Integrating AI and EdTech into Inclusive Learning: A Cross-Regional Study of Russia and Kazakhstan

Abstract

This article evaluates how artificial intelligence (AI) and educational technology (EdTech) support inclusive learning in Russia and Kazakhstan, two Eurasian countries that share post-Soviet educational legacies but differ in their levels of digital infrastructure and teacher preparedness. Using an asymmetric mixed-methods design, the study draws on a primary survey of N = 2570 educators and staff in four Russian cities (Moscow, Saint Petersburg, Yekaterinburg, Novosibirsk; October–December 2024; response rate 59.8%) and secondary policy/indicator analysis for Kazakhstan. Russia exhibits higher broadband access, AI/EdTech platform adoption, and teacher digital skill levels compared with Kazakhstan. Structural equation modeling (SEM; SmartPLS 4.1) tested four latent constructs—learning environment (LE), general digital competencies (HCg), specialized AI skills (HCs), and inclusion (I)—with satisfactory validity (AVE > 0.5; HTMT ≤ 0.85). A three-stage Measurement Invariance of Composite Models (MICOM) procedure confirmed configural, compositional, and full mean/variance invariance across Russian city subgroups, enabling pooled path analysis. Kazakhstan indicators from secondary sources are discussed as a descriptive benchmark. Semi-structured interviews with 24 stakeholders (12 Russia, 12 Kazakhstan; March 2025; analyzed with NVivo 14) revealed four themes: policy coherence, teacher readiness, infrastructure access, and ethical AI governance. Key SEM paths were LE → HCg (β = 0.278), HCg → HCs (β = 0.652), and HCs → I (β = 0.188), all p < 0.001. A formal mediation analysis confirmed a significant indirect effect across the full LE → HCg → HCs → I chain. The findings indicate that infrastructure is necessary but insufficient: the key to inclusion lies in sustained development of both basic and specialized digital skills, supported by coherent policies and continuous professional development. China and India are discussed as secondary international benchmarks drawn from published reports, not as sites of primary data collection.

1. Introduction

The global expansion of artificial intelligence (AI) and educational technology (EdTech) has intensified interest in their potential to foster inclusive learning environments—those capable of accommodating learners with diverse needs, abilities, and socioeconomic backgrounds (Adeleye et al. 2024; Mohammed and Watson 2019). AI-driven adaptive platforms can personalize instructional pathways for students with disabilities (Alkhawaldeh and Khasawneh 2023), while EdTech ecosystems support collaborative and technology-enhanced pedagogies (Su and Zou 2022; Zubiri-Esnaola et al. 2020). Systematic reviews confirm that adaptive learning systems in higher education improve differentiated instruction outcomes compared with traditional teacher-led approaches (Muñoz et al. 2022; Wang et al. 2023).

Yet these benefits are neither automatic nor evenly distributed. Infrastructure gaps, limited teacher preparedness, and a persistent “technology-first” mindset have been shown to constrain inclusive outcomes, particularly in low- and middle-income settings (Mohammed and Watson 2019; Tan 2023; Tapalova and Zhiyenbayeva 2022). Ethical concerns surrounding data-driven bias and algorithmic transparency add complexity: AI systems that are not designed with equity in mind may inadvertently marginalize the learner groups they are intended to support (Nguyen et al. 2023; Ntoutsi et al. 2020). International organizations have therefore called for trustworthy AI frameworks that embed fairness, accountability, and pedagogical relevance into educational technology adoption (Pedro et al. 2019; UNESCO 2019; Vincent-Lancrin and Van der Vlies 2020; OECD 2024).

Russia and Kazakhstan share a post-Soviet educational legacy and have both launched national digital transformation programs—Russia’s “igital Economy“ and ”Cadres for Digital Economy“ programs and Kazakhstan’s ”Digital Kazakhstan“ initiative—yet they differ substantially in infrastructure maturity, EdTech adoption rates, and teacher digital skill levels (Federal State Statistics Service 2023; Bureau of National Statistics of Kazakhstan 2023). Despite this natural comparative pair, no empirical study to date has modeled the pathways from learning environment quality through digital competencies to inclusive outcomes in Russian urban educational ecosystems while situating the findings against Kazakhstan’s secondary indicators and policy context.

This study addresses that gap through an asymmetric mixed-methods design combining primary survey data from Russian educators, secondary indicator and policy analysis for Kazakhstan, and stakeholder interviews in both countries. The quantitative SEM is based exclusively on Russian primary data, while Kazakhstan serves as a descriptive and qualitative benchmark. The following research questions guide the investigation:

• RQ1: What are the structural pathways linking learning environment quality, general digital competencies, specialized AI skills, and inclusive learning outcomes in Russian educational institutions?
• RQ2: To what extent is the measurement model invariant across Russian city subgroups, permitting pooled path analysis? How do Kazakhstan’s secondary indicators compare descriptively with the Russian findings?
• RQ3: What policy, institutional, and pedagogical factors do stakeholders identify as facilitating or hindering AI-supported inclusive education in each country?

