Within the sociotechnical perspective of information science, AI is conceptualised not as a neutral tool but as a mechanism of algorithmic mediation that shapes which information becomes visible, relevant, and legitimate, as well as how it circulates within educational environments. In terms of research on algorithmic governance and algorithmic relevance, algorithmic systems do not merely recommend content but actively structure conditions of participation in public life and practices of knowledge interpretation (Beer, 2017; Gillespie, 2014).

For information literacy, this implies a shift away from an understanding focused on a discrete set of skills toward information literacy as a social practice that is formed within specific contexts through interactions among people, norms, institutional rules, and material–technological infrastructures (Lloyd, 2017). Consequently, analysing AI use in education requires an examination of configurations of practices: who uses AI and for what purposes, which rules of appropriate use are established, what mechanisms of verification and explicating limitations are employed, and how these practices affect assessment, trust, and the distribution of responsibility.

The sociotechnical approach also foregrounds the relationship between algorithms and power asymmetries. Actors who define the rules of AI use, such as administrators, educators, or platform providers. effectively determine frameworks of access to knowledge, criteria of evidential validity, and acceptable modes of producing answers. In this context, the notion of epistemic responsibility becomes salient: responsibility for the quality, provenance, and consequences of information circulation, including AI-generated content, within educational processes. As shown in related sociotechnical studies of algorithmic management, algorithmic decisions emerge through the interaction of technological systems and organisational practices, rather than existing independently of them (Orlikowski, 2007; Orlikowski & Scott, 2008; Pasquale, 2015).

AI literacy is most commonly understood as a set of competencies that enable individuals not only to use AI systems, but also to evaluate their outputs, limitations, and impacts critically. A particularly influential framework is proposed by Long and Magerko (2020), who conceptualise AI literacy as a set of design-oriented competencies that do not require programming skills, but include an understanding of AI principles, constraints, and societal implications. Subsequent review studies broaden this perspective by framing AI literacy as an integration of technical understanding, ethical considerations, critical awareness, and responsible use practices in educational contexts (Ng et al., 2021).

At the same time, a growing body of work advances the concept of critical AI literacy, emphasising that it is insufficient merely to know how to use AI. Instead, learners and educators must be able to recognise algorithmic bias, risks, power effects, and contextual consequences of AI deployment (Veldhuis, 2025). At the policy level, international organisations stress the need for ethical frameworks, transparency, data protection, and targeted professional development for educators working with generative AI (UNESCO, 2023, 2025). The OECD further highlights that, in many jurisdictions, regulation of generative AI in education remains largely advisory, with significant decision-making delegated to institutions and individual teachers, thereby amplifying the importance of local practices and institutional cultures (OECD, 2023).

Crucially for the present study, international research points to the existence of distinct modes or regimes of AI use in education. These include critically reflective regimes oriented toward ethical discussion, explicating algorithmic limitations, and fostering critical thinking, as well as instrumental-operational or control-oriented regimes focused on automation of assessment, learning analytics, and managerial or disciplinary applications (Long & Magerko, 2020; Ng et al., 2021; Veldhuis, 2025). This regime-based logic provides a conceptual foundation for moving beyond descriptions of individual tools toward the empirical typologisation of educational AI practices based on measurable indicators.

The Ukrainian discourse on AI in education is developing rapidly but remains heterogeneous in terms of thematic focus and empirical depth. At the level of academic publications, a visible strand of research examines the integration of AI in higher education, with particular attention to academic integrity, changes in assessment practices, and institutional conditions of use (Androschuk & Malyuga, 2024). Closely related to this line of inquiry are emerging studies and policy-oriented discussions addressing institutional AI use policies in educational institutions, which draw upon both Ukrainian and international experience, and highlighting the need to formalise rules and requirements for declaring AI use (AISE, 2024).

In the school sector, an important empirical reference point is a nationwide study conducted by the Junior Academy of Sciences of Ukraine in cooperation with Projector Institute, with analytical support from Factum Group Ukraine and informational support from the Ministry of Education and Science of Ukraine. This research documents the scale and patterns of AI use in general secondary education institutions and identifies barriers to schools’ readiness for systematic AI integration (JASU, 2023; Melnyk, 2024).

