The role of Artificial Intelligence in amplifying educational inequalities in resource constrained mathematics classrooms

The role of Artificial Intelligence in amplifying educational inequalities in resource constrained mathematics classrooms | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article The role of Artificial Intelligence in amplifying educational inequalities in resource constrained mathematics classrooms Brantina Chirinda This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8679398/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Artificial intelligence (AI) is increasingly positioned as a transformative force in education, with particular promise for addressing persistent challenges in mathematics education, such as low achievement, limited instructional capacity, and unequal access to quality teaching. AI-driven systems—including adaptive learning platforms, intelligent tutoring systems, and automated assessment tools—are widely promoted as mechanisms for personalizing instruction at scale and advancing educational equity. This study critically examines these claims by investigating how AI-mediated mathematics education is enacted in South Africa’s schools located in contexts of socio-economic disadvantage. Drawing on qualitative interviews with mathematics teachers working across diverse institutional settings, the study explores how AI tools are selected, implemented, and experienced in everyday classroom practice. The findings reveal that, rather than functioning as neutral or equalizing technologies, current AI systems often amplify existing inequalities by disproportionately benefiting students who already possess strong self-regulatory skills, institutional support, and prior mathematical confidence. Personalization without sustained pedagogical and relational support often leads to observed student isolation, while automation prioritizes efficiency and monitoring over conceptual understanding. By foregrounding mathematics teachers’ perspectives, the study demonstrates that AI transforms educational practice primarily through the automation of pedagogical decision-making, the redistribution of learning responsibility, and the reconfiguration of what counts as legitimate mathematical activity. It argues that meaningful educational transformation through AI requires treating equity as a foundational design and governance condition, with efficiency serving as an instrumental outcome rather than a proxy for pedagogical improvement. Artificial Intelligence in Education (AIED) Algorithmic Equity Teacher Agency Educational Data Mining Mathematics Education Global South 1. Introduction AI is increasingly positioned as a transformative force in education, with growing influence over how teaching, learning, and assessment are organized. Across policy documents, commercial platforms, and academic research, AI-driven systems are promoted as solutions to longstanding educational challenges, including learner heterogeneity, teacher workload, and persistent achievement gaps. Crucially, this study adopts a functional rather than a strictly technical definition of Artificial Intelligence. While computer science distinguishes between dynamic machine learning algorithms and static, rule-based branching (often found in older educational software), this distinction is frequently invisible to the end-user. For a student isolated behind a screen or a teacher managing a data dashboard, the phenomenology of automation remains the same: an external system has assumed authority over the pacing, sequencing, and assessment of knowledge. Whether the underlying code is a neural network or a complex decision tree, the resulting pedagogical dynamic is one of ‘infrastructural control’ (Williamson, 2017 ). Therefore, the sanalyze these technologies not by their computational complexity, but by their capacity to automate instructional decisions and govern classroom activity. Proponents of educational technology frequently position algorithmic solutions—including adaptive courseware and automated assessment—as essential infrastructure for achieving equitable and efficient mathematics education (Holmes et al., 2019 ; Luckin et al., 2016 ). These claims resonate strongly within contemporary reform agendas that emphasize data-driven decision-making, scalability, and accountability. Within this discourse, personalization is often treated as both a technical and a moral achievement: by tailoring instruction to individual learners, AI systems are assumed to better support diverse needs than traditional classroom practices. As a result, AI is increasingly presented not merely as an instructional aid but as a catalyst for transforming education systems toward greater inclusivity and effectiveness. Yet alongside this optimism, a growing body of scholarship cautions that the educational effects of AI cannot be understood independently of the social, institutional, and pedagogical contexts in which these technologies are deployed. Research in critical AI studies and sociology of education demonstrates that technological systems are never neutral; they encode assumptions about learners, learning, and success, and tend to reflect the priorities of the institutions and markets that produce them (Selwyn, 2016 ; Williamson, 2017 ; Benjamin, 2019 ). In education, these assumptions often align with dominant norms of self-regulation, autonomy, and measurable performance—norms that are unevenly distributed across socio-economic, linguistic, and cultural contexts. Mathematics education provides a particularly salient site for examining these tensions. Long recognized as a gatekeeping subject, mathematics plays a central role in shaping access to advanced study, STEM careers, and broader social and economic opportunities (Boaler, 2016 ). Despite decades of reform, patterns of mathematical achievement remain persistently stratified along socio-economic and racial lines, with students in disadvantaged contexts more likely to encounter procedural, remedial, and test-oriented instruction (Nasir et al., 2014 ). These structural inequities form the backdrop against which AI-driven mathematics education is currently being introduced. Within policy and industry narratives, AI is frequently framed as uniquely capable of addressing these challenges. Adaptive systems promise to identify knowledge gaps, adjust pacing, and provide immediate