Based on prior evidence that digital skill development follows a staged progression (Ghosh 2024; Garg and Sharma 2020) and that infrastructure alone does not ensure inclusive outcomes (Shivani et al. 2024), we propose the following hypotheses:

H1: 

Learning environment quality is positively associated with general digital competencies (LE → HCg).

H2: 

General digital competencies are positively associated with specialized AI skills (HCg → HCs).

H3: 

Specialized AI skills are positively associated with inclusive learning outcomes (HCs → I).

H4 

(methodological validation): The measurement model is invariant across Russian urban contexts (MICOM partial or full invariance).

By triangulating quantitative path modeling with qualitative stakeholder perspectives, this research contributes empirically grounded recommendations for policymakers, educators, and technology developers seeking to advance inclusive digital learning in post-Soviet Eurasian contexts. Kazakhstan is included as a descriptive benchmark based on secondary indicators and qualitative interviews, not as a site of primary survey data collection. Where relevant, international benchmarks from China and India are drawn from published secondary sources to further contextualize the findings.

2. Literature Review

2.1. Infrastructure-Centered Studies: Digital Environments as Preconditions

A first strand of research emphasizes the role of digital infrastructure and learning environments as prerequisites for technology-enhanced inclusion. AI-powered assistive technologies have emerged as a central mechanism for supporting students with learning disabilities (Alkhawaldeh and Khasawneh 2023; Zdravkova 2022). Adaptive learning platforms adjust instructional content, pace, and difficulty to individual learner profiles, and systematic reviews indicate that they improve outcomes in differentiated instruction compared with uniform, teacher-led approaches (Muñoz et al. 2022; Wang et al. 2023). Speculative yet practical frameworks for leveraging AI in inclusive classroom transformation have been proposed (Mohammed and Watson 2019), alongside more targeted, locally developed tools for disabled learners in India (Shivani et al. 2024) and predictive support systems for special-needs students (Garg and Sharma 2020). However, these contributions tend to focus on technology design rather than on the educator competency pathways that mediate between infrastructure provision and inclusive practice.

2.2. Teacher Competency-Centered Studies: Digital Skills as Mediators

A second strand foregrounds teacher digital competencies as the critical link between available technology and pedagogical outcomes. As AI adoption in education accelerates, ethical considerations have moved to the foreground. Nguyen et al. (2023) outline core ethical principles for educational AI, while Ntoutsi et al. (2020) document biases embedded in data-driven systems that may inadvertently marginalize certain learner groups. International bodies emphasize the need for trustworthy AI that balances innovation with equity (Vincent-Lancrin and Van der Vlies 2020; Pedro et al. 2019; UNESCO 2019; OECD 2024). The role of AI in protecting vulnerable populations in post-pandemic contexts has also been explored, particularly in India (Rani 2024a, 2024b). These works highlight that ethical and competent use of AI tools requires staged skill development—from basic digital literacy to specialized AI capabilities—yet few studies empirically model this staged progression or examine how general competencies serve as prerequisites for specialized ones.

2.3. Inclusive Outcomes and Cross-Regional Policy Perspectives

A third strand addresses the translation of digital resources and skills into measurable inclusive outcomes across different policy contexts. Regional EdTech practices vary widely. Ghosh (2024) proposes a roadmap toward inclusive, AI-powered ecosystems in India. Singh and Agarwal (2015) provide earlier insights from Asian educational conferences, and Su and Zou (2022) contribute theoretical foundations for technology-enhanced language learning. Interactive group work supported by EdTech has been shown to foster inclusive participation (Zubiri-Esnaola et al. 2020). In the Russia–Kazakhstan context, the two countries’ national digital strategies present a natural comparative case: Russia’s federal programs prioritize large-scale teacher training and platform deployment, while Kazakhstan’s ”Digital Kazakhstan“ initiative focuses on building foundational digital infrastructure alongside inclusive learning modules (Federal State Statistics Service 2023; Bureau of National Statistics of Kazakhstan 2023). Taken together, these three strands demonstrate that infrastructure, teacher competencies, and inclusive outcomes are each necessary dimensions of digital education, yet they also reveal a significant gap in the existing literature that the following subsection addresses.

2.4. Research Gaps

The three strands reviewed above share a common limitation: they tend to investigate infrastructure, teacher competencies, and inclusive outcomes in isolation rather than as sequential elements of an integrated pathway. Infrastructure-centered studies establish that digital environments matter but do not trace how they shape educator skills; competency-centered studies demonstrate the importance of teacher digital literacy but rarely link it to measurable inclusion; and policy-oriented studies compare national strategies without modeling the mechanisms through which policy translates into classroom-level outcomes. This fragmentation leaves a critical gap: no empirical study to date has modeled the full sequential chain from learning environment quality through general and specialized digital competencies to inclusive learning outcomes in post-Soviet educational contexts. The present study addresses this gap by proposing and testing a staged pathway model (LE → HCg → HCs → I) using primary survey data from Russian urban educational institutions, with Kazakhstan serving as a descriptive benchmark based on secondary indicators and qualitative interviews.

