Generative AI and knowledge management in higher education: the impact of human development on student perceptions

Artificial intelligence (AI), particularly the emergence of large language models (LLMs), is fundamentally transforming the practice of knowledge management. Systems such as ChatGPT are capable of generating natural, context-sensitive responses, thereby supporting the acquisition, organization and application of information. These models not only automate tasks such as text generation and summarization but also contribute to enhancing organizational learning and decision-making efficiency (Chiarello et al., 2024). However, adopting this technology raises new concerns, particularly regarding data reliability, algorithmic bias and organizational trust (Nguyen, 2025).

Generative AI tools are playing an increasingly prominent role in the creation of new knowledge. This development is leading to a redefinition of classical knowledge management models. While traditional approaches have focused on transforming explicit and tacit knowledge, current systems enable new forms of collaboration and decision support (Leoni et al., 2024). Accordingly, the technology contributes not only as a tool. It also acts as a shaping factor in the development of sustainable knowledge processes (Jogezai et al., 2025). The role of ChatGPT in knowledge management is discussed in detail in Section 2.5. This is particularly relevant in higher education, where international students play an active role in the global flow of knowledge and bring new perspectives to educational processes through their diverse cultural backgrounds (Panda and Puri, 2025). However, these experiences are accompanied by differing socioeconomic conditions. The level of development in a student’s country of origin, commonly measured by the Human Development Index (HDI), is strongly associated with access to technology and the quality of educational infrastructure (Jaldemark et al., 2025). A detailed discussion of the HDI is provided in Section 2.2.

This study offers three key contributions. First, it is among the first empirical investigations to link the HDI with students’ perceptions of the impact of generative AI on higher education learning. Second, the research combines a multi-country sample with cluster analysis and a mediation model, providing a novel methodological approach to exploring learning patterns related to generative AI. Third, the study makes a theoretical contribution by offering an explanation for why students from developing countries tend to perceive the impacts of ChatGPT more positively, and how this relates to broader research on digital inequality.

Although an increasing number of studies explore the use of ChatGPT and other generative AI tools in higher education, there remains a lack of analyses that interpret students’ actual usage patterns and the national and individual factors influencing them. The aim of this study is to examine how socioeconomic development, particularly the HDI, and individual social background shape access to generative artificial intelligence, usage patterns and student attitudes in the context of higher education. The analysis is grounded in the theory of digital inequality, the socio-technical systems perspective and an interpretive framework based on knowledge management. A multi-level approach is applied, integrating individual, institutional and macro-level factors. The research questions are presented in detail in Section 2.6.

The structure of the study is as follows: Section 2 outlines the theoretical and conceptual frameworks; Section 3 presents the methodology; Section 4 reports the empirical findings; and Section 5 focuses on their interpretation and the limitations of the study. The final section offers conclusions and recommendations for future research.

This study makes three key contributions to the literature on generative artificial intelligence and digital inequality in higher education. First, it empirically demonstrates that socioeconomic development, operationalized through the HDI, is systematically associated with students’ perceived learning impacts of generative AI tools. Second, by combining mediation analysis and cluster analysis, the study shows that AI usage patterns function as a key mechanism through which socioeconomic context shapes educational outcomes. Third, the findings extend theories of digital saturation and compensatory technology use by providing cross-national evidence on how generative AI delivers differential value across development contexts.

This study builds on three interconnected theoretical frameworks that collectively support the interpretation of generative AI in higher education. These theoretical approaches offer explanatory perspectives at the levels of technological access, institutional integration and individual attitudes. First, Van Dijk’s (2020) theory of digital inequality highlights how socioeconomic background can influence access to technology, patterns of usage and the development of related skills. Different levels of access – physical, functional and outcome-related – can lead to varying user experiences. Within this framework, the HDI functions as a macro-level variable; it is discussed in detail in Section 2.2. The second theoretical pillar is the socio-technical systems approach, which argues that the impact of technology should be examined not in isolation, but in relation to institutional, pedagogical and cultural contexts. The theoretical background of usage intention and institutional factors is elaborated in Section 2.4. The third framework draws on the theoretical contributions of technology acceptance models, particularly the UTAUT model (Venkatesh et al., 2003). Although this model is not empirically tested in the present study, its core dimensions – such as perceived usefulness, social influence and perceived support – help interpret students’ varying attitudes toward generative AI tools.

These core theories are complemented by additional conceptual frameworks:

• The ethical use of AI incorporates principles of sustainability and equity (Banihashem et al., 2025).

• Theories of self-regulated learning (SRL) highlight the role of digital tools in supporting autonomous and reflective learning (Khalil et al., 2024).

• The contribution of generative chatbots – especially ChatGPT – to knowledge sharing and the development of metacognitive skills is discussed in detail in Section 2.5.

Another important theoretical perspective is digital saturation. This concept describes a state in which digital technologies are so pervasive and embedded in the educational environment that new tools – such as generative artificial intelligence – offer only limited added value. This saturation may result in students perceiving new technologies as less useful or using them less actively, especially when they feel overwhelmed by the abundance of existing tools or lack proper guidance for effective integration (Bulathwela et al., 2024; Yang et al., 2024). Although knowledge management does not serve as a standalone theoretical pillar in this study, its concepts and functions play a central role in the analysis. Generative AI tools – particularly ChatGPT – support activities such as acquiring, organizing, applying and sharing knowledge, which are closely linked to classical knowledge management processes. Accordingly, the study conceptualizes ChatGPT not merely as a technological tool but as an active participant in higher education knowledge processes.

The integrated approach of the three main theoretical pillars – structural (HDI), functional (technology use) and systemic (institutional integration) – enables a comprehensive examination of generative AI implementation across diverse national higher education contexts. The study identifies three interrelated levels of analysis: the macro level, represented by national development as measured by the HDI; the meso level, encompassing institutional and linguistic support; and the micro level, associated with individual background variables and attitudes. These levels exert their influence through a mediating variable – students’ knowledge management-oriented use of ChatGPT (USE_TYPE) – which links structural conditions to individual learning outcomes. The following section focuses on HDI as a key structural factor.