feedback, thereby compensating for limited instructional capacity and enabling individualized learning pathways. Such claims are especially compelling in under-resourced schools, where large class sizes, teacher shortages, and accountability pressures intensify the appeal of scalable technological interventions. However, emerging research raises critical questions about whether AI-driven personalization functions equitably in practice. Studies suggest that AI-based learning environments often work most effectively for students who already possess strong self-regulatory skills, prior academic knowledge, and familiarity with dominant learning norms (Azevedo et al., 2018 ; Kizilcec et al., 2017 ). Conversely, students who face linguistic barriers, limited academic support, or fragmented access to technology may experience AI systems as confusing, isolating, or demotivating. Rather than compensating for inequality, personalization may redistribute responsibility for learning in ways that are socially patterned and uneven. These concerns point to a broader issue in AI-in-education research: the tendency to evaluate AI systems primarily in terms of efficiency, predictive accuracy, or short-term learning gains, while paying less attention to how these systems reshape pedagogical relationships, learner identities, and opportunities to learn. In mathematics education, this often manifests through an emphasis on what is easily measurable—speed, accuracy, and task completion—rather than on reasoning, sense-making, and collaborative problem-solving. Such emphases risk reinforcing pedagogical approaches that have historically marginalized students in disadvantaged contexts. Crucially, the impacts of AI in education are mediated through teachers, who operate at the intersection of technology, curriculum, and institutional constraint. Teachers make day-to-day decisions about how AI tools are used, whom they support, and what forms of learning are prioritized. Yet teachers’ perspectives—particularly those working in disadvantaged contexts—remain underrepresented in research on educational AI, which often foregrounds system design or learner analytics over lived classroom practice. This study addresses this gap by examining how AI-driven mathematics education is experienced and enacted in contexts of socio-economic disadvantage in South Africa, drawing on qualitative interviews with mathematics teachers across diverse institutional settings. Given the Global South's distinct socio-economic disparities and resource constraints, this context offers a critical vantage point for interrogating the often-universalized claims of efficacy attached to AI technologies developed in the Global North. Rather than asking whether AI can personalize learning in principle, the study investigates how personalization, automation, and data-driven decision-making operate in practice and what consequences they have for educational equity. By foregrounding mathematics teachers’ perspectives, the study contributes to ongoing debates about the opportunities and challenges of AI in education in three ways. First, it challenges narratives of technological neutrality by showing how AI systems often align with and amplify existing inequalities. Second, it highlights the central role of mathematics teachers as constrained mediators of AI, whose professional judgment is shaped by accountability pressures, resource limitations, and institutional expectations. Third, it reframes equity not as an automatic outcome of AI adoption, but as a design, governance, and pedagogical challenge that must be addressed explicitly. In doing so, the study speaks directly to the aims of this special issue by offering a critical, empirically grounded account of how AI is transforming education—not only through new technical capabilities, but through its interaction with longstanding social and institutional structures. The guiding research question is: How does the introduction of AI-driven mathematics education reshape teaching and learning in contexts of disadvantage, and what challenges does this transformation pose for educational equity from the perspective of teachers? 2. Literature Review 2.1. AI-Driven Transformation in Mathematics Education AI has emerged as a central component of contemporary educational reform agendas, with increasing influence over curriculum delivery, assessment, and learner support. In mathematics education, AI-driven systems—such as adaptive learning platforms, intelligent tutoring systems, automated feedback tools, and predictive analytics—are widely promoted for their capacity to scale instruction, offering a mechanism to deliver standardized mathematics content to large student populations without the resource constraints associated with traditional human-led teaching (Holmes et al., 2019 ). These systems are often positioned as responses to persistent challenges, including heterogeneous classrooms, teacher workload, and unequal access to high-quality instruction. In policy and commercial discourse, AI-enabled personalization is frequently framed as a means of improving both effectiveness and equity. By tailoring instruction to individual learners’ performance data, AI systems are assumed to provide more precise and responsive support than traditional classroom practices (Luckin et al., 2016 ). This promise is particularly salient in under-resourced educational contexts, where structural constraints heighten the appeal of automated and scalable solutions. However, existing research suggests that claims of transformation must be examined in relation to the historical and institutional conditions of mathematics education. Longstanding patterns of inequality—shaped by tracking, assessment regimes, and deficit-oriented conceptions of ability—continue to structure who benefits from instructional innovations (Boaler, 2016 ). As such, AI enters mathematics classrooms not as a neutral intervention, but as a sociotechnical system embedded within already stratified educational environments. 