3. Materials and Methods

This study applied an asymmetric mixed-methods design to analyze the integration of artificial intelligence (AI) and educational technologies (EdTech). The design combined a primary cross-sectional survey (Russia), secondary indicator and policy analysis (Kazakhstan), structural equation modeling (Russian primary data), and semi-structured stakeholder interviews in both countries.

3.1. Research Questions and Hypotheses

The research questions (RQ1–RQ3) and hypotheses (H1–H4) are stated in Section 1. The quantitative component addresses RQ1 and RQ2 via SEM and MICOM (Russian city subgroups); the qualitative component addresses RQ3 via thematic analysis of interviews. Kazakhstan indicators are discussed descriptively.

3.2. Research Design and Sampling

The primary quantitative dataset was obtained from a large-scale survey conducted between October and December 2024 among 2570 university-level employees and educators across four major Russian cities: Moscow, Saint Petersburg, Yekaterinburg, and Novosibirsk. These cities were selected based on their digital infrastructure development and varying levels of EdTech adoption. Participants were recruited through institutional email lists and departmental coordinators at public universities; a stratified random sampling approach ensured representation of both teaching staff (68% of respondents) and administrative/technical employees (32%). The questionnaire included Likert-scale items (1–5) to evaluate basic and advanced digital skills, use of AI platforms in education, familiarity with adaptive learning systems, and perceptions of inclusivity. The instrument comprised 35 items organized into four construct blocks (LE: 8 items; HCg: 10 items; HCs: 8 items; I: 9 items). An additional block of seven items on student engagement (SE) was included in the survey instrument for exploratory purposes; however, SE was not part of the hypothesized structural model (H1–H3) and is therefore excluded from the reported SEM analysis. The response rate was 59.8% out of 4300 distributed questionnaires.

For cross-regional analysis, publicly available datasets and policy documents were reviewed from key educational institutions and government portals in Kazakhstan (specifically Almaty and Astana). Sources included UNESCO’s Inclusive Education Reports, OECD EdTech indicators, and national digital education strategies. Kazakhstan-level indicators (broadband coverage, AI/EdTech adoption rates, digital skill assessments) were extracted from the Bureau of National Statistics of Kazakhstan (2023) and the Ministry of Digital Development, Innovation and Aerospace Industry of Kazakhstan (2022). These secondary data enable a descriptive benchmarking comparison; however, no primary survey was administered in Kazakhstan, and this limitation is acknowledged throughout the analysis.

Additionally, published reports and statistical yearbooks from China (CAICT 2023; CNNIC 2023; CIE 2023; NDRC 2022; Ministry of Education PRC 2023) and India (NASSCOM 2024; NITI Aayog 2023; TRAI 2024; Ministry of Education India 2024; Venture Intelligence 2023) were consulted as international contextual benchmarks. These data are discussed in a dedicated subsection and are not used in the SEM or MICOM analyses.

3.3. Modeling and Analytical Tools

To estimate the relationship between digital competencies, AI tool adoption, and the effectiveness of inclusive education practices, we employed structural equation modeling (SEM) using SmartPLS 4.1. The model included four latent constructs: learning environment (LE, 8 indicators), general digital competencies (HCg, 10 indicators), specialized AI skills (HCs, 8 indicators), and inclusion (I, 9 indicators). The SEM followed a partial least squares path modeling (PLS-PM) approach, appropriate for exploratory models with formative and reflective indicators.

Measurement model assessment included outer loadings (≥0.70 threshold), composite reliability (CR ≥ 0.70), rho_A, and average variance extracted (AVE > 0.50) for convergent validity. Discriminant validity was assessed via the heterotrait-monotrait ratio of correlations (HTMT ≤ 0.85) and Fornell–Larcker criterion. Structural model evaluation included path coefficients (β), t-values, p-values via bootstrapping (5000 resamples, bias-corrected and accelerated confidence intervals), effect sizes (f²: 0.02 small, 0.15 medium, 0.35 large), predictive relevance (Q² via blindfolding), coefficients of determination (R²), and standardized root mean square residual (SRMR < 0.08). To assess the sequential mediating role of HCg and HCs in the LE → I relationship, a formal mediation analysis was conducted. The specific indirect effect (LE → HCg → HCs → I) was estimated via bootstrapping (5000 resamples) with bias-corrected confidence intervals. Additionally, the model included control variables—city (Moscow, Saint Petersburg, Yekaterinburg, Novosibirsk), years of professional experience, and job role (teaching staff vs. administrative/technical staff)—to account for potential heterogeneity in the sample. An alternative model specification with a direct path from LE to I was also tested to assess whether the mediated pathway provides a better theoretical and empirical fit than a direct-effect model.

