Dispositional foundations for AI-mediated learning: evidence from a latent profile analysis of grade-8 students in ICILS 2023

1 Introduction

1.1 Generative artificial intelligence (GenAI) and the transformation of learning environments

Artificial intelligence (AI) is rapidly transforming how students learn, teachers teach, and educational systems operate. While GenAI creates unprecedented opportunities for personalized learning, feedback, and knowledge construction, it also raises fundamental questions about equity, inclusion, educational purposes, and learners’ preparedness to participate effectively in AI-mediated environments (UNESCO, 2025). Beyond transforming teaching and learning, AI is reshaping educational systems, professional practices, and learners’ interactions with knowledge, creating both new pedagogical opportunities and significant societal challenges (OECD, 2026). The emergence of GenAI represents a major milestone in this evolution. Unlike previous generations of educational technologies, which primarily facilitated access to information and content delivery, GenAI systems can generate knowledge artifacts, provide contextualized explanations, support problem solving, and sustain interactive dialogue. These capabilities extend the pedagogical functions of digital technologies beyond information access, creating opportunities for personalized learning, formative feedback, and adaptive instructional support while simultaneously challenging established conceptions of teaching, learning, and assessment (Batista et al., 2024).

Evidence indicates that GenAI can effectively support self-directed learning when pedagogically integrated, with students reporting high levels of perceived usefulness and satisfaction following the use of AI-generated learning resources (Sánchez-Prieto et al., 2025). However, its educational effectiveness remains contingent upon how it is used. A recent systematic review reported mixed effects on student engagement, learning outcomes, and academic performance, suggesting that educational benefits depend less on the technology itself than on the pedagogical conditions under which it is implemented (Hon, 2026). Similarly, the OECD (2026) concludes that, although GenAI may improve immediate task performance, meaningful learning requires active cognitive engagement and explicit instructional guidance. Consistent with this perspective, effective educational engagement with AI depends not only on knowledge acquisition but also on learners’ confidence and their ability to critically evaluate AI-generated information (Hur, 2025). Consequently, the central educational question is no longer whether AI can support learning, but how learners can be prepared to benefit meaningfully from AI-mediated learning environments and which foundational competencies enable such participation.

1.2 From AI literacy to foundational preparedness for AI-mediated learning

The growing recognition that the educational value of GenAI depends on learners rather than technology has stimulated increasing interest in AI literacy. Initially conceptualized as the set of competencies required to understand, critically evaluate, communicate with, and effectively use AI systems, AI literacy marked an important shift from purely technical perspectives of digital competence toward a broader educational understanding of human-AI interaction (Long and Magerko, 2020). This perspective has subsequently been reinforced in school education. Reviewing AI literacy initiatives in primary and secondary education, Casal-Otero et al. (2023) argue that AI literacy extends beyond learning how to operate AI applications to include understanding how AI systems function, recognizing their growing presence in everyday life, and critically reflecting on their social, ethical, and educational implications. Together, these perspectives position AI literacy as a multidimensional educational construct rather than a set of technical or operational skills.

While AI literacy has become a central concept in educational research, attention has increasingly shifted toward the broader capacities that enable learners to engage effectively with AI. Building on a systematic review of K–12 AI education, Zhou et al. (2025) conceptualize AI competency as a multidimensional construct encompassing technological understanding, ethical awareness, collaboration with AI systems, and self-reflective capabilities. Consistent with this perspective, effective engagement with AI depends not only on technical competence but also on learners’ confidence and their capacity to critically evaluate AI-generated information (Hur, 2025). These developments broaden the educational understanding of AI by recognizing that successful participation in AI-mediated learning relies on cognitive, ethical, and metacognitive capabilities rather than AI knowledge alone. Nevertheless, they continue to emphasize the competencies that individuals possess, paying comparatively less attention to how these interact with learner dispositions and the learning resources available to students to shape their readiness for AI-mediated learning.

Building on this perspective, the present study conceptualizes the dispositional foundations for AI-mediated learning as the constellation of learner dispositions, namely ICT self-efficacy, attitudes toward technology and expectations of future ICT use, that may enable students to engage effectively with AI-mediated learning environments. We deliberately restrict the construct to its dispositional component. Digital competence and the learning resources available to students are treated as external conditions against which these dispositional configurations are validated and characterized, not as constituents of the construct itself. Rather than representing an alternative to AI literacy or AI competency, these dispositional foundations refer to enabling conditions that precede and may support the development and application of such competencies. Preparedness in this restricted sense is therefore understood not as a single capability but as the configuration of complementary dispositions that may shape students’ capacity to benefit from AI-mediated learning. This conceptualization provides the theoretical rationale for adopting a person-centered approach, allowing the identification of distinct dispositional patterns that cannot be captured through analyses of isolated competencies alone.