Taken together, the literature on digital inequality, generative AI adoption and knowledge management suggests that socioeconomic context does not merely condition access to technology but fundamentally shapes how AI tools are appropriated, valued and integrated into learning processes. This integrated perspective provides the theoretical basis for examining HDI as a structural determinant of AI usage patterns and perceived learning benefits

The HDI is a composite indicator developed by the United Nations to measure a country’s socioeconomic development based on health status, educational attainment and income levels (UNDP, 2023). In higher education, the HDI is not merely a background variable but a structural factor that influences the digital capacity of institutions and students’ opportunities for technology-based learning. Countries with a high HDI typically offer more advanced infrastructure, targeted competence development and broader opportunities for international mobility (Xie and Zhang, 2024). Van Dijk (2020) conceptualizes digital inequality across three dimensions: access, usage and digital skills. This framework is complemented in the present study by the concept of functional access, which considers not only the availability of technology but also students’ ability to engage in meaningful learning, considering user competencies and institutional support (Jin et al., 2025).

The development of functional access is shaped not only by infrastructure but also by the technical and pedagogical support provided by universities. In addition, linguistic accessibility is a key factor, especially in multilingual higher education environments, where the effectiveness of generative AI tools is influenced by language-related factors (Chiarello et al., 2024). Generative AI tools can have a dual impact: they may promote inclusive learning – for example, by improving access to educational materials – but in the absence of adequate support, they may also exacerbate existing disparities (Mimoudi, 2025).

The level of access is closely associated with knowledge processing, opportunities for knowledge sharing and educational performance (Walczak and Cellary, 2023). Therefore, the HDI is treated as a key structural variable in this study, as it may influence students’ access to AI tools and their usage patterns. To complement the macro-level assessment of the digital environment, the Information and Communication Technologies Development Index (IDI), developed by the International Telecommunication Union (ITU), also offers relevant insights (ITU, 2023), although it is not the primary focus of the present analysis. The following section examines the practical manifestations of functional access in learning and knowledge-sharing processes supported by generative AI.

Building on the theoretical framework of digital inequality, this section explores how access conditions and international learning environments influence the application of generative artificial intelligence (AI) in higher education. Knowledge sharing and knowledge management play a central role in developing key competencies such as critical thinking, intercultural sensitivity and SRL – all of which are essential for meaningful engagement in global learning spaces (OECD, 2021).

International mobility and inter-institutional cooperation facilitate the flow of tacit knowledge and foster a culture of mutual learning (Granić, 2025). These processes are also reflected in the use of generative AI tools – such as ChatGPT – which can support argumentation, source evaluation and the development of academic writing style, making them valuable components of the learning process. The specific impacts of ChatGPT on knowledge management are discussed in detail in Section 2.5. However, effective use of these tools requires an enabling learning environment and functional access. Students’ varying levels of technological preparedness, as well as demographic and instructional factors – such as gender, age or mode of study – also influence the perceived usefulness of generative AI tools (Zhao et al., 2024).

This section examines how knowledge management practices shape the educational integration of AI tools, with a particular focus on the following student perceptions:

• Academic impact: support for academic performance and task completion (Daniel et al., 2025);

• Developmental impact: enhancement of self-regulation and motivation (Iqbal et al., 2025); and

• Skill-enhancing impact: support for critical thinking and transferable skills (Daniel et al., 2025).

Inequitable access to digital resources remains a challenge and is a key issue for educational sustainability, aligning with the objectives of SDG4 (the United Nations goal to ensure inclusive, equitable and quality education for all) (Khlaif et al., 2025). The detailed conceptual and institutional connections are further elaborated in Section 2.5.

The educational application of generative artificial intelligence is not merely a technological issue; it is closely linked to the broader socioeconomic and institutional context. Students from less developed countries often face barriers – such as inadequate infrastructure or limited digital competence – that may reduce the educational utility of AI tools (Iqbal et al., 2025).

While generative AI – such as ChatGPT – can offer many advantages in the learning process, it does not inherently ensure equal opportunities. Mimoudi (2025) emphasizes that without institutional and policy support, these technologies may reinforce existing inequalities, particularly when they do not align with local educational practices. The hidden biases of LLMs can further exacerbate disparities, as they may favor digitally privileged groups while disadvantaging users with different linguistic or cultural backgrounds. This is a particularly important concern in international higher education, where equity is a core value.

Students’ engagement with AI tools is also shaped by individual factors – such as gender, age or mode of study – which may influence both usage frequency and perceived usefulness. Several frameworks have been developed to address these challenges, such as the SPADE model, which focuses on four key dimensions of responsible AI integration: sustainability, privacy, digital divide and ethics. Sustainability – in alignment with the goals of SDG4 – emphasizes inclusive and equitable use of technology, particularly in light of linguistic and cultural diversity (Khlaif et al., 2025).

Generative AI holds significant potential and risks, the extent of which is largely determined by the level of institutional support and existing inequalities (Cedeño-Salazar et al., 2025). Empirical research has also shown that students’ learning attitudes and behavior patterns can be clustered, often in association with their socioeconomic background and learning profiles. The role of institutional conditions – that is, how they enable or hinder the educational integration of digital tools – is examined in detail in the following Section 2.5.

The integration of generative artificial intelligence in higher education is not merely a matter of individual innovation but a process shaped by institutional and policy conditions. According to Jin et al. (2025), AI tools – especially collaborative systems – are transforming the ways knowledge is created, shared and applied, thereby redefining classical knowledge management functions. Sustainable integration is grounded in organizational adaptation. Institutional support extends beyond infrastructure to include curricular integration, teacher training and guidelines that promote ethical and inclusive technology use (Wang and Zhang, 2025). These tools become truly effective only when they are embedded in educational, research and administrative practices.

The digital competence of educators plays a crucial role. Jin et al. (2025) emphasize that such skills foster student engagement and personalized learning. In contrast, uncoordinated development efforts and unequal access may deepen disparities between institutions (Jiang et al., 2025). The concept of functional access can also be applied to systemic factors: targeted training, linguistic support and clear policy guidance are essential for effective AI integration. However, McDonald et al. (2025) highlight that many institutions still lack these components, which significantly affects learning outcomes.