2.2. Personalization, Adaptivity, and Assumptions About Learning Central to AI-in-education discourse is the concept of personalization. Adaptive systems aim to adjust content, pacing, and feedback based on learners’ interactions, often using performance metrics such as accuracy, speed, and persistence. Within mathematics education, personalization is frequently equated with responsiveness to learner diversity and, by extension, with equity (Luckin et al., 2016 ; Holmes et al., 2019 ). Yet critics argue that such approaches rely on narrow and implicit assumptions about learning. AI systems typically operationalize learning as individual progression through predefined tasks, privileging self-regulation, metacognitive awareness, and independent navigation of learning pathways (Williamson, 2017 ; Knox, 2020 ). These assumptions align more closely with the dispositions of students who already possess academic confidence, prior knowledge, and familiarity with dominant schooling norms. Empirical research supports this concern. Studies of adaptive and AI-supported learning environments consistently find that students with stronger prior achievement and learning strategies benefit more from personalization, while those who struggle academically are more likely to disengage or experience limited gains (Azevedo et al., 2018 ; Kizilcec et al., 2017 ). In mathematics education, where conceptual gaps can quickly compound, personalization may therefore stabilize existing hierarchies rather than disrupt them. 2.3. Algorithmic Systems, Datafication, and Educational Inequality A growing body of critical AI scholarship emphasizes that algorithmic systems are shaped by the data, values, and institutional priorities embedded in their design (Benjamin, 2019 ; Eubanks, 2018 ). In education, AI systems rely on data that reflect historical inequalities in access, opportunity, and support, raising concerns about the reproduction of deficit-based classifications through automated decision-making. Predictive analytics and adaptive pathways can label students as “at risk,” “behind,” or “low performing” based on patterns that are socially patterned rather than purely cognitive (O’Neil, 2016 ; Williamson & Eynon, 2020 ). In mathematics education, such classifications are particularly consequential, as early labeling can narrow curricular exposure and shape long-term trajectories. Research on digital inequality further underscores that educational outcomes are shaped less by access to technology than by differences in use, support, and institutional mediation (van Dijk, 2020 ). Without deliberate attention to equity, AI systems may function as mechanisms for automating existing inequalities under the guise of objectivity and precision. 2.4. Personalization, Pedagogical Mediation, and Student Isolation While AI-driven personalization is often promoted as learner-centered, critics note that it frequently coincides with a reduction in human interaction. In practice, personalization may involve students working independently on screens for extended periods, with limited opportunities for dialogue, explanation, or collaborative problem-solving (Selwyn et al., 2020 ). Research in mathematics education consistently highlights the importance of relational pedagogy—teacher questioning, peer interaction, and collective sense-making—for developing conceptual understanding and mathematical identity (Sfard, 2008 ; Cobb et al., 2009 ). When AI systems replace rather than supplement these interactions, students who require the most support may experience learning as isolating or confusing. Studies indicate that students facing academic difficulty are more likely to perceive automated feedback as opaque or demotivating, particularly when it lacks explanatory depth or opportunities for clarification (Eubanks, 2018 ). In disadvantaged contexts, where instructional support is already stretched, personalization without pedagogical mediation may intensify disengagement rather than foster inclusion. 2.5. Automation, Efficiency, and the Limits of Measurable Learning Another dominant justification for AI in education is efficiency. Automated grading, progress monitoring, and analytics are promoted as ways to reduce teacher workload and enhance accountability (Luckin et al., 2016 ). While such efficiencies can streamline administrative processes, scholars caution that they also reshape what counts as legitimate learning. Automation tends to privilege what is easily measurable, narrowing instructional focus toward short-answer tasks, procedural fluency, and alignment with standardized assessments (Biesta, 2015 ; Au, 2016 ). In mathematics education, this emphasis risks marginalizing reasoning, explanation, and exploratory problem-solving—practices shown to be particularly important for equitable learning. Predictive and monitoring systems may further contribute to rigid classifications of ability, especially when labels such as “at risk” become institutionalized within schools (O’Neil, 2016 ; Benjamin, 2019 ). In high-pressure environments, efficiency-driven AI use may therefore support institutional management more than student understanding. 2.6. Teachers as Human-in-the-Loop Mediators of AI Systems Increasingly, AI-in-education research recognizes that the effects of AI systems are mediated through teachers, who interpret, adapt, and contextualize technologies in practice. Teachers function as “human-in-the-loop” actors whose professional judgment shapes how AI is enacted in classrooms (Priestley et al., 2015 ). However, teacher agency is unevenly distributed. In disadvantaged contexts, teachers often operate under conditions of limited resources, prescriptive curricula, and high-stakes accountability, constraining their ability to integrate AI in equity-oriented ways (Darling-Hammond, 2010 ). Building on this literature, this study foregrounds mathematics teachers’ perspectives from contexts of disadvantage to examine how AI is enacted in everyday practice. By explicitly linking AI-driven personalization, automation, and teacher mediation to questions of equity, the study contributes a critical, empirically grounded account of how AI is transforming mathematics education—and for whom. 3. Theoretical Framework While much of the literature frames AI in education through the lens of teacher acceptance or efficacy, this study adopts a critical sociotechnical perspective focused on pedagogical automation and labor redistribution. The study draws on the work of Selwyn (2019) and Williamson (2021) not merely to argue that technology is social, but to examine how AI-driven systems restructure the essential dynamics of teaching and learning. Central to this framework is the concept of redistributing learning responsibility. Rather than functioning simply as neutral aids, adaptive platforms frequently offload the labor of pacing, remediation, and engagement management from the teacher to the student-technology dyad. As noted later in the analysis, this shift presupposes a level of self-regulation and digital literacy that is unequally distributed among learners. Therefore, AI was aligned not as a tool that teachers use, but as an actor that automates pedagogical decision-making. This lens shows how the efficiency promised by AI often comes at the cost of deepening structural inequalities, as the burden of navigating these automated systems falls disproportionately on students in under-resourced environments who lack the scaffolding to manage this new responsibility. 