1.3 Foundational preparedness for AI-mediated learning

AI literacy increasingly builds upon broader digital, cognitive, and motivational competencies rather than emerging as an isolated capability. This view is also reflected in recent international policy frameworks emphasizing that education systems should prepare learners with the digital competencies, ethical awareness, and adaptive capacities required to participate effectively in increasingly AI-enabled societies (UNESCO, 2025).

Digital competence consistently emerges as the principal foundation of AI-related learning. Students with stronger digital competencies demonstrate higher levels of AI literacy, particularly in information evaluation and creative problem solving (Aleksic et al., 2025; Shomotova et al., 2026). Likewise, computer and information literacy (CIL) has been associated with more successful educational outcomes (Muammar et al., 2025), while recent reviews conceptualize AI literacy as a construct grounded in digital literacy, computational thinking, critical data literacy, and AI ethics (Yim and Su, 2025). These findings indicate that AI-related competencies emerge from broader educational capabilities that precede learners’ interaction with AI technologies.

Yet digital competence alone is insufficient. Learners’ engagement with AI is also shaped by motivational dispositions, twenty-first-century skills, prior AI experience, educational background, and access to digital resources (Alenezi, 2026; López-Secanell et al., 2025; Salhab and Aboushi, 2025; Shomotova et al., 2026). Accordingly, students’ preparedness for AI-mediated learning reflects the interaction between digital competencies, learner dispositions, and educational opportunities that collectively enable effective participation in AI-mediated learning environments.

Despite this growing consensus, empirical research has largely relied on variable-centered approaches. Although these studies have advanced understanding of the individual predictors of AI literacy and digital competence, they provide limited insight into how these characteristics coexist within individual learners. Students are unlikely to differ in a single competency; rather, they are expected to display distinct configurations of competencies and learner dispositions that shape their preparedness for AI-mediated learning. This perspective is reinforced by recent OECD evidence suggesting that emerging educational inequalities associated with GenAI may increasingly reflect differences in how students use AI to support meaningful learning rather than differences in access to technology alone, pointing to the risk of a “second digital divide” (OECD, 2026). Person-centered approaches provide a complementary perspective by identifying groups of students who share similar configurations of learner characteristics, thereby revealing patterns that remain hidden in conventional analyses (Khajavy and Lou, 2025; Ziernwald et al., 2022). Identifying these latent patterns therefore represents a necessary step toward understanding educational inequalities in the AI era, where students may differ not only in individual competencies but also in the constellation of characteristics that determines their capacity to engage effectively with AI-mediated learning environments.

1.4 ICILS 2023 as a framework for examining foundational preparedness

ICILS 2023 offers a valuable framework for examining the conditions that may support AI-mediated learning. It is important to be explicit about what it does not contain. ICILS 2023 includes no measure of AI literacy, no record of AI or GenAI use, no assessment of students’ ability to evaluate AI-generated output critically, no measure of ethical reasoning about AI and no observation of human-AI interaction. The study therefore cannot measure preparedness for AI-mediated learning directly. What its internationally comparable measures of CIL, ICT self-efficacy, attitudes toward technology, educational aspirations, digital experiences and access to ICT resources do provide are digital and motivational foundations that recent research and international policy frameworks identify as plausibly relevant to meaningful and equitable participation in AI-mediated learning environments (OECD, 2026; UNESCO, 2025). Throughout this paper these are treated as candidate foundations whose relevance to AI-mediated learning is a hypothesis to be tested, not a demonstrated result.

Recent studies using ICILS 2023 have begun to examine the educational implications of AI and digital competence from complementary perspectives, including school leaders’ expectations regarding AI integration (Eryilmaz et al., 2025), the predictors of students’ digital competence (Ortega-Rodríguez, 2026), and the role of school-based ChatGPT policies in shaping inequalities in CIL (Liu et al., 2026). Collectively, these studies demonstrate the growing relevance of ICILS 2023 for understanding educational responses to AI while leaving largely unexplored how multiple learner characteristics combine to shape students’ preparedness for AI-mediated learning.

To address this gap, the present study adopts a person-centered approach to identify latent profiles of foundational preparedness for AI-mediated learning among Grade-8 students using ICILS 2023 data. It examines how ICT self-efficacy, attitudes toward technology, and expectations of future ICT use combine within individual learners, whether these profiles differ in CIL, and how profile membership relates to socioeconomic status, gender, educational aspirations, and ICT resources. By moving beyond variable-centered explanations, the study provides a more comprehensive understanding of students’ preparedness for AI-mediated learning and the inequalities associated with different patterns of preparedness.

The following research questions guided the study:

• RQ1. What latent profiles of foundational preparedness for AI-mediated learning can be identified among Grade-8 students in ICILS 2023?
• RQ2. How do the identified profiles differ in terms of CIL?
• RQ3. To what extent are socioeconomic status, gender, educational aspirations, and ICT access associated with membership in different preparedness profiles?
• RQ4. What do the identified profiles reveal about inequalities in students’ preparedness to participate in AI-mediated learning environments?