A three-stage MICOM (Measurement Invariance of Composite Models) procedure was used to test invariance across the four Russian city subgroups. Stage 1 assessed configural invariance (identical model specification and estimation). Stage 2 tested compositional invariance via permutation-based correlation of composite scores (c = 1.0 within the 5% quantile; 5000 permutations). Stage 3 evaluated equality of composite means and variances using permutation p-values (p > 0.05 for full invariance). Non-parametric permutation tests were applied to detect statistically significant differences between city-level subgroups. Full invariance at all three stages permits pooled path estimation across subgroups, as composite scores carry equivalent meaning across cities. Kazakhstan data, drawn from secondary sources, are presented as a descriptive benchmark and are not included in the MICOM or SEM analyses.

3.4. Qualitative Content Analysis

Qualitative data included policy reviews, institutional EdTech strategy documents, and stakeholder interviews conducted in March 2025 in Russia and Kazakhstan. A total of 24 semi-structured interviews were conducted: 12 in Russia (3 per city: Moscow, Saint Petersburg, Yekaterinburg, Novosibirsk) and 12 in Kazakhstan (6 in Almaty, 6 in Astana). Participants included university administrators (n = 8), EdTech coordinators (n = 6), classroom educators (n = 6), and policymakers from regional education ministries (n = 4). Participants were selected via purposive sampling based on their involvement in digital education initiatives. Interviews lasted 35–60 min and followed a protocol covering: (a) institutional AI/EdTech adoption experiences, (b) teacher training adequacy, (c) infrastructure access, (d) policy coherence, and (e) ethical AI governance concerns. Texts were analyzed thematically using NVivo 14.0 to identify recurring patterns related to AI-supported inclusion, educator training gaps, infrastructure challenges, and regulatory frameworks. Coding proceeded in two rounds: open coding of all transcripts followed by axial coding to consolidate initial codes into themes. Inter-coder reliability (Cohen’s κ = 0.81) was established between two coders on a 20% subsample. Although the interview protocol contained five initial topical domains, the axial coding process yielded four higher-order themes because institutional AI/EdTech adoption experiences (topic a) did not emerge as an analytically distinct category; rather, adoption-related observations were distributed across the themes of policy coherence, teacher readiness, and infrastructure access. This thematic consolidation may have reduced the visibility of implementation-specific nuances and is acknowledged as a limitation.

3.5. Ethical Considerations

This study did not involve direct experimentation with human subjects. Informed consent was obtained from all survey and interview participants. The research protocol was approved by the Ethical Committee of Peter the Great St. Petersburg Polytechnic University (approval code: ECR/2024/08-PTPU-EdAI).

3.6. Data and Code Availability

Survey data, questionnaires, and analysis code are available upon request from the corresponding author. Secondary datasets used in the study (UNESCO, OECD, Rosstat) are publicly accessible and referenced in the manuscript. No proprietary or restricted data sources were used.

4. Results

This section presents the empirical outcomes of the study organized by: infrastructure and access (descriptive comparison of Russia and Kazakhstan), human capital, SEM (Russian primary data), measurement invariance (MICOM across Russian city subgroups), qualitative findings, forecast analysis, and international contextual benchmarking.

4.1. Infrastructure and Digital Access

Access to reliable digital infrastructure represents a key prerequisite for the effective implementation of AI and EdTech solutions. Comparative data from Rosstat (Federal State Statistics Service 2023) and the Ministry of Digital Development, Innovation and Aerospace Industry of Kazakhstan (2022) highlight a clear disparity: as of 2022, 74.1% of Russian organizations reported broadband internet availability, whereas in Kazakhstan this figure was lower at 64.2%, pointing to a noticeable gap in digital infrastructure development between the two countries.

As shown in

Table 1. Comparative Indicators of Digital Competence and AI/EdTech Integration (Sources: Federal State Statistics Service 2023, form 3-inform; Bureau of National Statistics of Kazakhstan 2023; Ministry of Digital Development, Innovation and Aerospace Industry of Kazakhstan 2022).

Indicator	Russian Federation	Kazakhstan
Level of digitalization	High (74.1% organizations with internet access)	Moderate (64.2% organizations with internet access)
Share of institutions using AI/EdTech platforms (%)	68%	42%
National digital inclusion strategies	“Digital Economy” Program; “Cadres for Digital Economy”	“Digital Kazakhstan” Program with inclusive learning modules
Average level of digital skills (1–5)	3.35 (general), 2.65 (specific)	3.10 (general), 2.20 (specific)
Government support	Strong (federal grants, broadband, EdTech platforms)	Moderate (limited grants, infrastructure under development)

, Russia demonstrates a higher level of digitalization (74.1%) compared to Kazakhstan (64.2%). The adoption of AI/EdTech platforms shows a notable disparity, with 68% of Russian institutions versus 42% in Kazakhstan. Russia’s ”Digital Economy“ programs provide a more comprehensive policy framework than Kazakhstan’s ”Digital Kazakhstan“ initiative. Digital skills assessment reveals Russia maintaining a moderate advantage in general (3.35 vs. 3.10) and specific skills (2.65 vs. 2.20).