Digital competence – particularly AI literacy – is essential for the conscious, critical and ethical use of technology. A higher level of competence facilitates the integration of AI tools into learning processes, thereby supporting adaptation at the institutional level. To understand the systemic implementation of AI, the socio-technical approach offers a valuable framework. According to Zamir et al. (2025), the success of AI integration depends not only on technological factors but also on organizational and cultural dimensions – such as the quality of decision-making and the regulatory environment. The following section builds on these theoretical foundations to present the research model and questions.

The aim of this research is to explore how the level of national development – as measured by the HDI – and individual socioeconomic factors are associated with students’ perceptions of generative AI use in higher education and its perceived learning impacts. The analysis is guided by five theoretically grounded research questions (RQ1–RQ5), based on the frameworks of digital inequality and socio-technical systems, and informed by the interpretive lens of knowledge management. The first research question focuses on the perceived learning benefits:

Previous studies suggest that generative AI tools – particularly through personalized and adaptive functions – can significantly enhance academic performance, support the development of SRL and strengthen self-confidence and critical thinking skills. ChatGPT, in particular, has been found useful in areas such as academic writing, content deepening and feedback provision (Jiang et al., 2025). The present study investigates these benefits across three dimensions: academic usefulness, personal development and skill enhancement. The second research question focuses on the macro-level impacts of national development:

This research question focuses on a continuous association between the national HDI score and perceived benefits, as opposed to RQ4, which explores categorical relationships via cluster membership. According to the theory of digital inequality, socioeconomic development can influence the quality of access to technology, user experiences and how those experiences are interpreted (Van Dijk, 2020). The HDI is a composite indicator that reflects a country’s educational, health and income conditions, and is therefore expected to be associated with the perceived educational benefits of AI tools (Holderegger and de Almeida Duarte, 2025). The third research question examines the diversity of student perceptions through cluster analysis:

The second-level digital divide – which focuses on the quality and purpose of technology use – provides the theoretical foundation for this question (Van Dijk, 2020). Cluster analysis enables the identification of student groups that interpret the learning impacts of generative AI tools in different ways. Previous research has shown that such differences may align with demographic (e.g. gender and age), academic (e.g. mode of study and field of specialization) and socioeconomic characteristics (Holderegger and de Almeida Duarte, 2025). The fourth research question interprets these clusters at the national level:

RQ4 investigates whether usage profiles based on perceived ChatGPT impacts tend to concentrate in countries with different HDI levels, thereby placing the clusters into a macro-level context and revealing potential associations between AI usage patterns and national development. The theory of digital stratification emphasizes that technology use is not merely a matter of individual choice, but is also shaped by cultural, institutional and structural factors (Neufeld et al., 2025). The fifth research question examines the role of generative AI from the perspective of knowledge management practices:

According to the socio-technical systems perspective, AI tools are effective when embedded in the learning environment and support students in structuring, processing and applying information (Grüneke et al., 2024). The literature also suggests that these tools play a role in various phases of SRL – including planning, execution and reflection (Mannuru et al., 2023). However, ethical, data privacy and transparency issues remain key regulatory challenges (Zhou et al., 2025). Together, the research questions form a multi-level analytical framework. RQ1 and RQ5 focus on the micro-level functioning of AI-supported knowledge management. RQ2 and RQ4 address macro-level effects associated with HDI, while RQ3 serves as a bridge between the micro and macro levels: although cluster formation is based on individual perceptions, the relationship of these clusters to socioeconomic background variables reveals structural-level patterns. This integrated approach enables the study to provide a comprehensive picture of how generative AI contributes to the evolving landscape of global educational inequality.

This study used a quantitative approach with a multi-stage analytical strategy to examine how students from different socioeconomic backgrounds perceive the learning-related impacts of ChatGPT. Data analysis was conducted using IBM SPSS Statistics 28 and SmartPLS 4. SPSS was used to perform descriptive statistics, correlation and regression analyses and cluster analysis (Field, 2022), while the PLS-SEM method enabled the exploration of both direct and indirect effects within a predictive, multivariate framework. It is particularly suited to mediation models and nonnormally distributed data (Cheah and Hair, 2025). The detailed analytical steps are presented in Section 3.4. The study included the following main constructs:

• ChatGPT Use (USE_TYPE): A three-item scale measured how frequently and for what purposes (e.g. structuring information, content generation and question answering) students used ChatGPT during their studies. This construct functioned as a behavioral pattern mediating the perceived learning impacts in the structural model.

• Perceived Learning Impacts: These were assessed across three dimensions: academic usefulness (Q26), personal development (Q27) and skill enhancement (Q28–29). The reliability of these constructs was confirmed using standard indicators (Cronbach’s alpha, AVE).

• Background Variables: Gender, age, level of study, field of specialization, student status and self-reported financial situation were not included as input variables in the cluster analysis but were examined in relation to cluster membership.

• Macro-Level Development (HDI): Country-level HDI values were used as indicators of the socioeconomic context (UNDP, 2023).

The aim of the analysis was to identify how students’ perceived benefits of ChatGPT use relate to both their individual socioeconomic backgrounds and national levels of development (HDI), with particular attention to the dimensions of learning support, accessibility and digital equity. The analysis, guided by the five research questions (RQ1–RQ5), followed these main steps: testing the reliability of scales, conducting correlation and regression analyses, performing cluster analysis and building a structural model.

The study is based on the international, public and anonymized data set Higher Education students’ Early Perceptions of ChatGPT (Ravšelj et al., 2025), organized by the CovidSocLab group at the University of Ljubljana under the leadership of Prof. Dr Aleksander Aristovnik. Data collection took place between October 2023 and February 2024, involving nearly 150 international researchers. One of the current study’s coauthors, Prof. Dr Andrea Bencsik, joined the scientific analysis phase after the data collection had concluded. The original sample included 23,218 student responses from 109 countries. After data cleaning, 11,910 valid responses from 58 countries were retained for analysis, based on the inclusion criterion of at least 35 responses per country. The data set and questionnaire are publicly available: Higher Education Students' Early Perceptions of ChatGPT: Global Survey DataLink to the cited article.