2 Materials and methods

2.1 Data and participants

This study is a secondary analysis of the public use files of the IEA International Computer and Information Literacy Study 2023 (ICILS 2023), which assessed the CIL of Grade-8 students in 35 education systems (Fraillon, 2025). ICILS employs a two-stage stratified cluster design in which schools are sampled with probability proportional to size and students are randomly selected within sampled schools. The combined international student database contains 135,615 students. The analytic sample comprised the 118,764 students (87.6%) with complete scores on the four scales used as profile indicators; models involving covariates were estimated on the 115,537 students with additionally complete predictor data. As reported in Section 3.8, the excluded students differ systematically from those retained, so the analytic sample is positively selected. Sample sizes by education system, before and after case selection, are provided in the Supplementary Material.

2.2 Measures

Consistent with the conceptual framework presented in the Introduction, the dispositional foundations for AI-mediated learning were operationalized through the four ICILS scales that capture ICT self-efficacy, attitudes toward technology and expectations of future ICT use. Digital competence, educational aspirations and ICT access enter the analysis as external criteria and correlates, not as components of the construct.

Profile indicators were four questionnaire scales derived by IEA through item response theory modeling and reported as weighted likelihood estimates on an international metric with a mean of 50 and a standard deviation of 10: general ICT self-efficacy (students’ confidence in everyday digital tasks), specialized ICT self-efficacy (confidence in advanced technical tasks such as programming or configuring systems), positive attitudes toward ICT in society, and expectations of future ICT use in study and work. Because latent profile solutions are sensitive to indicator variances, all four scales were z-standardized using weighted means and standard deviations from the analytic sample.

CIL was used as a distal outcome rather than as a profile indicator. ICILS reports CIL as five plausible values on a scale with an international mean of 500 and a standard deviation of 100. Plausible values are random draws from each student's posterior proficiency distribution; they are designed for estimating population parameters, not for classifying individuals. Treating CIL as a distal outcome avoids this problem while preserving CIL as an external indicator of the digital competence component of foundational preparedness.

Predictors of profile membership were the international index of students’ socioeconomic background (IISB), student gender, educational aspirations (expecting to complete a university degree), and the number of computers or laptops available at home as an indicator of ICT access.

2.3 Statistical analysis

All analyses accounted for the ICILS complex sampling design. Pooled analyses used senate weights, which scale each system's contribution to a constant so that all systems contribute equally regardless of population size. Standard errors were estimated by jackknife repeated replication using the 75 replication zones defined in the database, and statistics involving CIL were computed separately for each of the five plausible values and combined using Rubin’s (1987) rules.

Latent profile models with one to six profiles were estimated by weighted maximum likelihood using the expectation-maximization algorithm implemented in the mclust package (Scrucca et al., 2016) for R 4.4.2 (R Core Team, 2024). This range was considered sufficient to capture meaningful heterogeneity while avoiding overextraction of very small or poorly interpretable classes in large-scale educational datasets. Two diagonal covariance structures were compared: equal variances across profiles and freely estimated variances. Because the weighted likelihood estimates take a limited number of distinct values, unregularized free-variance models produced degenerate solutions in which a profile collapsed onto a mass of identical scores; free-variance models were therefore estimated with the Bayesian regularization prior proposed by Fraley and Raftery (2007), and degenerate solutions were discarded.

Each model was estimated from 200 random initializations, half from k-means partitions and half from random posterior matrices, retaining the solution with the highest weighted log-likelihood; the number of admissible solutions and the proportion of starts reaching that optimum are reported in Table 2 so that the stability of each solution can be judged directly. Model selection combined the Akaike information criterion (AIC), the Bayesian information criterion (BIC), the sample-size-adjusted BIC (aBIC), relative entropy, a minimum class size of 5% and substantive interpretability. Because information criteria do not reach a minimum in samples of this size (Nylund et al., 2007), they were used to describe the fit of competing solutions rather than to select among them.

Classification quality was assessed with relative entropy, the average posterior probability of the assigned profile (AvePP) and the odds of correct classification (OCC), against the conventional criteria of AvePP above .70 and OCC above 5. Profiles were characterized and validated without assigning students to classes deterministically. For distal outcomes we used the BCH approach (Bolck et al., 2004), in which the inverse of the classification error matrix provides case weights that correct profile-specific means of external variables for classification uncertainty. Pairwise contrasts between profiles in CIL were tested on this basis, with standard errors from jackknife replication, the five plausible values combined under Rubin's rules, and effect sizes expressed in units of the international standard deviation of CIL (100 points). For predictors of membership we used the maximum likelihood three-step approach (Vermunt, 2010): the modal assignment serves as a single indicator of the latent profile with error rates fixed at the values estimated in the enumeration step, and a multinomial logistic model for profile membership is estimated by expectation-maximization. The largest profile served as the reference category. Standard errors for both analyses were obtained by repeating the estimation within each jackknife replicate.