4.2. Human Capital and Skill Development

General digital skills in Russia average 3.35 (on a scale from 1 to 5), compared to 3.10 in Kazakhstan. Specific (AI/EdTech-related) skills show a wider gap: 2.65 in Russia versus 2.20 in Kazakhstan. These findings reflect Russia’s stronger institutional investment in digital pedagogy. The skill gap, particularly in specific digital competencies, indicates a need for targeted professional development in Kazakhstan. This observation motivates H2 and H3, which propose that general competencies serve as a prerequisite for specialized capabilities.

4.3. Structural Equation Modeling (SEM)

The structural equation model, developed using SmartPLS 4.1, explored how four latent factors interact to shape inclusive learning outcomes. The measurement model assessment is presented in Table 2.

Table 2. Measurement Model Summary (Convergent Validity and Internal Consistency).

Construct	Items	Loadings Range	CR	rho_A	AVE
LE	8	0.72–0.88	0.91	0.90	0.57
HCg	10	0.70–0.85	0.92	0.91	0.54
HCs	8	0.74–0.89	0.93	0.92	0.62
I	9	0.71–0.86	0.90	0.89	0.53

Note: All outer loadings exceed the 0.70 threshold; CR ≥ 0.70 and AVE > 0.50 for all constructs, confirming convergent validity. HTMT values for all construct pairs were ≤0.85, satisfying discriminant validity (full HTMT matrix available upon request). VIF values for all indicators were <3.5, indicating no multicollinearity concern. The structural model results are presented in Table 3.

Table 2. Measurement Model Summary (Convergent Validity and Internal Consistency).

Construct	Items	Loadings Range	CR	rho_A	AVE
LE	8	0.72–0.88	0.91	0.90	0.57
HCg	10	0.70–0.85	0.92	0.91	0.54
HCs	8	0.74–0.89	0.93	0.92	0.62
I	9	0.71–0.86	0.90	0.89	0.53

Note: All outer loadings exceed the 0.70 threshold; CR ≥ 0.70 and AVE > 0.50 for all constructs, confirming convergent validity. HTMT values for all construct pairs were ≤0.85, satisfying discriminant validity (full HTMT matrix available upon request). VIF values for all indicators were <3.5, indicating no multicollinearity concern. The structural model results are presented in Table 3.

Table 3. Standardized Path Coefficients in the Structural Model.

Path	Coefficient (β)	t-Value	p-Value	f2	R2	Q2
LE → HCg	0.278	14.2	0.000	0.08 (S)	0.077	0.041
HCg → HCs	0.652	49.1	0.000	0.74 (L)	0.425	0.254
HCs → I	0.188	12.5	0.000	0.04 (S)	0.035	0.019

Note: S = small, M = medium, L = large. Bootstrapping: 5000 resamples, BCa confidence intervals. SRMR = 0.062 (<0.08 threshold). All three hypothesized paths (H1, H2, H3) are supported.

Table 3. Standardized Path Coefficients in the Structural Model.

Path	Coefficient (β)	t-Value	p-Value	f2	R2	Q2
LE → HCg	0.278	14.2	0.000	0.08 (S)	0.077	0.041
HCg → HCs	0.652	49.1	0.000	0.74 (L)	0.425	0.254
HCs → I	0.188	12.5	0.000	0.04 (S)	0.035	0.019

Note: S = small, M = medium, L = large. Bootstrapping: 5000 resamples, BCa confidence intervals. SRMR = 0.062 (<0.08 threshold). All three hypothesized paths (H1, H2, H3) are supported.

Based on the structural equation modeling results, the quality of learning environments is significantly associated with the development of general digital skills among educators (β = 0.278 (p < 0.001)), supporting H1. In Russia, this pattern corresponds to a 23% increase in teacher digital competence since 2019, when the national ”Digital Pedagogical Competence“ program was launched, reaching over 345,000 educators (Ministry of Education RF 2023).

The strong association between general and specific digital skills (β = 0.652, p < 0.001) supports H2 and demonstrates a critical developmental pathway. This suggests that foundational digital literacy is a necessary precondition for developing advanced AI-related competencies—a staged progression that should inform training program design in both countries.

The R² value of 0.077 for HCg indicates that learning environment quality accounts for approximately 7.7% of the variance in general digital competencies. While statistically significant, this modest explanatory power suggests that numerous other factors not captured in the model—such as institutional culture, prior training history, individual motivation, and organizational support—also contribute substantially to general digital skill development. The R² of 0.035 for inclusion (I) similarly indicates that specialized AI skills explain only a small portion of variance in inclusive outcomes, reflecting the multifaceted nature of inclusion. A formal test of the specific indirect effect across the full chain LE → HCg → HCs → I yielded a significant indirect effect (β_indirect = 0.034, 95% BCa CI [0.027, 0.041], p < 0.001), confirming that the sequential mediation is statistically significant. An alternative model with a direct path from LE to I was tested; the direct effect was non-significant (β = 0.024, p = 0.417), supporting the fully mediated specification. City, years of professional experience, and job role (teaching vs. administrative/technical) were included as control variables. City showed no significant effect on any endogenous construct, consistent with the MICOM results confirming measurement invariance. Years of experience was positively associated with HCg (β = 0.058, p = 0.011) but not with HCs or I. Job role showed no significant associations. The inclusion of controls did not substantively alter the magnitude or significance of the hypothesized paths. The positive but small association of specialized digital skills with inclusive learning outcomes (β = 0.188, p < 0.001) supports H3. The effect size is small (f² = 0.04), which is noteworthy given the complexity of inclusive education. The small effect size for LE → HCg (f² = 0.08) suggests that while infrastructure is necessary, it is insufficient on its own—a finding consistent with the literature on technology-first approaches in developing education systems (Mohammed and Watson 2019; Tan 2023).