3.1. AI as a Socio-technical System in Education The study approaches AI not as a neutral pedagogical aid, but as a political artifact that creates new asymmetries of power in the classroom. From this perspective, algorithmic platforms do not simply support learning; they operationalize specific ideologies about knowledge and ability, effectively deciding whose ways of thinking are valued and whose are marginalized. AI systems are designed, trained, procured, and implemented within policy regimes that prioritize scalability, efficiency, accountability, and datafication (Williamson, 2017 ). These priorities shape how learning is modeled, what forms of knowledge are valued, and how success is defined within AI-driven environments. From this perspective, the effects of AI cannot be attributed solely to algorithmic capability. Instead, they emerge through interactions between technical design, institutional constraints, and human actors. In mathematics education, AI systems often encode assumptions about learning as individualized, measurable, and self-directed—assumptions that align unevenly with the realities of disadvantaged classrooms. Conceptualizing AI as sociotechnical enables the analysis to move beyond questions of effectiveness toward questions of power, governance, and equity. From this perspective, the educational significance of AI lies less in whether systems employ advanced machine learning, and more in how algorithmic automation, classification, and datafication reorganize pedagogical authority and learning opportunity. 3.2. Educational Inequality and the Redistribution of Learning Responsibility The framework draws on sociological theories of educational inequality, particularly those that emphasize how schooling redistributes responsibility for success and failure in socially patterned ways. Mathematics education has long been identified as a site of social reproduction, where access to valued forms of knowledge is shaped by socio-economic status, language, race, and institutional sorting mechanisms (Bourdieu & Passeron, 1990 ). The study posits that when the responsibility for learning is delegated to an algorithm, the human elements of teaching—care, motivation, and the diagnosis of root causes of error—are often stripped away, leaving vulnerable students to face a deficit based solely on data metrics. AI-driven personalization plays a key role in this redistribution. By emphasizing self-paced progression, independent problem-solving, and continuous performance monitoring, AI systems shift responsibility for learning onto individual students. This shift advantages learners who possess forms of cultural capital aligned with dominant academic norms—such as self-regulation, confidence, and familiarity with abstract mathematical representations—while disadvantaging those who require relational, dialogic, or scaffolded support. This component of the framework directly informs the interpretation of the finding that AI works best for students who already know how to learn, positioning this outcome as a structural effect of personalization rather than a deficit of individual learners. By automating feedback and pacing, these systems risk mechanizing the Matthew Effect in education—where those with prior advantages (connectivity, high self-regulation) gain the most, while those relying on the school for structure are further marginalized. 3.3. Critical Mathematics Education and the Nature of Mathematical Learning To examine how AI reshapes what counts as legitimate mathematics learning, the framework draws on critical mathematics education. This tradition emphasizes that mathematics education is not neutral or purely technical, but involves normative decisions about which forms of reasoning, participation, and knowledge are valued (Skovsmose, 2011 ; Frankenstein, 2012). AI-driven systems frequently prioritize procedural correctness, efficiency, and task completion, aligning with what critical scholars describe as exercise paradigms of learning. Such paradigms limit opportunities for explanation, dialogue, and collective sense-making—practices shown to be essential for developing conceptual understanding and mathematical identity, particularly for students in disadvantaged contexts. Within this framework, the finding that automation increases efficiency, not understanding, is interpreted as a consequence of how AI systems operationalize learning, rather than as an unintended side effect. The framework thus highlights tensions between AI-supported efficiency and equity-oriented pedagogy. 3.4. Algorithmic Power, Classification, and Datafication The framework also incorporates insights from critical data and algorithm studies, which examine how classification, prediction, and automation shape social outcomes. Educational AI systems rely on historical data that reflect existing inequalities in access, opportunity, and support (O’Neil, 2016 ; Benjamin, 2019 ). As a result, adaptive pathways and predictive analytics may reproduce deficit-based narratives about students, particularly those from marginalized backgrounds. In mathematics education, algorithmic labeling—such as identifying students as “at risk” or “low performing”—can narrow learning opportunities and shape expectations in ways that become self-reinforcing. Datafication further privileges what is easily measurable, reinforcing curricular narrowing and accountability-driven pedagogies. This aspect of the framework informs the analysis of how automation and monitoring reshape teaching and learning, contributing to all three findings themes by explaining how inequality becomes embedded in algorithmic processes. 