Because digital competence is conceptually adjacent to the construct but is reported only as plausible values, its exclusion from the profile indicators was examined in a sensitivity analysis. The final model was re-estimated five times, once with each of the five plausible values of CIL added as a fifth standardized indicator, the resulting profiles were matched one-to-one with the retained solution by minimizing the distance between indicator means, and classification agreement was quantified with Cohen's kappa. Two further robustness analyses were conducted: a comparison of students retained in and excluded from the analytic sample, and three checks of the cross-system generalizability of the pooled solution, namely independent estimation within each education system, re-estimation leaving out one system at a time, and profile prevalence by system.

2.4 Ethics

The study analyzed publicly available, de-identified secondary data from ICILS 2023. According to institutional regulations, analyses of anonymized public datasets do not require additional ethical approval.

3 Results

3.1 Descriptive statistics

Before identifying latent preparedness profiles, we examined the relationships among the four profile indicators. Weighted correlations ranged from .20 to .61, with the strongest associations observed within the two self-efficacy measures (r = .58) and between positive attitudes toward ICT and expected future ICT use (r = .61). The moderate magnitude of these correlations suggests that the indicators represent related but non-redundant dimensions of foundational preparedness, supporting the use of a person-centered approach capable of identifying qualitatively different preparedness profiles (Table 1).

Table 1. Weighted descriptive statistics and correlations among the profile indicators.

N = 118,764 (analytic sample); descriptive statistics computed on the full international database (N = 135,615). Senate weights, so that each education system contributes equally. Profile indicators are reported on the international WLE metric (M = 50, SD = 10) and CIL on the IEA scale (M = 500, SD = 100); CIL descriptives average the five plausible values. All correlations are significant at p < .001.

3.2 Profile enumeration

The first research question asked whether distinct dispositional profiles could be identified among Grade-8 students. Latent profile models with one to six classes were estimated, each from 200 random initializations. Free-variance models consistently outperformed their equal-variance counterparts. As is characteristic of samples of this size, the information criteria decreased monotonically as classes were added and did not reach a minimum at any point in the range examined, so they cannot by themselves determine the number of profiles (Nylund et al., 2007). The proportion of starts reaching the best solution was low throughout, between 0.5 and 2.5 percent for the free-variance models with three or more profiles, which is why the large number of initializations matters: with a smaller search the retained solutions would have been local optima.

The four-profile solution was retained. We state openly that this is a judgment rather than a determination made by the data, and it rests on four grounds. It has the highest relative entropy of the candidate solutions (.78, against .71 for five profiles and .76 for six). It is parsimonious and each class exceeds five percent of the sample (smallest = 7.9%). It is the solution that replicates most consistently across education systems, as reported in Section 3.7. And it separates configurations that differ in shape, which is what the study is about, whereas the fifth class subdivides the majority average group into two groups that differ in level rather than in shape. Cross-tabulating the four- and five-profile classifications confirms this: the three profiles defined by their shape persist, while the average group splits in two. Most importantly, the efficacy-dominant profile, which carries the main finding of this study, is present in the four-, five- and six-profile solutions alike, so that finding does not depend on this decision. Model-fit statistics for all solutions, together with the number of admissible solutions and the proportion of starts reaching the optimum, are reported in Table 2.

Table 2. Fit indices for weighted latent profile solutions with one to six profiles.

The retained solution is the four-profile free-variance model. Free-variance models were estimated with the Bayesian regularization prior of Fraley and Raftery (2007). Each solution was estimated from 200 random starts; “Admissible” is the number of starts yielding a non-degenerate solution and “% at optimum” the proportion reaching the best log-likelihood. AIC is omitted for space; it ranks the solutions in the same order as BIC. Entropy is relative entropy. Information criteria decrease monotonically across the range examined and do not reach a minimum, so the number of profiles was not selected on fit; see Section 3.2.

Classification quality in the retained solution is adequate by conventional standards. Relative entropy is .78, the average posterior probability of the assigned profile ranges from 0.82 to 0.98 across the four profiles, and the odds of correct classification range from 5.7 to 778.9. All four profiles therefore exceed the usual thresholds of .70 and 5 respectively, and the efficacy-dominant profile is separated almost perfectly from the others. These indicators are reported by profile in Table 3, and the full classification matrix is provided in the Supplementary Material.

Table 3. Profile sizes, BCH-adjusted CIL means, and profile composition.

AvePP is the average posterior probability of the assigned profile and OCC the odds of correct classification; conventional criteria are AvePP above .70 and OCC above 5. CIL means combine the five plausible values with Rubin’s (1987) rules, with standard errors by jackknife repeated replication (75 zones). Composition columns use BCH weights, which correct for classification uncertainty.

