interaction beyond completing a survey, no clinical trials, biomedical procedures, or experimental treatments were administered, and participants were not subjected to any experimental conditions that could cause harm or distress. Prior to data collection, all participants were provided with a detailed informed consent form explaining the purpose of the study, the voluntary nature of their participation, and the measures taken to ensure anonymity and confidentiality. Informed consent was obtained implicitly through the completion and submission of the questionnaire, as explicitly stated in the informed consent form. Participants were assured that their responses would be used solely for academic research purposes and that no personally identifiable information would be collected or reported. Furthermore, participants were informed of their right to withdraw from the study at any time before submitting the questionnaire, without consequences. All data were stored securely on password-protected devices accessible only to the research team, and aggregated findings were reported to prevent the identification of individual respondents. These measures ensured that the study was conducted in accordance with the ethical principles of the Declaration of Helsinki and the institutional guidelines for research involving human participants, despite the formal requirement for IRB review not being applicable.

Research design

This study employs a quantitative, cross-sectional research design to investigate the complex relationships between generative AI usage, trust in AI, cognitive load, and academic performance among higher education students. This design was operationalized through a survey strategy, chosen for its efficiency and effectiveness in collecting standardized data from a large sample, thereby enabling robust statistical testing of the hypothesized relationships. The survey method facilitates the systematic measurement of each construct using validated scales, ensuring that the data collected is both reliable and comparable across all respondents. Furthermore, the adoption of a cross-sectional approach, wherein data is gathered at a single point in time, is particularly appropriate for this investigation as it allows for the examination of the current prevalence of generative AI usage and trust among students, while simultaneously assessing the interconnections between these variables and their collective influence on cognitive load and academic performance within the study's target population.

Participants and sampling procedure

The target population for this study comprised students enrolled in higher education institutions who had active experience using generative AI tools, specifically ChatGPT and similar large language models, for academic purposes. To ensure a representative cross-section of this population and enhance the generalizability of findings, a stratified random sampling technique was employed. The sampling frame was obtained from the university registrar's office, comprising a complete list of currently enrolled students across all faculties. Stratification was based on two key demographic variables: academic discipline and year of study. Academic discipline was categorized into four main strata: humanities and social sciences, natural sciences, engineering and technology, and business and economics. Year of study was stratified into four levels: first-year, second-year, third-year, and fourth-year undergraduate students, as well as postgraduate students (Master's and Doctoral). This dual-stratification approach ensured proportional representation from each subgroup within the student body, preventing over-representation or under-representation of any academic field or academic level. From each stratum, participants were randomly selected using a random-number generator, with the number drawn proportional to that stratum's representation in the overall student population. This systematic approach to participant selection ensured that the final sample accurately reflected the diversity of the higher education student population, including academic backgrounds and educational progression.

Regarding participant demographics, the sample of 390 respondents comprised a diverse mix of students across various age groups, typically ranging from 18 to 35 years or older, with a balanced representation of gender identities proportionate to the institutional demographics. The sample included students from all four undergraduate years as well as postgraduate students, ensuring coverage of different stages of academic development and varying levels of exposure to AI tools. Academic disciplines were represented proportionally, with students from humanities, sciences, engineering, and business backgrounds included in the sample. Additionally, basic demographic information, such as prior technology experience, frequency of generative AI use, and native language status, was collected to provide context for the primary variables under investigation. This comprehensive demographic profiling served two purposes: first, it allowed for descriptive characterization of the sample, and second, it enabled researchers to control for potential confounding variables in subsequent analyses. The detailed documentation of the sampling procedure and demographic characteristics ensures that other researchers can replicate this study precisely, thereby fulfilling the scientific requirement of reproducibility and enabling meaningful comparisons across different institutional and cultural contexts. Furthermore, this rigorous sampling approach strengthened the external validity of the findings, allowing for more confident generalization of results to the broader higher education student population. The data for this study were collected using a structured, self-administered online questionnaire. The questionnaire is divided into five sections, designed to measure the study's core constructs using established and validated scales. All items, unless otherwise stated, will be measured on a five-point Likert scale, ranging from 1 ("Strongly Disagree") to 5 ("Strongly Agree").