3.5. Teachers as Constrained Human-in-the-Loop Mediators A final and central component of the framework positions teachers as human-in-the-loop mediators between AI systems and students. Teachers interpret, adapt, and sometimes resist AI tools based on their professional judgment, pedagogical commitments, and contextual knowledge (Priestley et al., 2015 ). However, their capacity to do so is shaped by institutional conditions, including curriculum mandates, assessment regimes, resource constraints, and professional development opportunities. In disadvantaged contexts, teachers often face heightened accountability pressures and limited autonomy, constraining their ability to integrate AI in equity-oriented ways. This framework rejects deficit explanations that locate responsibility for inequitable outcomes in teachers’ practices, instead emphasizing how upstream design and governance decisions shape what teachers can do in practice. This perspective is essential for understanding why personalization without support can lead to isolation, as teachers’ ability to provide relational mediation is frequently undermined by the very systems designed to increase efficiency. 3.6. Integrative Role of the Framework Together, these theoretical perspectives provide a coherent lens for analyzing how AI-driven mathematics education operates within unequal educational systems. The framework links mathematics teachers’ lived experiences to broader sociotechnical dynamics, enabling the study to move beyond surface-level descriptions of AI use toward a critical examination of how AI redistributes responsibility, reshapes pedagogy, and reconfigures equity. By aligning theories of sociotechnical systems, educational inequality, critical mathematics education, and algorithmic power, the framework supports a nuanced analysis of AI as both an opportunity and a risk. It also provides a foundation for rethinking how AI in education might be designed and governed to support—not undermine—equitable mathematics learning. 4. Methods 4.1. Research Design This study employed a qualitative research design to examine how AI–driven mathematics education is enacted in contexts of socio-economic disadvantage. Qualitative methods were selected to capture mathematics teachers’ situated experiences, professional judgments, and interpretations of AI systems in everyday classroom practice—dimensions that are often obscured in system-centered or outcome-focused evaluations of educational AI. Guided by a critical sociotechnical framework, the study conceptualized AI not as a neutral instructional tool, but as a system embedded within institutional constraints, pedagogical traditions, and historical patterns of educational inequality. This framework informed all stages of the research process, including participant selection, interview design, and analytic strategy. While this study centers on teachers' perspectives rather than direct student data, mathematics teachers were positioned as critical expert witnesses in the learning process. Consequently, references to student 'confusion' or 'isolation' in the findings reflect the teachers' professional diagnosis of student affect based on longitudinal observation. Unlike platform analytics—which capture only click-rates, time-on-task, and accuracy—teachers possess the longitudinal and ecological context necessary to interpret why a student pauses or disengages. Teachers observe the affective and behavioral correlates of data points (e.g., the difference between a student pausing to think versus pausing due to disengagement). Consequently, this analysis treats teacher observation not merely as opinion, but as a form of situated professional judgment that illuminates the classroom dynamics often invisible to the algorithmic gaze. 4.2. Participants and Context Participants were 26 mathematics teachers working in South Africa’s plublic schools characterized by socio-economic disadvantage, ensuring the analysis captured the realities of AI implementation outside of well-resourced, elite private institutions. South African school quintiles (Q1-Q5) classify public schools by community poverty, with Q1 being the poorest (non-fee paying) and Q5 the least poor (fee-paying). Mathematics teachers were recruited using purposive sampling to ensure variation across school levels, institutional contexts, and degrees of exposure to AI-driven instructional tools. All participants had direct experience using AI-based platforms for mathematics instruction, including adaptive learning systems, automated assessment tools, or data analytics dashboards. The platforms ranged from gamified interfaces requiring rapid-fire input to dashboard-centric systems presenting static problem sets. While the commercial platforms utilized by participants were broadly marketed as AI, the study’s analysis distinguishes between the specific forms of algorithmic agency encountered in practice. The majority of tools relied primarily on rule-based branching and automation; these operated effectively as digital worksheets that routed students based on static if-then performance thresholds rather than dynamic machine-learning models. A smaller subset of systems employed predictive analytics, using opaque algorithms to generate risk labels or individualized pathways without transparency to users. This technical distinction is critical, as the study’s findings reveal that the limitations of the former (e.g., frustration with repetitive loops) and the opacity of the latter (e.g., fatalistic labeling) contributed distinctively to the inequities observed in the classroom. This distinction is not merely technical, but analytically central to the study. The transformative effects examined here do not arise primarily from the sophistication of machine learning models, but rather from algorithmic governance mechanisms embedded in these systems—namely, automated instructional sequencing, data-driven learner classification, and continuous performance monitoring. Even relatively simple rule-based systems exert substantial sociotechnical power when they structure access to content, regulate pacing, and shape how teachers and institutions interpret student ability. Consequently, this study treats AI not as a question of computational complexity, but as a form of infrastructural control over pedagogical decision-making. Of the 26 interviews, 17 were conducted virtually and 9 face-to-face. This mixed interview modality reflected both participants' availability and contextual constraints, enabling the inclusion of mathematics teachers across geographically dispersed settings. No substantive differences were observed in the depth or focus of responses across interview modes. 