3.3 The four preparedness profiles

The retained four-profile solution identified four distinct dispositional configurations (Figure 1). These profiles were not arranged along a simple continuum from low to high. They differed both in their overall level and in the relative balance between ICT self-efficacy and attitudinal dispositions.

The majority profile, labeled average (55.9% of students), scored slightly below the pooled mean on all four indicators (z between −0.37 and −0.11). The attitude-oriented profile (25.2%) combined clearly positive attitudes toward ICT (z = 1.04) and high expectations of future ICT use (z = 0.80) with only modest levels of general and specialized ICT self-efficacy (both around z = 0.25). The low-disposition profile (11.0%) scored markedly below the pooled mean on all four indicators (z between −1.12 and −0.81). Finally, the efficacy-dominant profile (7.9%) reported very high levels of general (z = 2.16) and specialized (z = 1.66) ICT self-efficacy while showing only moderately positive attitudes and future expectations (both around z = 0.60).

The two intermediate profiles are particularly noteworthy, and we deliberately avoid describing either of them as a high-preparedness profile. One is characterized primarily by positive attitudes and expectations toward ICT; the other by exceptionally high ICT self-efficacy that is not matched by measured competence, as Section 3.4 shows. Their crossover pattern across the four indicators indicates that these dispositional configurations differ not only in overall level but in how the dispositions are combined.

3.4 Profiles and CIL

The second research question examined whether the identified profiles differed in CIL. BCH-adjusted comparisons indicated statistically reliable but substantively modest differences. Profile means ranged from 472.1 (SE = 1.48) in the low-disposition profile to 495.4 (SE = 0.88) in the attitude-oriented profile, a difference of approximately 23 ICILS points or about 0.23 international standard deviations. The corresponding BCH-adjusted means are reported in Table 3.

Despite reporting ICT self-efficacy more than two standard deviations above the mean, the efficacy-dominant profile (486.7, SE = 1.34) did not meaningfully outperform the average profile (485.0, SE = 0.83). The highest CIL scores were observed instead among students in the attitude-oriented profile. We interpret this as a confidence-competence mismatch: for this group, self-reported capability and demonstrated capability diverge.

Pairwise contrasts confirm this formally. The difference between the efficacy-dominant and the average profile is 1.7 points (SE = 1.38, d = 0.017, p = 0.19), that is, statistically indistinguishable from zero: students reporting ICT self-efficacy more than two standard deviations above the mean do not differ in measured competence from students reporting unremarkable dispositions. The remaining contrasts are statistically reliable but small, the largest being between the attitude-oriented and low-disposition profiles (23.3 points, d = 0.233), and the efficacy-dominant profile scores significantly below the attitude-oriented profile (−8.8 points, d = −0.088). No contrast between profiles exceeds a quarter of an international standard deviation. All contrasts are reported in Table 4.

Table 4. Pairwise contrasts between profiles in computer and information literacy.

BCH-adjusted means; standard errors by jackknife repeated replication with the five plausible values combined under Rubin’s (1987) rules. Effect sizes are expressed in units of the international standard deviation of CIL (100 points).

These findings indicate that students’ dispositional preparedness for AI-mediated learning is related to, but distinct from objectively assessed CIL.

Having established that the profiles differed only modestly in CIL, the next step was to examine which student characteristics were associated with membership in each profile.

3.5 Predictors of profile membership

The third research question examined the extent to which socioeconomic status, gender, educational aspirations, and ICT access were associated with membership in the four preparedness profiles. Before estimating the multinomial model, BCH-weighted descriptive comparisons suggested that the efficacy-dominant profile tended to have the highest socioeconomic background and university aspirations, whereas the low-disposition profile included the highest proportion of girls and the lowest university aspirations. Home ICT access differed little across profiles.

The multinomial logistic model confirmed these patterns while considering all predictors simultaneously (Table 5). Because the predictors are measured on different scales, and odds ratios are therefore not comparable across them, we report average marginal effects on the predicted probability of profile membership, expressed in percentage points: binary predictors move from 0 to 1, the standardized socioeconomic index increases by one standard deviation, and the device count by one device. Gender showed the largest and most consistent contrasts. Being a girl was associated with a 5.3 point higher probability of membership in the low-disposition profile and with 7.9 and 3.6 point lower probabilities of membership in the attitude-oriented and efficacy-dominant profiles respectively.

Table 5. Average marginal effects on the probability of profile membership, in percentage points (jackknife standard errors).

N = 115,537 students with complete covariate data. Entries are average marginal effects in percentage points, with jackknife standard errors in parentheses. Binary predictors move from 0 to 1, the standardized socioeconomic index increases by one standard deviation and the device count by one device.