In parallel, surveys disseminated in the media and professional field indicate very high levels of teacher engagement with AI tools. In particular, results published by Osvitoriia Media suggest that AI tools have become a routine component of teachers’ everyday professional practices (Osvitoriia Media, 2024). Another group of Ukrainian sources focus on normative and ethical dimensions, including academic integrity, open science, responsibility for AI use, and ethical dilemmas, often presented through professional development programmes and conference proceedings (AISE, 2024).

From an applied and analytical perspective, recent reports on digital transformation and institutional readiness in Ukrainian education emphasise managerial and cultural barriers, as well as asymmetries between the rapid pace of technological adoption and institutions’ capacities to embed AI within coherent pedagogical and ethical frameworks (BDO Ukraine, 2025).

Overall, the Ukrainian context of AI integration in education is characterised by three persistent patterns: (1) uneven infrastructure and practices across regions and institutions; (2) the predominance of instrumental uses of AI, such as material preparation and automation, alongside growing concerns regarding academic integrity; and (3) a lack of comprehensive institutional policies and methodological frameworks capable of transforming AI use from individual life hacks into responsible educational practice (Androschuk & Malyuga, 2024; Melnyk, 2024).

Despite the rapid expansion of research on AI in education, the existing Ukrainian and international literature tends to follow several dominant trajectories. Most studies (a) focus on specific tools or user attitudes, (b) analyse individual risks or ethical concerns, or (c) formulate policy recommendations. Far less frequently do they offer an empirically grounded typology of AI use models understood as coherent configurations of practices that can be reproduced, compared, and applied for diagnosing institutions or groups of educators, particularly in high-risk informational contexts (Ng et al., 2021; Melnyk, 2024).

To address these gaps, this article proposes a sociotechnical research design that combines a conceptualisation of AI as an algorithmic information mediator with a quantitative operationalisation of educational information practices through indices of critically reflective AI use and instrumental–operational (control–automation) AI use. Based on a survey of educators from frontline regions of Ukraine, the study constructed a 2×2 typology of AI use models that enabled an empirical distinction between critically reflective and instrumental-operational configurations of practice and supported their interpretation in the context of heightened information risks associated with war.

In the sample (N = 208), the mean value of CR_index (critically reflective practices) was M = 3.64, SD = 0.86, while Control_index (instrumental-operational and control-automation practices) averages were M = 2.95, SD = 0.86 (scale 1-5). Median thresholds used for the typology were 3.83 for CR_index and 3.00 for Control_index. A positive association between CR_index and Control_index was observed (r ≈ 0.61; see Table A2), indicating that educators who used AI more intensively in a critically reflective manner also tended to employ it for operational tasks. As a robustness check, Spearman rank correlations were additionally calculated; the results were substantively consistent with the Pearson correlations reported in the manuscript. At the same time, the distribution revealed distinct poles—high CR with low Control and low CR with relatively high Control—supporting the analytical relevance of typologisation. Correlation analysis further indicated heterogeneity within the operational dimension: CR_index was more strongly related to Efficiency/Analytics (r ≈ 0.77) than to Assess Control (r ≈ 0.36), suggesting that efficiency-oriented uses more often coexisted with critically reflective practices than assessment- and control-oriented logics. Overall, the observed structure suggested that AI use practices were not dichotomous, but formed overlapping configurations, supporting the analytical relevance of typological modelling.

The distribution of respondents across the four models is presented in Table 2. The hybrid model (CR↑ /Control↑) was dominant, accounting for 36.1% of the sample (n = 75). This indicated that the most prevalent configuration combined critically reflective practices with intensive instrumental-operational uses of AI. The second most common pattern was the low-integration model (CR↓/Control↓), representing 32.2% of respondents (n = 67), pointing to a substantial segment of educators with limited AI engagement across both dimensions. The instrumental–control model (CR↓ /Control↑) comprised 17.3% of the sample (n = 36), while the critically–reflective model (CR↑/Control↓) was the least prevalent, accounting for 14.4% (n = 30).

Comparative profiles of the four models are reported in Table 3. The critically-reflective model (CR↑/Control↓) combined a high CR_index (4.27) with a relatively low Control_index (2.41), with efficiency-oriented uses prevailing over assessment-control (Assess Control = 1.70). The hybrid model (CR↑/Control↓) showed the highest levels on both dimensions (CR_index = 4.34; Control_index = 3.72), with the highest values for both Assess Control (3.33) and Efficiency/Analytics (4.30), representing a maximally integrated regime. The instrumentalcontrol model (CR↓/Control↑) was characterised by a lower CR_index (3.26) and a relatively high Control_index (3.31), with elevated Assess Control (3.19) and moderately high Efficiency/Analytics (3.49). The low-integration model (CR↓/Control↓) exhibited the lowest values across both dimensions (CR_index = 2.76; Control_index = 2.15), indicating minimal AI engagement in educational information practices.