The questionnaire consisted of eleven sections and provided a comprehensive overview of students’ ChatGPT usage, attitudes and educational and social backgrounds. It covered demographic characteristics, frequency and purposes of use, perceptions of ethical and regulatory issues and both emotional and academic impacts of technology use. Of the 33 items included in the analysis, only those with adequate factor loadings and internal consistency were retained in the final model. As a result of the filtering process, 19 items were selected for construct development, as detailed in Section 3.4.

To comprehensively assess students’ socioeconomic environment, the HDI was used as a country-level aggregated background variable. Each respondent was assigned the HDI value corresponding to their country of origin, based on data from the 2023 United Nations Development Programme (UNDP).

In the analysis, the HDI served as an ecological indicator, reflecting the socioeconomic development of the respondent’s country. While the use of such variables entails methodological risks – particularly the potential for ecological fallacy – this approach is widely accepted and well-established in international research on education and digital inequality. It enables the examination of macro-level factors – such as national digital infrastructure or technology access – in relation to students’ perceived learning outcomes and AI usage patterns.

To examine students’ perceptions of ChatGPT’s impacts, we developed a comprehensive measurement framework. Based on 33 Likert-scale items, exploratory factor analysis identified three main dimensions:

• Academic impact – for example: “ChatGPT helps me complete my assignments.”

• Developmental impact – for example: “ChatGPT supports my ability to manage the learning process.”

• Skill-enhancing impact – for example: “ChatGPT improves my problem-solving skills.”

These constructs reflect students’ subjective perceptions of how ChatGPT contributes to their studies, personal development and skill acquisition. They do not capture objective academic performance but rather capture perceived learning benefits. The reliability of all three scales was high, with Cronbach’s alpha values exceeding 0.85. The factor structure was supported by principal component analysis (PCA) with Varimax rotation (Tabachnick and Fidell, 2013). Prior to PCA, we conducted a multistep item reduction of the original 33 Likert-scale items to exclude underperforming items. First, items with low communality (<0.40) or weak factor loadings (<0.50) were removed from the preliminary analyses. Next, items with high cross-loadings (≥0.40 on at least two components) were excluded due to poor alignment with the theoretically expected factor structure. Additional items were dropped due to content redundancy or to improve internal consistency (Cronbach’s alpha). As a result of this filtering process, 19 items were retained for the final analysis, forming a clear three-factor structure representing academic, developmental and skill-enhancing impacts.

The skill-enhancing dimension originally covered several subdomains (critical thinking, problem-solving and collaboration), but was modeled as a unidimensional, reflective construct due to strong correlations and high internal consistency among the items (AVE, CR). While a formative interpretation could also be theoretically justified, future studies may explore this as an alternative model (Cheah and Hair, 2025). In the structural model, the construct of knowledge management–oriented ChatGPT use (USE_TYPE) was included separately. This variable captures the extent to which students use ChatGPT to structure, process and apply information. It was modeled as a mediating and reflective construct due to high inter-item correlations, although a formative interpretation is also plausible. Findings by Kasneci et al. (2023) support the notion that students perceive ChatGPT as facilitating the development of workforce-relevant skills such as critical thinking and problem-solving.

To comprehensively address the research questions and test the proposed hypotheses, a multi-method statistical approach was applied. This analytical framework enabled both variable-centered (correlations, regressions) and person-centered (e.g. cluster analysis) strategies, culminating in the use of Partial Least Squares Structural Equation Modeling (PLS-SEM) to assess direct and indirect relationships within the model. The combination of these techniques allowed for a robust examination of patterns of ChatGPT use, their perceived impacts and their relationship with socioeconomic background.

As a first step, Pearson correlation and linear regression analyses were conducted to examine the relationship between the perceived academic, developmental and skill-enhancing impacts of ChatGPT use and the socioeconomic development level of the students’ country of origin, as measured by the HDI. These analyses are directly related to the first two research questions:

The analyses revealed significant differences associated with HDI level, with a negative relationship: students from lower-HDI countries reported greater perceived learning benefits from artificial intelligence tools, particularly ChatGPT. This finding supports the theoretical perspective of digital inequality.

To explore students’ varying experiences and usage patterns of ChatGPT, we conducted a cluster analysis. Hierarchical clustering (Ward’s method, agglomerative approach) was first performed based on the mean scores of the three reflective constructs – academic, developmental and skill-enhancing impacts. After determining the optimal number of clusters, the results were refined using K-means clustering. The selection of three clusters was based on the elbow method and the interpretability of results, in accordance with established methodological recommendations for cluster analysis (Ketchen & Shook, 1996). To assess the robustness of the cluster solution, we compared the results of hierarchical clustering (Ward’s method) with those of K-means clustering. Ward’s method also identified three clusters that matched the K-means solution exactly. The consistency between these two distinct methods confirms the stability and robustness of the cluster structure. The resulting groups were then compared across demographic and socioeconomic variables (gender, age, level of study, employment status and self-reported financial situation) to address the third research question:

To address the fourth research question, we also analyzed the extent to which cluster membership is associated with HDI level:

Based on the 2023 HDI data, countries were categorized into three groups (low, medium and high development), and a cross-tabulation analysis was conducted to examine the relationship between cluster membership and HDI level. This allowed us to investigate whether certain usage patterns are more prevalent among students from countries with different levels of development – providing an empirical test of the structural model of digital inequality.

This two-step procedure (hierarchical clustering followed by K-means refinement) was explicitly applied as a robustness check. The convergence of two conceptually different clustering approaches confirms that the identified cluster structure is not method-dependent but reflects stable and meaningful patterns in students’ perceived learning impacts.

The theoretical basis for the mediation analysis is drawn from the TAM and UTAUT models, which propose that technology use serves as a mediating variable between background factors and perceived outcomes. According to digital divide theory, socioeconomic development (HDI) influences the intensity of technology use, which may in turn be related to perceived learning benefits. This provides the rationale for including usage patterns as a mediator. Using PLS-SEM, we examined whether knowledge management–oriented ChatGPT usage (USE_TYPE) mediates the relationship between HDI and perceived learning impacts.