Socioeconomic background showed considerably smaller effects once educational aspirations and home ICT access were taken into account: a one standard deviation increase changed the probability of membership by at most one percentage point in any profile. University aspirations were associated with a 6.2 point higher probability of the attitude-oriented profile and a 4.4 point lower probability of the low-disposition profile, and the number of devices at home with changes of up to 2.5 points. Averaged across profiles, the absolute marginal effect was 5.8 percentage points for gender, 4.1 for university aspirations, 1.9 for home devices and 0.6 for socioeconomic background.

Taken together, and on a metric that permits comparison across predictors, gender showed larger and more consistent contrasts in profile membership than socioeconomic background. We describe the pattern in these terms rather than claiming that gender is a stronger determinant, since these are cross-sectional associations among correlated characteristics.

3.6 Sensitivity analysis

To assess robustness, the four-profile model was re-estimated five times, once with each of the five plausible values of CIL added as a fifth indicator. Using a single plausible value would have been inconsistent with the rationale given in Section 2.2. Across the five replications the profile structure remained essentially unchanged, with agreement between 89.9 and 98.7 percent and Cohen's kappa between .83 and .98.

Averaged over the five replications, standardized CIL means were 0.15 for the attitude-oriented profile, −0.15 for the low-disposition profile, −0.19 for the efficacy-dominant profile and −0.03 for the average profile. The efficacy-dominant profile fell below the pooled mean in four of the five replications and sat at the mean in the fifth. This reinforces the interpretation of that profile as combining exceptionally high perceived self-efficacy with unremarkable measured competence.

Overall, the sensitivity analysis supports the robustness of the four-profile solution and confirms that the distinction between dispositional preparedness and measured digital competence is not an artifact of treating CIL as an external outcome.

3.7 Robustness of the pooled solution across education systems

Because the analysis pools 35 education systems, three checks were carried out. First, the pooled model was re-estimated 35 times, leaving out one system at a time; the largest departure from the retained solution across all re-estimations was 0.007 standard-deviation units in the profile means, so the solution is not driven by any particular system. Second, all four profiles occur in all 35 systems, with prevalence ranging from 9.3 to 43.5 percent for the attitude-oriented profile, 3.9–17.5 for the low-disposition profile, 3.8–22.2 for the efficacy-dominant profile and 46.7–70.4 for the average profile. Third, and most demanding, the four-profile model was estimated independently within each system: the mean absolute departure from the pooled profile means was 0.38 standard-deviation units, and 23 of the 35 systems reproduced the three shape-defined profiles with a congruence of at least .90, while in the remaining systems the correspondence was weaker.

We report this mixed result as such. The pooled solution is robust in the sense that it does not depend on particular systems and that every profile is found everywhere, but independent within-system estimation does not recover the same configuration everywhere. The profiles should therefore be read as patterns in the pooled international sample rather than as universal student types, and a formal test of measurement invariance across systems remains a priority for future work.

3.8 Complete-case selection

Because the analytic sample excludes students with missing profile indicators, retained and excluded students were compared. The 16,851 excluded students (12.4%) scored 98.7 CIL points lower than those retained (387.5 against 486.2, close to a full international standard deviation), came from families 0.32 standard deviations lower on the international index of socioeconomic background, were ten percentage points less likely to expect to complete a university degree and had less access to digital devices at home. Exclusion rates also varied markedly across systems, from 0.9 percent in Chinese Taipei to 31.0 percent in Azerbaijan.

The analytic sample is therefore positively selected. Two consequences follow. The low-disposition profile is likely to be underrepresented, so its prevalence should be read as a lower bound; and because the students who drop out are disproportionately the least advantaged, the socioeconomic gradients reported here are likely to understate rather than overstate the true inequalities.

4 Discussion

4.1 Foundational preparedness as a multidimensional construct

The present study provides empirical support for conceptualizing foundational preparedness for AI-mediated learning as a multidimensional configuration of learner characteristics rather than as a single continuum. Recent international frameworks have argued that students’ readiness to engage meaningfully with AI depends on a combination of digital competence, self-efficacy, attitudes, and other learner characteristics rather than on any single capability (OECD, 2026; UNESCO, 2025). However, empirical evidence has relied predominantly on variable-centered approaches that examine these characteristics independently. By adopting a person-centered perspective, the present study demonstrates that students with comparable overall levels of preparedness may nevertheless differ substantially in how key learner characteristics are configured. This finding highlights the value of examining foundational preparedness as a constellation of interacting dispositions rather than as a single latent trait and illustrates how person-centered analyses can reveal meaningful heterogeneity that remains hidden in conventional variable-centered approaches (Morin et al., 2018; Wormington and Linnenbrink-Garcia, 2017).