4.3. Data Collection Data were collected through semi-structured interviews designed to elicit mathematics teachers’ perspectives on the selection, implementation, and pedagogical consequences of AI-driven mathematics tools. Interview questions were explicitly informed by the study’s theoretical framework and treated its core components as sensitizing concepts rather than prescriptive categories. Interview protocols explored five interrelated domains: Institutional context and AI adoption , including how and why AI tools were introduced and the problems they were expected to address. Pedagogical use of AI , focusing on how AI systems shaped instructional practices and mathematical content. Personalization and differentiation , with particular attention to which students benefited most or least from AI-mediated instruction. Automation, data, and classification , including experiences with dashboards, analytics, and labels such as “at risk.” Teacher agency and mediation , examining mathematics teachers’ capacity to adapt, resist, or supplement AI tools in practice. Interviews typically lasted 45–75 minutes and were audio-recorded with participants’ consent. All interviews were transcribed verbatim prior to analysis. 4.4. Data Analysis Data analysis followed a theoretically informed thematic analysis, where the study’s critical sociotechnical framework explicitly guided coding and theme development. Rather than treating theory as a post hoc interpretive lens, the framework provided sensitizing concepts that oriented analytic attention toward specific dimensions of AI-mediated practice, including the redistribution of learning responsibility, forms of pedagogical mediation, and the role of algorithmic systems in shaping opportunity. The analytic process proceeded in three iterative stages: First-Order Coding (Initial Coding) Analysis began with an inductive focus on teachers’ concrete descriptions of AI use, such as specific classroom practices, student behavioral responses (e.g.,“students work alone,” “system flags them as behind”), and institutional expectations. Focused Coding (Second-Order Categories) Initial codes were then clustered around higher-order categories aligned with the theoretical framework. For example, concrete observations of student autonomy were grouped under Redistribution of Learning Responsibility , while descriptions of dashboards were analyzed through the lens of Datafication and Classification . Thematic Synthesis : Finally, these categories were integrated into three overarching themes that reflect broader socio-technical dynamics: (a) AI works best for students who already know how to learn ; (b) Personalization without support becomes isolation ; and (c) Automation increases efficiency, not understanding . These themes were not treated as mutually exclusive but as interrelated manifestations of how AI systems reshape pedagogy and equity within unequal educational settings. Throughout the analysis, codes were iteratively refined through constant comparison across interviews and contexts. Analytic memos were used to document decisions and trace connections between empirical patterns and theoretical constructs, ensuring the findings remained grounded in teachers' lived experiences while speaking directly to the study's sociotechnical focus. 4.5. Mapping Framework Components to Methods and Analytic Codes Table 1 Theoretical Framework–Method–Analysis Alignment Theoretical Framework Component How It Informed Data Collection (Methods) Key Analytic Codes (Examples) Linked Findings Theme AI as a Sociotechnical System Interview questions probed not only AI tools themselves, but also institutional contexts of adoption (policy mandates, accountability pressures, resource constraints). Teachers were asked how and why AI tools were introduced and what problems they were expected to solve. Institutional pressure , AI as management tool , policy-driven adoption , data demands , standardization Cross-cutting across all themes Educational Inequality & Redistribution of Learning Responsibility Teachers were asked which students benefited most from AI tools and which struggled, and why. Questions focused on learner autonomy, prior knowledge, and access to support. Self-regulation required , independent learning burden , student responsibility , prior knowledge gaps , confidence divides AI works best for students who already know how to learn Critical Mathematics Education (Nature of Mathematical Learning) Interview prompts explored how AI changed what mathematics was taught and how learning was structured (procedural vs conceptual, practice vs explanation). Procedural emphasis , answer-focused learning , lack of reasoning , repetitive practice , narrowed curriculum Automation increases efficiency, not understanding Algorithmic Power, Classification, and Datafication Teachers were asked about dashboards, analytics, and labels (e.g., “at risk”), and how these influenced instructional decisions and expectations. Algorithmic labeling , data-driven grouping , rigid classifications of ability , monitoring and surveillance Automation increases efficiency, not understanding (also informs Theme 1) Teachers as Human-in-the-Loop Mediators Interview questions explicitly addressed teacher agency: involvement in adoption, ability to adapt tools, time for intervention, and professional judgment. Teacher mediation , workarounds , resistance/adaptation , time constraints , reduced relational support Personalization without support becomes isolation Relational Pedagogy vs Individualization Teachers were asked how AI affected classroom interaction, student talk, collaboration, and help-seeking behaviors. Students working alone , reduced dialogue , silent disengagement , loss of collective learning Personalization without support becomes isolation 4.6. Trustworthiness and Reflexivity Several strategies were employed to enhance the trustworthiness of the analysis. First, variation in participants’ institutional contexts and teaching experiences supported analytic depth and comparative insight. Second, systematic memoing documented how theoretical assumptions