The use of mediation is justified by the possibility that the relationship between HDI and perceived learning benefits may be influenced by usage patterns. This is consistent with the socio-technical systems perspective, as well as with prior research demonstrating the mediating role of technology use (Cheah and Hair, 2025). The model addresses the following research question:

In the model, HDI was included as an exogenous variable, while knowledge management-oriented ChatGPT use (USE_TYPE) served as the mediating construct. The USE_TYPE mediator was derived from three items in the questionnaire’s Q18 block: Q18a, Q18c and Q18e. These items measured students’ knowledge management-oriented ChatGPT use. The three items formed a single reflective construct, as they were conceptually aligned and showed strong inter-item correlations. The USE_TYPE variable was created by averaging the item scores, and internal consistency was satisfactory, with both composite reliability (CR) and average variance extracted (AVE) values exceeding the recommended thresholds.

The academic, developmental and skill-enhancing impacts were defined as endogenous, reflective variables. This approach aligns with recent studies, such as the work of Al-Sharafi et al. (2022), which also demonstrated that factors supporting knowledge application can act as mediators in the sustainable use of AI tools in education. Our findings support the theoretical assumption that knowledge management – related usage patterns play a key role in the mechanism of impact within generative AI – supported learning environments.

The hypotheses formulated in this study are closely aligned with the five research questions (RQ1–RQ5) and are grounded in well-established theoretical foundations, including the theory of digital inequality, the socio-technical systems approach and relevant knowledge management literature. The primary objective is to examine how the socioeconomic development level of students’ countries of origin (as measured by the HDI) is associated with their ChatGPT-related learning experiences and the perceived impacts of those experiences. The research model not only seeks to identify associations but also considers underlying mechanisms, such as patterns of knowledge management-oriented usage (USE_TYPE), as potential mediators.

For RQ1 – which aims to explore general academic, developmental and skill-enhancing impacts – no formal hypothesis is required. However, this question serves as a foundation for subsequent analyses. According to the literature, such self-reported evaluations can offer relevant insights into the educational potential of generative AI (Chaparro-Banegas et al., 2024). In response to RQ2, we hypothesize that students come from countries with varying HDI levels, which may be associated with differences in how they perceive their learning experiences with ChatGPT. Empirical evidence – including findings from Suárez and García-Mariñoso (2025), Dhakal et al. (2025) and Szabó (2024) – supports the idea that socioeconomic background, infrastructure and individual competencies all influence the effective use of AI tools. Based on these considerations, we formulated the following hypothesis and its sub-hypotheses:

The first two research questions examine whether students’ perceived learning impacts of ChatGPT vary across countries with different levels of socioeconomic development. Drawing on the concept of the secondary digital divide, which emphasizes differences in the quality of technology use and perceived benefits rather than mere access, we expect that students’ experiences with generative AI tools are associated with the HDI level of their country of origin. Accordingly, the following hypotheses are formulated:

The RQ3 focuses on whether students form distinct user groups based on their learning-related experiences with ChatGPT. Prior research suggests that the use and perceived benefits of AI tools are shaped by multiple factors, including digital competencies, trust and socioeconomic background, resulting in heterogeneous user patterns rather than uniform adoption. Based on this reasoning, we propose the following hypotheses:

The RQ4 examines whether the distribution of these ChatGPT-use clusters varies across countries with different HDI levels. Previous studies highlight that beyond technical access, educational background and digital skills play an important role in shaping how students benefit from AI-based tools. Therefore, we formulate the following hypothesis:

Finally, the RQ5 explores whether the relationship between socioeconomic development and perceived learning impacts is direct or mediated by the way ChatGPT is used. Socio-technical systems theory and prior empirical findings suggest that usage patterns may function as an intermediate mechanism through which contextual factors influence educational outcomes. Given the mixed evidence in the literature, no specific directional assumption is made. Accordingly, we test the following mediation hypothesis:

To ensure conceptual clarity and methodological transparency, Figure 1 illustrates the theoretical model, summarizing the study’s conceptual framework, the research questions (RQ1–RQ5) and the corresponding hypotheses (H1–H5). The model visualizes the relationships among national development level (HDI), perceptions of generative AI and usage patterns. The theoretical foundation draws on approaches to digital inequality and the social embeddedness of technology, both of which suggest that the effectiveness of AI applications depends on social, institutional and infrastructural conditions. According to Hammerschmidt and Stolz (2025), AI can only be inclusive if it is embedded within systems that ensure access. Hayes et al. (2024) argue that achieving democratic educational goals requires more than the presence of AI; adequate infrastructure, support and regulation are also essential. Suárez and García-Mariñoso (2025) emphasize that digital inequality manifests not only at the technical level, but also in terms of skills and institutional capacity.

At the core of the model are three reflective constructs: the perceived academic, developmental and skill-enhancing impacts of ChatGPT use. These constructs were measured using Likert-scale items and served as the basis for the descriptive analyses in RQ1. Based on these constructs, we conducted a cluster analysis to identify distinct user types (RQ3, H2), which were subsequently compared based on demographic and socioeconomic characteristics (H3). The model also investigates whether cluster membership is associated with the HDI classification of students’ countries (RQ4, H4), that is, whether certain usage patterns are more prevalent among students from countries at different levels of development.

Finally, the model examines whether HDI indirectly influences perceived learning benefits through knowledge management-oriented ChatGPT use (USE_TYPE) (RQ5, H5), with USE_TYPE serving as a mediating variable. This interpretive framework aligns with the socio-technical systems theory, which posits that the impact of digital learning technologies is shaped not only by technical factors but also by the social and institutional context in which they are applied.

HDI is not positioned at the far left of the model because it plays both direct and indirect roles in shaping perceived educational outcomes. The visual structure reflects the sequential and interdependent nature of the analytical design rather than a strictly linear causal flow.