The identification of two contrasting intermediate profiles illustrates this contribution particularly well. Although both display relatively favorable dispositions, they differ markedly in internal configuration. One is characterized primarily by positive attitudes toward ICT and expectations of future ICT use; the other by exceptionally high ICT self-efficacy that measured competence does not corroborate. Rather than representing successive points along a single continuum, these findings indicate that dispositional configurations can take qualitatively different forms. They should therefore be understood not only in terms of how much of each resource students possess but also in terms of how those resources are combined.

4.2 Foundational preparedness extends beyond CIL

A second finding concerns the relationship between these dispositional configurations and CIL. Differences between profiles were modest, and the efficacy-dominant profile did not achieve the highest measured CIL despite reporting exceptionally high ICT self-efficacy. This reinforces the distinction between objectively assessed digital competence and students’ perceptions of their own capabilities, and it is the reason we describe that profile as a confidence-competence mismatch rather than as a highly prepared group.

This interpretation aligns with previous ICILS research showing that ICT self-efficacy and CIL are related but empirically distinct constructs (Hatlevik et al., 2018). As Hatlevik et al. argued, it is important to distinguish between what students believe they are capable of accomplishing with ICT and what they are actually able to demonstrate through performance-based assessments. Likewise, Rohatgi et al. (2016) showed that different dimensions of ICT self-efficacy relate differently to measured CIL, highlighting that confidence in using digital technologies should not be interpreted as a direct indicator of competence. Together, these findings reinforce the interpretation that foundational preparedness for AI-mediated learning extends beyond digital competence alone.

This distinction has important implications for understanding students’ readiness to learn in AI-supported educational environments. Discussions surrounding AI in education often assume that students with stronger digital competence are also better prepared to benefit from AI-mediated learning. The present findings suggest that this assumption should be treated with caution. While CIL provides an essential foundation, preparedness for AI-mediated learning also depends on motivational beliefs, self-perceptions, and positive dispositions toward technology. In this sense, foundational preparedness should be understood as a broader construct that incorporates, but is not reducible to, objectively assessed digital competence.

As AI becomes increasingly integrated into educational contexts, this distinction may have practical implications, although we state them as hypotheses rather than as demonstrated results. No student in this study was observed using AI, and no profile has been shown to perform differently in an AI-mediated learning environment. What the evidence supports is the more modest claim that confidence, motivation and dispositions toward technology are candidate foundations, distinct from measured competence, whose relevance to AI-mediated learning should now be tested directly in AI-mediated contexts.

4.3 Gender and educational aspirations shape preparedness profiles more strongly than socioeconomic background

The third finding concerns the student characteristics associated with different dispositional configurations. On a common metric of average marginal effects, gender showed larger and more consistent contrasts than socioeconomic background, while educational aspirations consistently differentiated the profiles and home ICT access showed smaller associations. This pattern should be interpreted with care. Home internet access is close to a ceiling in this sample, above 95 percent in 28 of the 35 systems, so its limited explanatory role is not evidence that material access is unimportant; the number of devices at home, which is the measure that entered the model, is spread across all response categories.

The prominent role of gender is broadly consistent with previous research showing that gender differences in digital learning extend beyond objectively assessed competence to include students’ attitudes toward technology and other learner characteristics (Campos and Scherer, 2024; Hatlevik et al., 2018). In particular, Campos and Scherer (2024) showed that differences in attitudes toward technology partially explain gender differences in digital knowledge and skills, highlighting the interconnected nature of cognitive and attitudinal dimensions of digital learning. This pattern is particularly telling in ICILS 2023 itself, in which girls outperformed boys in measured CIL in 28 of the 32 education systems with comparable data, scoring 486 points on average compared with 467 points for boys (Fraillon, 2025). The profiles identified here therefore describe students who are, on average, more competent yet less confident, so that reading their preparedness from self-reports alone would systematically understate it.

An equally noteworthy finding concerns the role of educational aspirations. Students expecting to complete a university degree were consistently more likely to belong to the attitude-oriented and efficacy-dominant profiles, even after accounting for socioeconomic background and ICT access. Although educational aspirations have received comparatively limited attention in research on students’ digital dispositions, this suggests that future educational expectations may be linked to the motivational dispositions examined here, capturing broader patterns of engagement, confidence and future orientation.

Finally, the comparatively small contribution of home ICT access is also noteworthy. While equitable access to digital technologies remains an essential prerequisite for participation in AI-mediated learning, these findings suggest that material access alone is unlikely to explain why some students appear substantially better prepared than others. As AI becomes increasingly integrated into educational practice, inequalities in preparedness may depend less on the availability of technological resources and more on differences in students’ motivational, attitudinal, and self-efficacy-related characteristics. This interpretation aligns with recent international recommendations emphasizing that educational policies seeking to promote equitable AI adoption should move beyond questions of access and place greater emphasis on fostering the learner characteristics that underpin foundational preparedness for AI-mediated learning (OECD, 2026; UNESCO, 2025).