4.4. Measurement Invariance (MICOM)

A three-stage MICOM procedure was conducted to test whether the measurement model is invariant across the four Russian city subgroups, permitting meaningful pooled analysis of path coefficients. The MICOM results for the Russian city subgroups are presented in

Table 4. MICOM Results: Russian City Subgroups.

Construct	Stage 1	Stage 2 (c)	Stage 2 p-Value	Stage 3 Mean Diff	Stage 3 Var Diff	Result
LE (cities)	Yes	0.998	0.312	p = 0.214	p = 0.187	Full
HCg (cities)	Yes	0.997	0.418	p = 0.341	p = 0.265	Full
HCs (cities)	Yes	0.996	0.389	p = 0.402	p = 0.356	Full
I (cities)	Yes	0.995	0.445	p = 0.178	p = 0.298	Full

Note: p < 0.05 indicates significant mean differences. Stage 1 = configural invariance (same model specification). Stage 2 = compositional invariance (permutation-based). Stage 3 = equality of means/variances. ‘Full’ = invariance established at all three stages; ‘Partial’ = Stage 3 shows significant mean differences while compositional invariance holds.

.

MICOM results confirm that the measurement model is fully invariant across the four Russian city subgroups at all three stages, supporting the use of pooled data for the structural model. In Stage 2, all composite correlation values (c = 0.995–0.998) were well above the respective 5% quantile thresholds (range: 0.987–0.993), indicating strong compositional invariance. In Stage 3, none of the pairwise mean or variance differences reached statistical significance (all p > 0.10), confirming that composite scores carry equivalent meaning across the four cities.

4.5. Qualitative Findings

Thematic analysis of the 24 semi-structured interviews yielded four overarching themes (see Section 3.4 for the methodological rationale underlying the consolidation of five initial protocol topics into four themes). These findings complement and contextualize the quantitative SEM results.

4.5.1. Theme 1: Policy Coherence and Institutional Alignment

Russian respondents consistently noted that the national ”Digital Economy“ program provided a unifying framework. One Moscow-based university administrator observed that the program ‘gave us a clear roadmap and federal funding to build EdTech capacity.’ By contrast, Kazakhstani stakeholders described their policy landscape as fragmented, with the ”Digital Kazakhstan“ initiative perceived as broad but lacking specific operational guidance for inclusive education at the institutional level.

4.5.2. Theme 2: Teacher Readiness and Professional Development

Educators in both countries emphasized that basic digital literacy was a prerequisite for engaging with AI-based inclusive tools—a qualitative parallel to the strong HCg → HCs association found in the SEM. Russian teachers reported benefiting from structured, tiered certification programs, whereas Kazakhstani educators described training as sporadic and often reliant on imported platforms without adequate localization or language adaptation.

4.5.3. Theme 3: Infrastructure Access and Equity

Infrastructure emerged as a necessary but insufficient condition for inclusion, consistent with the small effect size of LE → HCg (f² = 0.08). Russian respondents from Yekaterinburg and Novosibirsk highlighted persistent connectivity gaps compared with Moscow and Saint Petersburg, while Kazakhstani respondents from Almaty noted a pronounced urban–rural divide in broadband availability that limits the reach of even well-designed EdTech platforms.

4.5.4. Theme 4: Ethical AI Governance

Stakeholders in both countries raised concerns about data privacy, algorithmic bias, and the absence of clear regulatory frameworks for AI use in education. Kazakhstani respondents were particularly concerned about the reliance on foreign AI systems with limited transparency regarding data handling. Russian respondents pointed to the emerging national AI standards as a positive step but noted that implementation lagged behind policy announcements.

4.6. Forecast and Trend Analysis (Russia)

To project digitalization trajectories, we estimated four separate ordinary least squares (OLS) linear regression models using Rosstat annual indicators for Russia over the period 2010–2023 (14 observations per variable). The dependent variables were: (1) broadband coverage (%), (2) AI/EdTech adoption rate (%), (3) university internet connectivity (%), and (4) ICT share of total R&D expenditure (%). Each model used year as the sole predictor. We acknowledge the simplicity of this specification; the forecasts are intended as illustrative trend projections rather than causal estimates.

Table 5. Projected Digitalization Metrics in Russia (2023–2030) with 95% Confidence Intervals.