Item-level results (Table B1, Appendix B) indicated that differences between models were primarily associated with ethical discussion, verification practices, and the use of control-oriented applications.

K-means clustering (k = 5) on standardised CR_index and Control_index values indicated that the data contained not only four broad quadrants, but also more fine-grained practice profiles (Table A4). Two hybrid clusters emerged: one with moderately high Control, and another with very high Control, alongside clusters characterised by very low integration and by moderate CR combined with low Control. Mapping these clusters onto the 2×2 typology (Table A5) showed that one cluster aligned entirely with the CR↓/Control↓ quadrant (low integration, 100%), while the high-control hybrid cluster fell almost exclusively within CR↑/Control↑ (≈87%). The remaining clusters spanned adjacent quadrants, indicating transitional configurations. Overall, the results confirmed the 2×2 typology as a parsimonious representation, while demonstrating that AI use regimes formed a continuum with intermediate modes.

χ2-tests examining associations between the 2×2 models and respondent characteristics (type of institution, ownership form, position, and years of experience) revealed no statistically significant relationships (all p > 0.05; see Table B2). Cramer’s V values predominantly fall within the range of weak effects (≈0.11–0.27). For variables with a large number of categories, particularly institution type and position, some cells exhibited very low expected frequencies. Accordingly, χ² results for these tables are interpreted with caution, relying primarily on effect size estimates (Cramer’s V) and diagnostics of expected cell counts (Table B2). To ensure robustness, Fisher’s exact test was additionally applied in cases with small, expected cell frequencies, yielding results consistent with the χ² tests reported.

The empirical findings allow us to address the research questions formulated in the introduction.

RQ1 was addressed through the identification of four empirically distinguishable models of AI use in educational information practices: critically reflective, hybrid, instrumental–control, and low-integration.

RQ2 was addressed by demonstrating that critically reflective practices and instrumental–operational uses of AI frequently coexist, particularly in the hybrid model, which represented the dominant configuration in the sample. At the same time, control-oriented uses showed weaker alignment with critically reflective practices than efficiency-oriented applications.

RQ3 was addressed through the sociotechnical interpretation of these models in the context of wartime informational risk, where critically reflective practices functioned as mechanisms of epistemic security and responsible engagement with AI-generated information.

The 2×2 typology (CR × Control) demonstrated that AI integration in educational information practices cannot be reduced to a simple for or against position. Instead, the hybrid model (CR↑/Control↑) predominated, indicating that critically reflective practices frequently coexisted with logics of operational efficiency, automation, and, to some extent, control. From a sociotechnical perspective, this outcome suggested that the same technology became embedded in different regimes of organising information work, functioning simultaneously as a learning support tool and as an instrument of administrative ordering.

The data further showed that operationality is not monolithic. Efficiency- and analytics-oriented uses of AI more often aligned with critically reflective practices, whereas assessment- and control-oriented sub-logics exhibited weaker alignment with the CR dimension. Thus, sociotechnical tension arose not between criticality and operationality per se, but primarily between critically reflective practices and AI applications oriented toward control, standardisation, and performance-based governance. The typology, therefore, captured a competition between two institutional logics: a) a pedagogical logic, centred on meaning-making and critical engagement, and b) a technocratic logic, oriented toward productivity, manageability, and control.

This interpretation aligned with STS accounts of algorithmic mediation and sociomateriality, which emphasise that algorithms actively reconfigure regimes of visibility, authority, and accountability, rather than merely supporting existing practices (Gillespie, 2014; Orlikowski & Scott, 2008). From this perspective, the expansion of control-oriented AI applications exemplifies the diffusion of responsibility within algorithmic systems, where governance increasingly operates through metrics and opaque decision infrastructures (Pasquale, 2015).

Finally, clustering results (k = 5) indicated that real-world practices are often transitional. Profiles of moderate and high-control hybridity suggested that sociotechnical regimes are not binary, but assembled from multiple practice elements, with institutional rules and professional expectations, shifting the balance toward either critical engagement or operational–control-oriented reduction.