informed coding decisions, supporting reflexivity and transparency. Third, attention was paid to disconfirming cases that complicated dominant patterns, particularly instances where AI tools appeared to support student engagement under specific conditions. Throughout the study, the researcher remained attentive to their own positionality and professional experience in disadvantaged educational contexts, treating this background as an analytic resource while critically interrogating its influence on interpretation. Future research would benefit from triangulating teacher perspectives with additional data sources, such as student interviews, classroom observations, and platform interaction logs, to further illuminate how AI systems shape learning across analytical levels. Such triangulation is not required to validate teachers’ accounts, but to extend understanding of how algorithmic processes, pedagogical mediation, and student experience intersect over time. This study, therefore, positions qualitative teacher inquiry as a foundational step in a broader research agenda on equity-centered AI in education 4.7. Ethical Considerations Ethical approval was obtained in accordance with institutional guidelines. All participants provided informed consent prior to participation. Pseudonyms were used, and identifying details were removed to protect confidentiality. 5. Findings Analysis of the teacher interviews revealed three interrelated themes that illuminate how AI-driven mathematics education is experienced and enacted in contexts of disadvantage. Across interviews, mathematics teachers did not reject AI outright; rather, they articulated nuanced accounts of how these systems interacted with existing inequalities, pedagogical constraints, and institutional priorities. The findings suggest that AI tools often function less as transformative interventions and more as amplifiers of pre-existing conditions. 5.1. AI Works Best for Students Who Already Know How to Learn Consistently, teachers reported that AI-based mathematics tools disproportionately benefit students who already possess strong learning strategies, prior knowledge, and confidence in mathematics. This pattern was closely linked to adaptivity logic embedded in AI systems, particularly performance-threshold routing, automated pacing, and feedback mechanisms that presupposed learner self-regulation and interpretive capacity. Mathematics teachers consistently reported that students who were organized, motivated, and academically successful were able to navigate adaptive platforms effectively, interpret feedback, and progress through content with minimal support. As one secondary mathematics teacher working in a low-income urban school explained: “The students who do well are the ones who already know how to sit down, focus, and figure things out on their own. For them, the program is like an extra tutor. But for the others, it just becomes another thing they don ’ t understand.” Participant mathematics teachers noted that many AI systems assume a level of self-regulation and metacognitive skill that cannot be taken for granted in disadvantaged contexts. Students who struggled with reading comprehension, language barriers, or gaps in foundational knowledge appeared to find AI feedback confusing or overwhelming. Rather than adapting meaningfully to these learners, the systems tended to recycle simpler tasks or flag students as “behind,” reinforcing deficit-oriented classifications. Several teachers described a pattern in which AI tools appeared to confirm existing hierarchies in the classroom: It doesn’t really change who succeeds. It just gives more practice to the kids who were already going to get it. This theme highlights how personalization, when detached from broader pedagogical and relational support, can reproduce rather than disrupt inequitable learning trajectories. In line with 5.2. Personalization as Transformative Individualization Under Constraint This theme illustrates not a failure of AI-driven personalization to transform classroom practice, but the specific form that transformation takes under conditions of institutional constraint. While AI systems were frequently introduced under the banner of personalized learning, mathematics teachers emphasized that personalization often translated into students working alone for extended periods, with reduced opportunities for interaction and explanation. This was particularly pronounced in under-resourced classrooms, where AI tools were sometimes used to manage large class sizes or compensate for limited instructional capacity. A primary school teacher described this tension as follows: They call it personalized, but what it really means is each child on a screen, working through questions by themselves. If they get stuck, the system gives them another question, not a conversation. Mathematics teachers expressed concern that students who were already disengaged or struggling seemed to experience AI-mediated learning as isolating rather than supportive. Without consistent teacher intervention, these students were more likely to disengage silently, guess their way through tasks, or abandon the activity altogether. In contrast, higher-achieving students were more likely to seek clarification independently or use feedback productively. Importantly, mathematics teachers framed this isolation not as a flaw of individual students, but as a structural consequence of how AI was implemented: In theory, I ’ m supposed to be circulating and supporting. In reality, I ’ m troubleshooting technology. This theme underscores that personalization, in the absence of relational and instructional support, can exacerbate signs of alienation—particularly for students who already feel marginalized within mathematics classrooms. In this sense, AI does transform pedagogy—shifting classrooms from collective, dialogic engagement toward individualized, screen-mediated participation—but this transformation remains pedagogically regressive in contexts where relational support cannot be sustained. 