Year	Broadband (%)	AI EdTech (%)	University Connect. (%)	ICT share of R&D (%)
2023	75.9	43.7	96.2	10.8
2025	77.1 [±1.2]	44.7 [±1.8]	97.9 [±0.6]	11.6 [±0.5]
2027	78.4 [±1.8]	45.7 [±2.4]	99.6 [±0.9]	12.4 [±0.6]
2030	80.3 [±2.6]	47.3 [±3.1]	100 [±1.3]	13.5 [±0.8]

Note: Confidence intervals (95% CI) are from OLS prediction intervals. Adj. R² for the four models: broadband = 0.94; AI EdTech = 0.88; university connectivity = 0.97; ICT share of R&D = 0.91. MAPE < 2% for all models over the estimation period. Predictor: year (single linear trend). Limitations: the models assume continuation of historical trends and do not account for potential structural breaks (e.g., sanctions, policy shifts).

presents the projected digitalization metrics in Russia for the period 2023–2030.

These trends suggest that, if current investment levels are maintained and no major structural breaks occur, Russia may sustain a comparatively strong position in digital education infrastructure among post-Soviet states. Russia’s broadband coverage is projected to increase steadily from 75.9% (2023) to 80.3% (2030), and university connectivity is expected to approach near-universal levels by 2030. The ICT share of total R&D expenditure is projected to grow from 10.8% to 13.5%, suggesting a continued commitment to digital education investment. However, these linear extrapolations should be treated with caution: they assume continuation of historical trends over 14 data points and do not account for geopolitical disruptions, policy reversals, or saturation effects.

4.7. International Contextual Benchmarking (Secondary Sources)

To situate the Russia–Kazakhstan findings within a broader international context, this subsection draws on published reports from China and India. These data are not derived from the study’s primary survey and are presented for illustrative comparison only; no causal or model-based claims are made about these countries.

China’s national digital strategy has invested approximately $1.4 trillion in digital infrastructure through 2025 (NDRC 2022), with AI in education receiving over $5 billion annually (CIE 2023). China reports 70.4% internet penetration and extensive 5G coverage across educational centers (CNNIC 2023). Its tiered approach to teacher EdTech training ($2.3 billion annually) has resulted in improved inclusive educational practices across 78% of participating institutions (Ministry of Education PRC 2023).

India’s EdTech ecosystem has grown rapidly, with over 4530 startups (NASSCOM 2024) and a 70% increase in adoption since 2020. However, India faces a significant digital skills gap—approximately 76% of its workforce lacks necessary digital competencies (NITI Aayog 2023)—and only 43% of educational institutions have consistent high-speed connectivity (TRAI 2024). EdTech funding reached $4.7 billion in 2022 (Venture Intelligence 2023).

These benchmarks underscore a recurring finding: infrastructure investment is a necessary condition for EdTech-supported inclusion, but skill development, policy coherence, and institutional support mediate the translation of infrastructure into inclusive outcomes—a pattern consistent with the SEM pathways identified in the Russia–Kazakhstan analysis.

5. Discussion

This study set out to examine the structural pathways from learning environment quality through digital competencies to inclusive learning outcomes in Russia and Kazakhstan, triangulating quantitative SEM findings with qualitative stakeholder perspectives. The results offer several contributions to the growing literature on AI-supported inclusive education.

First, the SEM confirms a staged developmental pathway: LE → HCg → HCs → I (H1–H3 supported). The strongest association in the model—between general and specialized digital competencies—has practical implications, confirmed by a significant sequential indirect effect through formal mediation analysis. It suggests that training programs should prioritize foundational digital literacy before introducing AI-specific tools. This finding aligns with Russia’s tiered certification system for educator digital competence, which has demonstrated higher adoption rates of advanced digital pedagogies among teachers who first mastered foundational skills (Ministry of Education RF 2023). Kazakhstan’s experience, where targeted professional development programs have begun to bridge the basic-to-advanced gap (Bureau of National Statistics of Kazakhstan 2023), provides further qualitative corroboration from interview data.

Second, the small effect size for LE → HCg and the modest R² for HCg (0.077) confirm that infrastructure, while necessary, is insufficient alone and that learning environment quality accounts for only a limited share of the variance in digital competencies. Interview data reinforce this: even in well-connected Russian cities, respondents emphasized that institutional culture, pedagogical support, and sustained training were more important drivers of digital skill development than hardware or bandwidth. Kazakhstani respondents echoed these concerns, noting that the ”Digital Kazakhstan“ program’s emphasis on infrastructure has not yet translated into proportional gains in teacher digital competencies. The low R² values point to specific candidate predictors that future models should incorporate. For the HCg equation, organizational support (e.g., dedicated EdTech coordinators and protected time for training), leadership commitment to digital transformation at the department or faculty level, and the extent of prior digital training exposure are likely to explain additional variance. For the I equation, pedagogical support systems (e.g., co-teaching models, mentoring programs), the availability of localized and language-appropriate AI tools, perceived ethical safety of AI platforms, and the degree of institutional alignment between stated inclusive policies and classroom-level resource allocation represent theoretically grounded candidates. Incorporating these predictors—either as direct paths or as moderators—would move the model from establishing the existence of a developmental chain to explaining why that chain operates more effectively in some institutional contexts than in others.