The majority of respondents were full-time students (86%), primarily enrolled in undergraduate programs (84%), with a slight female majority in the sample (54%). The most common fields of study were social sciences (40%) and applied sciences (37%), while both the humanities and natural sciences accounted for approximately 12% each. Based on self-reports, half of the students rated their financial situation as average, nearly one-quarter considered it below average and about one-fifth assessed it as above average (see Table 1).

As an initial step in the analysis, we examined the relationship between students’ country-level development (measured by HDI) and their positive experiences with ChatGPT use. Pearson correlation and linear regression analyses revealed a weak but significant negative association between HDI and the perceived academic, developmental and skill-enhancing impacts (see Table 2). This suggests that students from lower-HDI countries tend to view the learning-supportive role of generative AI – particularly ChatGPT – more positively.

The results indicate that HDI values – although only to a limited extent – are statistically significantly correlated with the degree to which students perceive the academic, developmental and skill-enhancing impacts of ChatGPT. Given the noncausal nature of the study, these relationships should not be interpreted as evidence of cause-and-effect. In addition, several uncontrolled factors (e.g. educational infrastructure, digital access and student competencies) may also influence these associations. The negative but statistically significant associations indicate that the lower a student’s country-of-origin HDI value, the more positively they tend to perceive the learning-supportive role of artificial intelligence. This finding provides strong support for the first hypothesis (H1) and its sub-hypotheses (H1a–H1c). Based on student responses, three clearly distinguishable groups were identified, each showing significant differences in perceived impact. According to the ANOVA results, all three examined dimensions – academic, developmental and skill-enhancing – differed significantly across the clusters (Table 3). It is important to note that the regression models did not include control variables, so the results reflect statistical associations that may be influenced by potential confounding factors.

The theoretical foundation for the cluster analysis was provided by the concept of the secondary digital divide, which suggests that inequalities in higher education are not limited to access but also manifest in how technologies are used and the benefits users perceive. By clustering students based on their experiences with ChatGPT, the goal was to identify usage types that go beyond demographic characteristics and reflect distinct user patterns. This approach offered a sound theoretical and methodological framework for answering RQ3 and RQ4.

The cluster analysis was performed using the K-means algorithm via SPSS (Analyze → Classify → K-Means Cluster). Input variables consisted of the aggregated means of the scales measuring perceived impacts of ChatGPT (Q26–Q29). The number of clusters was determined in advance based on hierarchical clustering, using the largest jumps in the agglomeration schedule to identify three well-separated groups. Demographic variables (e.g. gender, age and level of study) were not included in the clustering process, as the aim was to identify experience-based user types. These background variables were examined post hoc as independent variables.

The procedure resulted in three clearly interpretable clusters. The robustness of the cluster solution is supported by the high consistency between the hierarchical (Ward’s method) and K-means clustering results:

• Cluster 1 (N = 4,512): “Techno-Optimists", who reported high perceived benefits of ChatGPT across all dimensions.

• Cluster 2 (N = 1,465): “Skeptics", who reported low scores and attributed few advantages to AI use.

• Cluster 3 (N = 5,933): “Development-Oriented", who reported moderate academic impacts but notably high developmental and skill-enhancing impacts.

Cluster membership was significantly associated with several background variables, as confirmed through tests of independence or ANOVA, thus supporting H3 and its sub-hypotheses (H3a–H3d):

• Gender: Male students were overrepresented in the techno-optimist cluster, while female students were more common in the development-oriented group (Chi² = 980.854; p < 0.001).

• Age: Development-oriented students were statistically significantly older (F = 18.221; p < 0.001).

• Financial background: Techno-optimists more frequently reported being in below-average financial situations (Chi² = 1366.488; p < 0.001), suggesting that ChatGPT may have been perceived as especially useful and accessible for them.

• Employment status: Development-oriented students were more likely to be working alongside their studies (Chi² = 1649.960; p < 0.001).

In addition, a significant association was found between cluster membership and the HDI value of students’ country of origin, based on a test of independence (Chi² = 1286.474; p < 0.001). Students from lower-HDI countries were disproportionately represented in the Techno-Optimist cluster, providing further support for H4.

To examine the dimensional structure of students’ perceived impacts of ChatGPT use, we conducted a PCA on 19 items to establish the constructs for the mediation model and to assess their factorial validity. The Kaiser–Meyer–Olkin (KMO) measure indicated excellent sampling adequacy (KMO = 0.93), and Bartlett’s test of sphericity was highly significant (χ² ≈ 18,500; p < 0.001), confirming the suitability of the data for factor analysis. PCA with Varimax rotation revealed three distinct components: Academic Impact, Developmental Impact and Skill-Enhancing Impact. Together, these components explained 69.1% of the total variance. All factor loadings exceeded the threshold of 0.76, and no item exhibited substantial cross-loading, supporting both the conceptual clarity and discriminant validity of the constructs (Tables 4 and 5).

The aim of the mediation analysis was to explore how the level of development of students’ countries of origin (HDI) is directly and indirectly related to the perceived academic (ACAD_IMPACT), developmental (DEV_IMP) and skill-enhancing (SKILL_DEV) benefits associated with ChatGPT use. In the model, HDI was specified as the exogenous variable, ChatGPT usage patterns (USE_TYPE) as the mediator, and the perceived benefits as endogenous constructs. The mediator variable USE_TYPE was based on the average of three items from the Q18 block (Q18a, Q18c, Q18e), while the outcome variables were constructed using the scales from Q26 to Q29 (see Figure 2).

The structural model was estimated using SmartPLS 4 and was based on a reflective measurement model. All item loadings exceeded the recommended threshold of 0.70 for each construct, the AVE values were all above 0.50 and the CR values exceeded 0.70 across all constructs, indicating strong convergent validity. Discriminant validity was also confirmed, as all Heterotrait–Monotrait (HTMT) ratios remained below the 0.85 threshold.