Instrumentation and measures

Generative AI usage, specifically ChatGPT, was quantified using an 8-item scale adapted from a study⁴⁵. This scale moves beyond simple frequency of use to capture the multifaceted nature of student interaction with AI. The items focus on three key dimensions: (a) the frequency of use for various academic tasks, (b) the specific purposes for which it is employed (e.g., brainstorming, content summarization, proofreading), and (c) the students' perceived efficacy of the tool in enhancing their language-related academic work. Trust in AI was conceptualized as a multidimensional construct, drawing from established frameworks of interpersonal trust⁴⁶, and technology trust. The scale comprises two distinct dimensions. Human-like trust: This 6-item subscale measures the extent to which students attribute human-like qualities to the AI, such as benevolence, integrity, and predictability. For example, items assess whether students believe the AI acts in their best interest or provides unbiased information. Functionality trust: This 5-item subscale assesses trust in the AI's functional performance. It focuses on the reliability, competence, and dependability of the technology itself, capturing the belief that the AI has the functionality required to perform academic tasks effectively.

Cognitive load was measured using a scale adapted from another study⁴⁷, which is grounded in CLT. To capture the triarchic structure of CLT, the scale is divided into three subscales: Intrinsic Load (3 items): this subscale assesses the perceived complexity and difficulty inherent in the learning materials and tasks themselves. A sample item might be, "The topics covered in my coursework are very complex." Extraneous load (3 items): This subscale measures the cognitive burden imposed by the instructional design and the way information is presented. An example is, "The instructions for assignments are often unclear and difficult to follow. Self-perceived learning (4 items): Following prior educational research, this subscale was used as an approximate indicator of germane cognitive processing rather than a direct measure of germane load itself⁴⁸. The items capture students’ perceptions of the mental effort invested in understanding, comprehension, and schema construction processes associated with meaningful learning. However, because self-perceived learning may conceptually overlap with broader perceptions of academic capability and achievement, it should be interpreted cautiously as a partial operational representation of productive cognitive engagement rather than a definitive measure of germane load. It includes items related to understanding, comprehension, and the feeling of having mastered the material. Following established precedents in educational research⁴⁶, academic performance was operationalized using a self-perception scale. A 4-item instrument was employed. This approach is appropriate for capturing students' subjective assessment of their learning efficacy and academic achievement. The scale focuses on perceived academic efficacy and includes statements such as, "I am confident in my ability to complete my coursework successfully," and "I feel I have learned to manage my academic tasks efficiently and effectively."

Data analysis strategy

Partial least squares structural equation modeling (PLS-SEM) was employed to examine the relationships among generative AI usage, trust in AI, cognitive load, and self-perceived academic performance. The selection of PLS-SEM was guided by both methodological and analytical considerations. First, the study adopts a prediction-oriented and exploratory modeling perspective aimed at examining associative pathways among relatively emerging constructs within AI-enhanced learning contexts rather than testing a fully established causal theory. Second, PLS-SEM is considered appropriate for complex models involving multiple latent constructs and mediation pathways, particularly when the primary objective is variance explanation and predictive analysis. In addition, PLS-SEM is robust under conditions of non-normal data distributions and is suitable for social science survey research using moderate sample sizes. Given the exploratory nature of generative AI research in higher education and the study’s emphasis on examining predictive associations among perceptual constructs, PLS-SEM was considered methodologically appropriate for the present analysis.

The collected data will be analyzed using a two-step approach, employing partial least squares structural equation modeling (PLS-SEM) with SmartPLS 4 software. PLS-SEM is a variance-based technique suitable for complex predictive models and does not require the data to be normally distributed. Step 1: Assessment of the measurement model. The first step involves evaluating the reliability and validity of the constructs. Internal consistency reliability will be assessed using Cronbach's alpha and composite reliability, with a threshold of 0.70. Convergent validity will be established by examining the outer loadings of the indicators (should be > 0.70) and the average variance extracted (AVE) for each construct (should be > 0.50). Discriminant validity, which ensures that the constructs are empirically distinct, will be assessed using the Fornell-Larcker criterion and the Heterotrait-Monotrait (HTMT) ratio of correlations. Step 2: Assessment of the structural model. Once the measurement model is confirmed as reliable and valid, the structural model will be evaluated to test the study's hypotheses. This involves examining the path coefficients (β) for significance and relevance using a bootstrapping procedure with 5,000 resamples. The key criteria for assessment will include the coefficient of determination (R²): To measure the model's predictive power (i.e., the variance explained in the endogenous variables, particularly academic performance). Predictive relevance (Q²): Using the blindfolding procedure to assess the model's predictive accuracy. Effect sizes (f²): To evaluate the substantive impact of each independent variable on the dependent variables. Mediation analysis: To test the mediating role of cognitive load in the relationship between the AI-related antecedents (generative AI usage, trust in AI) and academic performance. The significance of the indirect effects will be assessed using bootstrapped confidence intervals.

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.