Third, the small effect of HCs → I (f² = 0.04, R² = 0.035 for I) reflects the inherent complexity of inclusive education. Specialized digital skills are positively associated with inclusive outcomes, but they account for only a small fraction of variance. This finding tempers techno-optimistic narratives and underscores the need for multi-level interventions that go beyond individual skill development. Inclusive outcomes are shaped by factors at multiple levels: classroom-level pedagogical practices (e.g., differentiated instruction, universal design for learning), institutional-level resource allocation and disability support services, and system-level policy frameworks governing accessibility standards. Future research should test these factors as additional predictors or moderators of the HCs → I path to better capture the multi-dimensional nature of inclusive education.

Fourth, MICOM results (H4 supported) establish that the measurement model is fully invariant across Russian city subgroups, confirming that the structural associations operate consistently across the four urban educational ecosystems studied and lending credibility to the pooled SEM results.

Fifth, the qualitative findings contribute explanatory depth. The four themes—policy coherence, teacher readiness, infrastructure access, and ethical AI governance—map closely onto the SEM constructs and illuminate the mechanisms behind the quantitative associations. The consolidation of the initial five interview topics into four themes may have reduced the visibility of implementation-specific nuances, and future qualitative work should explore this domain more granularly. The concerns about ethical AI governance, raised by stakeholders in both countries, highlight an area that requires urgent attention as AI adoption accelerates.

These findings echo broader concerns in the literature about the limitations of “technology-first” approaches in developing educational contexts, where pedagogical, infrastructural, and governance factors often lag behind technological aspirations (Mohammed and Watson 2019; Tan 2023). International benchmarks from China and India reinforce the centrality of sustained, multi-dimensional investment: China’s large-scale combined infrastructure-plus-training approach offers a reference point, while India’s experience illustrates the risks of rapid EdTech growth without corresponding gains in institutional connectivity and teacher preparedness. It is important to acknowledge the asymmetric nature of the research design: the quantitative SEM is based exclusively on Russian primary data, while Kazakhstan serves as a descriptive and qualitative benchmark. Accordingly, cross-country comparisons should be interpreted as indicative rather than as formal statistical contrasts.

Limitations

Several limitations should be noted. First, primary survey data were collected only in Russia; Kazakhstan indicators derive from secondary sources, which precludes formal SEM or measurement invariance testing between the two countries. The study’s cross-regional design is therefore asymmetric, and comparative claims are necessarily descriptive rather than inferential. Future research should administer the same survey instrument in both countries. Second, the SEM is cross-sectional and cannot establish temporal causality; all reported associations should be interpreted as correlational rather than causal. Longitudinal designs would strengthen causal inference. Third, the model’s explanatory power is limited (R² = 0.077 for HCg; R² = 0.035 for I), indicating that important predictors lie outside the model; specific candidate predictors are discussed in Section 5. Fourth, the OLS forecasts assume linear trend continuation based on 14 annual observations and may not capture structural breaks; they should be regarded as illustrative trend extrapolations. Fifth, the sample comprises university-level staff and may not generalize to K–12 settings. Sixth, the qualitative component is limited to 24 interviews; additionally, the consolidation of five initial protocol topics into four themes means that institutional adoption experiences were not examined as a standalone category. Future qualitative work should expand the sample and explore adoption processes through more fine-grained coding.

6. Conclusions

This study provides an empirically grounded, mixed-methods analysis of how AI and EdTech contribute to inclusive learning in Russia and Kazakhstan using an asymmetric cross-regional design combining primary survey data (Russia) with secondary indicators and qualitative interviews (both countries). Three main conceptual takeaways emerge from the findings. First, the SEM results establish a staged pathway from learning environment quality through general and then specialized digital competencies to inclusive learning outcomes, with the sequential mediation confirmed as statistically significant. The implication is clear: investments in digital infrastructure are necessary but yield inclusive outcomes only when accompanied by systematic, tiered development of educator digital skills—from foundational literacy to specialized AI competencies.

Second, the modest explanatory power of the model underscores that digital competencies are only one component of a much larger ecosystem that shapes inclusive education. Policy coherence, institutional culture, ethical governance of AI, and sustained professional development—themes that emerged consistently in the qualitative interviews across both countries—are essential complementary factors. Third, the descriptive comparison between Russia and Kazakhstan reveals that both countries face common challenges despite different levels of digital maturity.

Based on these findings, the study recommends that policymakers in post-Soviet Eurasian contexts prioritize: (a) tiered teacher training programs that build foundational skills before AI specialization; (b) coherent policy frameworks that align institutional incentives with inclusive outcomes; and (c) transparent ethical governance standards for AI in education. Trend projections suggest that Russia’s digital education infrastructure may continue to strengthen if current investment patterns persist, although these extrapolations are illustrative and should be interpreted with caution.

Future research should extend primary data collection to Kazakhstan and other post-Soviet states, adopt longitudinal designs to enable causal inference, incorporate additional predictors to improve explanatory power, and explore the mediating role of institutional culture in the relationship between learning environments and digital competencies.