4.4 Implications for research and educational practice

Taken together, the findings of this study suggest that preparing students for AI-mediated learning requires a broader conceptualization than is typically adopted in discussions of digital competence alone. While CIL remains an essential foundation, the identified preparedness profiles indicate that students’ readiness to benefit from AI-mediated learning also depends on how motivational, attitudinal, and self-perceptual resources combine within individuals. Consequently, educational initiatives designed to support AI integration should consider these learner characteristics alongside the development of digital competence itself.

From a research perspective, the present study illustrates the value of person-centered approaches for advancing the study of AI-related educational readiness. Most previous research has examined learner characteristics independently, making it difficult to understand how different dimensions of preparedness coexist within students. By identifying distinct preparedness profiles, the present findings demonstrate that students with similar overall levels of readiness may nevertheless require different forms of educational support. This approach provides a complementary perspective to traditional variable-centered analyses and offers a useful framework for future research examining how different learner profiles respond to AI-mediated learning environments.

The findings also highlight the potential value of ICILS 2023 for research on AI in education. Although the study was not designed to measure AI literacy directly, its internationally comparable measures of digital competence, ICT self-efficacy, attitudes toward technology, educational aspirations, and contextual characteristics provide a robust foundation for examining the learner conditions that may support meaningful engagement with AI-mediated learning. As international large-scale assessments continue to incorporate AI-related constructs, combining these emerging measures with established indicators of digital competence and learner dispositions may offer a more comprehensive understanding of students’ preparedness for learning in increasingly AI-enabled educational systems.

4.5 Limitations

Several limitations should be considered when interpreting the findings of this study. First, the analyses were based on cross-sectional data, which precludes conclusions about the developmental relationships among the learner characteristics examined or about their influence on subsequent engagement with AI-mediated learning. Future longitudinal studies should examine how foundational preparedness evolves over time and whether different preparedness profiles are associated with distinct trajectories of AI use, learning processes, and educational outcomes.

Second, the construct was operationalized through four existing ICILS 2023 scales capturing general and specialized ICT self-efficacy, attitudes toward ICT and expectations of future ICT use. ICILS 2023 contains no measure of AI literacy, AI use, GenAI experience, critical evaluation of AI-generated output, ethical reasoning about AI or human-AI interaction, so these scales do not constitute a measure of preparedness for AI-mediated learning. Future research should develop dedicated instruments and examine how this construct relates to emerging measures of AI literacy.

Third, the analyses were conducted using pooled international data. As Section 3.7 reports, the pooled solution does not depend on any particular education system and all four profiles occur everywhere, but independent estimation within systems recovers the same configuration in about two thirds of them. The profiles are therefore patterns in the pooled international sample rather than universal student types.

Fourth, the analytic sample excludes the 12.4 percent of students with incomplete indicator data, and Section 3.8 shows that these students differ systematically from those retained: they score close to a full international standard deviation lower in CIL and come from less advantaged backgrounds. The sample is therefore positively selected, the low-disposition profile is likely underrepresented, and the socioeconomic gradients reported here are conservative.

Finally, the number of profiles is a judgment rather than a determination made by the data. Information criteria decreased monotonically across the range examined without reaching a minimum, and solutions with five and six profiles are admissible. We retained four on grounds of entropy, parsimony, cross-system replicability and interpretability, and we note that the efficacy-dominant profile, on which the main finding rests, is present in all three solutions.

5 Conclusion

This study provides one of the first international person-centered examinations of the dispositional foundations that may support AI-mediated learning, using ICILS 2023. Four distinct configurations emerged, showing that students differ not only in overall level but in how ICT self-efficacy, attitudes toward technology and expectations of future ICT use combine within individuals. These configurations are related to, but clearly distinct from, objectively assessed CIL, and one of them combines very high confidence with unremarkable measured competence.

The results further indicate that gender and educational aspirations are more strongly associated with preparedness profiles than socioeconomic background or home ICT access. Consequently, promoting equitable participation in AI-mediated learning is unlikely to depend solely on expanding access to technology or strengthening digital competence. It also requires educational policies and practices that foster the motivational, attitudinal, and self-perceptual learner characteristics that underpin students’ foundational preparedness for AI-mediated learning.

As AI continues to reshape educational systems worldwide, understanding how students differ in these characteristics becomes increasingly important. Whether the configurations identified here actually predict who benefits from AI-mediated learning is a question this study cannot answer and that we hope it helps to frame: the profiles are candidate foundations, and the next step is to test them in genuinely AI-mediated settings.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this work, the authors used Fable 5 (Anthropic) to assist with the preparation of the ICILS 2023 database, the development of the R code for the statistical analyses, and the drafting and editing of the manuscript text. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feduc.2026.1956540/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Sascha Schneider, University of Zurich, Switzerland

Reviewed by: Dech-Siri Nopas, Kasetsart University, Thailand

Ali Albashir Mohammed Alhaj, King Khalid University, Saudi Arabia