The validity and reliability of the measurement model were assessed using standard indicators recommended in the PLS-SEM literature. Convergent validity was evaluated using AVE, which reflects the proportion of variance captured by a construct relative to variance due to measurement error. CR was preferred as the indicator of internal consistency in PLS-SEM because it better reflects the weighted reliability of indicators than Cronbach’s alpha. Discriminant validity was assessed using the HTMT ratio, widely regarded as the most reliable criterion for assessing whether constructs are adequately distinct from one another (Cheah and Hair, 2025).

Based on the structural model, HDI showed a weak but statistically significant negative relationship with all three dimensions of perceived impact:

1. Academic Impact (ACAD_IMPACT: β = –0.050; p = 0.003);

2. Developmental Impact (DEV_IMP: β = –0.053; p = 0.002); and

3. Skill-Enhancing Impact (SKILL_DEV: β = –0.072; p = 0.001).

Additionally, HDI had a significant negative effect on knowledge management-oriented ChatGPT use (USE_TYPE) (β = −0.162; p < 0.001).

The usage patterns (USE_TYPE) exerted a strong and statistically significant positive effect on all three outcome variables:

1. Academic Impact (β = 0.345; p < 0.001);

2. Developmental Impact (β = 0.363; p < 0.001); and

3. Skill-Enhancing Impact (β = 0.339; p < 0.001).

The indirect (mediated) impacts were negative and statistically significant for all three dimensions:

1. Academic Impact (β = –0.056);

2. Developmental Impact (β = –0.059); and

3. Skill-Enhancing Impact (β = –0.055); with all p-values < 0.001.

Since the model is based on cross-sectional data, the path coefficients indicate statistical associations, not causal relationships. However, in all cases, the 95% confidence intervals did not include zero, suggesting that the indirect impacts are robust and statistically reliable. In terms of total impacts, HDI exhibited a negative association with perceived learning outcomes (β = −0.106 to −0.127; p < 0.001), whereas USE_TYPE consistently showed the strongest positive effect (β ≈ 0.34–0.36; p < 0.001).

These findings indicate that students from higher-HDI countries are less likely to use ChatGPT for knowledge management purposes, which may partly explain their lower levels of perceived academic and developmental benefits. Nevertheless, given that the model does not control for confounding variables, the mediation relationships should not be interpreted as causal. All direct, indirect, and total impacts are summarized in Table 6.

It is important to note that several uncontrolled factors – such as differences in educational systems, digital access and student competencies – may influence the observed relationships and should therefore be considered when interpreting the results. Table 7 summarizes the outcomes of all tested hypotheses (H1–H5), indicating whether they were supported or rejected based on the statistical analyses presented above.

Taken together, the regression-based hypothesis testing and the cluster analysis provide complementary insights into students’ experiences with ChatGPT. While H1 and H2 capture average-level relationships between HDI and perceived impacts, the cluster analysis (RQ3–RQ4) reveals qualitatively distinct usage patterns that would remain hidden in variable-centered models.

Taken together, these results provide a broader picture of how socioeconomic context shapes both the perceived learning benefits and usage patterns of generative AI in higher education.

The present study examined how socioeconomic development (measured primarily by the HDI shapes the use of generative artificial intelligence in higher education, as well as how it influences students’ attitudes and perceived learning experiences in the context of digital inequality. The findings of the mediation model (RQ5) indicate that students from lower-HDI countries evaluated the educational utility of ChatGPT significantly more positively than their counterparts from more developed regions. This pattern highlights systematic differences in how students integrate artificial intelligence into their learning processes across socioeconomic contexts. The results align with the theory of digital saturation (Bulathwela et al., 2024), which suggests that in high-HDI environments, novel technologies such as ChatGPT may offer diminishing perceived added value due to technological saturation or institutional fatigue. Recent research also indicates that exposure to AI-generated knowledge may activate cognitive and trust-related mechanisms that reduce perceived usefulness or even lead to knowledge avoidance, underscoring the importance of contextual and user-specific factors in AI-supported learning (Li and Yuan, 2026).

H5 tested a nondirectional mediational pathway, which was statistically supported. Students from countries with higher HDI values were significantly less likely to use ChatGPT for knowledge management purposes (USE_TYPE), and because this form of usage had a strong positive effect on perceived learning benefits, the indirect effect of HDI was negative. This finding is consistent with recent knowledge management research, which shows that the performance impacts of human–AI interaction are largely indirect and operate through knowledge-related processes such as knowledge sharing rather than through direct technological adoption (Chen et al., 2026). While this finding challenges linear models that directly link technological adoption to levels of socioeconomic development, it provides empirical evidence for the presence of a mediating mechanism. This interpretation is consistent with prior research suggesting that in less developed settings, digital technologies often serve a compensatory function, offering relatively greater perceived educational value. Nonetheless, a more nuanced understanding of this relationship requires further investigation that accounts for institutional infrastructure, learning environments and students’ prior technological experiences.

From a theoretical perspective, these findings can be interpreted through the lenses of digital saturation and compensatory technology use.

Synthesizing findings across methods, the results suggest that HDI influences not only the magnitude of perceived learning impacts (H1) but also the configuration of usage patterns identified through cluster analysis (H4).

Taken together, these findings indicate that socioeconomic context is associated with both the intensity and the forms of students’ engagement with generative AI tools.

Bulathwela et al. (2024) emphasize that generative artificial intelligence can only reduce educational inequalities if it is supported by adequate infrastructure, institutional capacity and regulatory frameworks – a finding that highlights the exceptional value of generative AI in resource-constrained contexts. The results of the cluster analysis support this claim by identifying three distinct student groups – technology optimists, skeptics and development-oriented users – which illustrates that the use of AI is not homogeneous and that usage patterns differ significantly. Similarly, Yang et al. (2024) note that the acceptance and perceived benefits of AI are strongly shaped by the socioeconomic environment. According to the theory of compensatory technology use, students with fewer resources tend to adopt AI more consciously and strategically to achieve their learning goals. Socio-technical systems theory also highlights that the successful integration of artificial intelligence depends not only on individual digital competencies but also on the quality and alignment of institutional conditions. Feuerriegel et al. (2024) stresses the need for harmonizing technological and social systems to ensure responsible AI use – a view echoed by Khullar et al. (2025), who argue that the diffusion of AI hinges not only on technical infrastructure but also on human-centered and organizational factors. Ethical and regulatory considerations have increasingly been formalized at the institutional level – for example, the University of Cambridge (2023) regards the submission of AI-generated content without proper citation as academic misconduct, while New York University requires students to disclose how they used AI tools in their coursework (Chan, 2023). This study also affirms that appropriate use of generative AI can support the development of higher-order thinking and critical reflection. Bates et al. (2020) emphasize that fostering autonomous and reflective learning is closely linked to institutional pedagogical practices. This is particularly evident among development-oriented users, who use AI not just to retrieve information but to engage in deeper processing, self-reflection, and skill development.