The empirical results provided partial support for the hypotheses formulated in the study. Hypothesis H1 (Multiplicity of models) was supported by the identification of four distinct AI use configurations in the data. Hypothesis H2 (Sociotechnical tension) was supported insofar as control-oriented uses are less strongly associated with critically reflective practices than efficiency-oriented uses. Hypothesis H3 was supported conceptually, rather than directly tested empirically, highlighting the importance of critically reflective practices in high-risk informational contexts. Hypothesis H4 remained interpretive in nature and indicated a direction for future research, focusing on the role of educators as algorithmic information mediators. Thus, the proposed typology not only described empirical patterns, but also provided an analytical lens for understanding how algorithmic mediation restructures educational information practices under conditions of uncertainty.

The context of war foregrounded not only the efficiency of AI use but also epistemic security—the capacity of educational environments to distinguish reliable information from manipulative content, withstand information attacks, work with uncertainty, and sustain critical practices. Under such conditions, critically-reflective practices — ethical discussion, explicating limitations, verification of outputs, and the cultivation of critical thinking — acquired not merely a pedagogical but also a protective societal function, reducing the risk of uncritical reproduction of AI-generated errors, hallucinations, or manipulative narratives.

Accordingly, the role of the educator in AI-mediated environments can be conceptualised as that of an algorithmic information mediator: an actor who establishes rules for interacting with AI-generated information, defines verification criteria, moderates boundaries of acceptable use, and fosters a culture of explainability within existing constraints. From this perspective, the hybrid model (the most prevalent in the data) appeared inherently ambivalent. The hybrid model may represent a productive configuration in which operational efficiency does not displace critical engagement, but it may also constitute a potential risk point when increasing operational dependence on AI reinforces tendencies toward standardisation and control, thereby narrowing the space for critical information work. The wartime context intensified this ambivalence since the cost of fast but unchecked responses was substantially higher than in stable information environments.

From the perspective of critical information literacy, this mediating role extended beyond technical competence to include responsibility for shaping how authority, credibility, and trust are negotiated in AI-mediated learning environments (Drabinski, 2019; Elmborg, 2006). In wartime contexts, such mediation becomes a core dimension of information literacy itself, reframing it as a practice of epistemic risk management, rather than merely efficient information use.

The practical value of the proposed typology lies in its potential use as a diagnostic tool for educational institutions and governance bodies, enabling them to identify dominant AI use regimes and to assess which competencies require reinforcement. At the level of institutional policy (for schools and universities, the Ministry of Education, and professional development programmes), the findings supported three main directions for action.

This study had several limitations. While the indices demonstrated satisfactory internal consistency and their empirical structure was additionally examined using exploratory factor analysis, the current study did not employ confirmatory factor analysis or latent-variable modelling. The design, therefore, prioritised reliability, interpretability, and typological usefulness over full psychometric validation. Future research should test the stability of the factor structure in larger samples and across different educational contexts. In addition, while the theoretical framework emphasised wartime and disinformation risks, the current design lacked a dedicated risk-sensitivity index that would allow direct testing of links between disinformation sensitivity and AI use models. Future research should develop and validate such a scale, incorporating behavioural indicators (e.g., fact-checking practices, detection of manipulation, source-evaluation rules, and responses to suspicious content), or complement the survey with qualitative methods, such as interviews or practice diaries.

The self-reported nature of the data may have introduced social desirability bias, particularly regarding ethical and critically reflective practices. Furthermore, the cross-sectional design precluded causal inference, capturing the coexistence of practice regimes without explaining transitions between them. Finally, the sample was drawn from frontline regions, which enhanced relevance for high-risk contexts but warrants caution in generalising to other settings.

Despite these limitations, the index-based operationalisation and typologisation of AI use models provided a robust basis for comparative and longitudinal research. Future studies could examine how institutional policies, professional development, and contextual change shape shifts in AI use regimes toward either greater critical engagement or increased control-oriented reduction.

Although grounded in a wartime Ukrainian context, the proposed CR × Control typology is transferable to other high-uncertainty information environments. Such as post-pandemic education, data-driven school systems, or AI-mediated governance settings, where educators similarly navigate tensions between critical engagement and operational automation. Future research may also employ latent profile analysis or longitudinal designs to examine transitions between AI use regimes over time.