From a practical standpoint, countries with high HDI values should focus on promoting strategic and knowledge-building uses of AI in education – thereby maximizing its added value. In contrast, in lower-HDI contexts, the priority remains to ensure equitable access without deepening existing educational disparities. The true learning-support potential of generative AI can only be realized through pedagogical innovation, context-sensitive integration and strong institutional support (Bulathwela et al., 2024; Yang et al., 2024). Nevertheless, it is important to acknowledge that the study’s findings are based on statistically significant correlations rather than causal relationships. The analyses did not control for potential confounding variables – such as institutional regulations, instructor attitudes, cultural differences students’ prior digital experiences – all of which may influence the frequency and perceived benefits of AI use. Furthermore, the self-reported nature of the data limits objectivity, and the perspectives of educators were not included in the study.

While the HDI serves as a critical structural indicator, it does not – by itself – determine learning outcomes. What ultimately matters is the extent to which generative AI tools are embedded within broader social, pedagogical and institutional systems. Although these technologies offer substantial potential to democratize education, they cannot fulfill this promise independently. Their transformative impact depends on responsible, intentional integration – aligned with institutional practices and pedagogical goals. This interpretation is consistent with recent scholarship, which highlights that the true educational value of AI lies not in the technology per se, but in its purposive and contextually grounded implementation (Hammerschmidt and Stolz, 2025). At the same time, the potential risks of generative AI must not be underestimated – including algorithmic bias, the production of inaccurate or misleading content, and threats to academic integrity. Overdependence or misuse may also result in increased plagiarism or superficial learning. Accordingly, the integration of generative AI should be transparent, pedagogically sound and ethically grounded – to mitigate negative consequences and enhance its contribution to student development and higher-order thinking (Zhang et al., 2025).

Beyond higher education practice, the findings also carry broader societal and policy relevance. The results highlight that inequalities in the benefits derived from generative AI tools are not solely driven by access but are strongly shaped by usage patterns, digital competencies and institutional contexts. From a policy perspective, this suggests that national and institutional AI strategies should move beyond infrastructure provision and explicitly support pedagogically guided and inclusive forms of AI use. In countries with different HDI levels, targeted policies may be required to prevent generative AI from reinforcing existing educational and socioeconomic inequalities. More broadly, the study contributes to ongoing debates on digital inequality by demonstrating that the societal impact of AI in education depends not only on technological diffusion but also on how learning-related practices are supported and regulated.

The study highlights that generative artificial intelligence, particularly ChatGPT, can become a truly effective tool to support learning in higher education only if it is integrated not merely as a technological innovation but as a deliberately embedded pedagogical element. The results indicate that the HDI shows a significant relationship with students’ AI usage patterns and their perceived learning benefits. While higher levels of technological saturation in high-HDI countries may reduce the perceived added value of new tools, in lower-HDI regions, generative AI often plays a compensatory role due to limited access to learning opportunities. Based on these findings, the following recommendations can be made:

• In developed countries, it is advisable to encourage the use of AI for knowledge-building purposes, especially by supporting problem-based, research-driven and self-directed learning formats.

• In developing regions, priority should be given to ensuring equitable access, establishing digital mentoring programs and developing foundational skills, particularly digital and language competencies.

• For educators, methodological and ethical training is essential to support the responsible integration of generative AI, recognize its potential dilemmas and guide students in using the technology consciously.

• At the institutional level, clear and consistent regulations are needed, especially in the areas of assessment and examinations, so that students clearly understand the boundaries of AI use.

• Developers must take into account social and cultural differences to ensure that AI-based tools are well adapted to local linguistic and digital environments.

Although the study is not representative and relies on self-reported data, it offers valuable insight into global student perceptions. The results are better interpreted as trends rather than generalizable conclusions. Future research should involve longitudinal, controlled studies and include the perspectives of educators and institutions. The theoretical framework of knowledge management (KM) served in this case not as a formal model, but as an interpretive background that helped to understand students’ AI usage and learning experiences. Future studies could explore KM processes, such as knowledge creation, sharing, and application, in the context of generative AI integration in education. The application of generative AI goes beyond institutional practice. It plays a strategic role in educational policy, economic development, and promoting social equity. It is especially important that the technology be integrated into national digital inclusion strategies, particularly in countries where low HDI levels reflect systemic educational inequality. Development initiatives should be aligned with local labor market needs and competency development goals in a fair and inclusive manner. At the societal level, generative AI may contribute to lifelong learning and improved employability, especially in regions where access to formal education or qualified instructors is limited. Strengthening digital and language skills is essential for this to be effective. Generative AI is not only a technological tool but also a social mirror. It reflects how we think about learning, how we respond to change, and what actions we are willing to take toward the meaningful, fair, and sustainable integration of technology.

Overall, this study contributes to ongoing debates on digital inequality by demonstrating that the educational value of generative AI is not uniform but context-dependent. By linking socioeconomic development to AI usage patterns and perceived learning outcomes, the findings highlight the importance of theory-informed and context-sensitive integration of generative AI in higher education.

This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

No ethical approval was required for this study as it involved anonymous survey data collection and adhered to the principles of voluntary participation and data confidentiality.

The authors confirm that the submitted manuscript is original, has not been published before and is not currently under consideration for publication elsewhere. All authors have read and approved the final version of the manuscript.

Generative AI tools such as ChatGPT were not used for data analysis, but were used to support language editing and clarity in manuscript drafting, under the full supervision and final review of the authors.