5.3. Automation Increases Efficiency, Not Understanding Mathematics teachers widely acknowledged that AI tools increased efficiency in certain respects, particularly in grading, progress tracking, and administrative reporting. Importantly, teachers identified bounded but tangible pedagogical opportunities arising from these efficiencies. Automated assessment and instant feedback were widely valued for supporting procedural fluency, particularly in foundational numeracy and exam-oriented practice. Several teachers described how AI systems created a low-stakes environment for repetition, enabling students to practice without fear of public failure and reducing teachers’ administrative burden for routine marking. In these cases, AI did not deepen conceptual understanding, but it reconfigured instructional time, creating potential—albeit unrealized—for teachers to redirect attention toward explanation, discussion, and targeted intervention. These opportunities were therefore not inherent to AI itself, but contingent on how efficiency gains were pedagogically reinvested. A recurring concern was that AI systems prioritized speed and correctness over reasoning and explanation: The system knows if the answer is right or wrong, but it doesn’t ’ t know why the student is thinking that way. That part still falls on the teacher—and we don ’ t always have time. In high-pressure environments shaped by standardized testing and accountability demands, several participant teachers reported using AI tools primarily for exam preparation or remediation. This often narrowed instructional focus to repetitive practice aligned with assessment formats, limiting opportunities for exploratory or conceptually rich mathematics. Teachers also raised concerns about predictive features that labeled students as “at risk” based on performance data: Once a student is flagged, everything they do is seen through that lens. The system expects them to struggle, and so does the school. Rather than supporting growth, automation sometimes reinforced static categorizations of competence, particularly among students from disadvantaged backgrounds. Teachers emphasized that while AI made certain processes more manageable, it did not replace the pedagogical work required to foster understanding, confidence, and mathematical identity. 5.4. Cross-Cutting Observation: Mathematics Teachers as Mediators Under Constraint Across all three themes, mathematics teachers positioned themselves as mediators attempting to reconcile equity-oriented commitments with tools that were not designed for their students’ realities. While many teachers adapted or resisted aspects of AI systems to better support learners, their agency was constrained by limited time, resources, and institutional expectations. As one teacher summarized: AI doesn’t ’ t decide what happens in my classroom—but it definitely shapes it. And it shapes it more when you don ’ t have many other options. These findings should not be read merely as a rejection of AI, but as a roadmap for more context-sensitive design. Teachers explicitly valued the technology when it operated as a low-stakes environment for repetition, allowing students to build foundational fluency without the social cost of public error. This suggests that the most equitable application of current AI in disadvantaged contexts may be its most modest: not as a substitute for teacher-led conceptual instruction, but as a dedicated tool for consolidating procedural knowledge. By designing systems that explicitly aim to support these bounded tasks—automating routine feedback to liberate teacher time for relational intervention—developers can realign AI tools with the actual needs of resource-constrained classrooms. 6. Discussion This study examined how AI-driven mathematics education is enacted in South African contexts of socio-economic disadvantage, foregrounding mathematics teachers’ perspectives to interrogate dominant claims about personalization, efficiency, and equity. Data analysis identified three primary dynamics that challenge these claims: the tendency of current tools to disproportionately favor self-regulated learners, the pedagogical risk of student isolation inherent in personalized pathways, and the prioritization of procedural efficiency over conceptual understanding. By aligning these findings with the conceptual structure established in the literature review, this section situates the empirical results within broader debates in AI-in-education and critical mathematics education. Collectively, the data suggest that AI does not simply transform education through technical innovation; rather, it reshapes existing pedagogical and institutional dynamics, often reproducing inequality. 6.1. AI-Driven Transformation in Mathematics Education: Reform Without Redistribution AI is frequently framed as a transformative intervention capable of fundamentally reshaping teaching and learning. The findings of this study suggest that AI does transform mathematics education, but primarily by reconfiguring pedagogical relations, redistributing responsibility for learning, and reshaping what counts as legitimate mathematical activity. In disadvantaged contexts, these transformations occur within tightly constrained institutional environments, resulting in forms of change that prioritize efficiency, monitoring, and individualization rather than conceptual depth or relational support. The findings are in line with Țîru et al. ( 2025 ), who argue that teachers' engagement with AI is influenced by institutional limits that lead to cautious, adaptive, or resistant practices, showing that AI-driven change tends to reshape teaching roles and responsibilities rather than deliver inherently progressive or equitable improvements in teaching and learning. Rather than enabling new pedagogical possibilities, AI systems in this study transformed existing practices by automating inequality-sensitive processes: accelerating the advancement of already-advantaged learners, individualizing struggle, and rendering underperformance more visible and administratively manageable. Transformation, therefore, did not take the form of innovation, but of intensification—making longstanding patterns of stratification more efficient, scalable, and difficult to contest. This interpretation reinforces the literature's positioning of AI as a socio-technical system embedded within existing educational regimes, rather than as a neutral or autonomous driver of change. Importantly, the findings demonstrate that low-level or rule-based AI systems can exert high-level socio-technical power. Even in the absence of advanced machine learning, these systems reshape educational practice by automating curricular pathways, stabilizing deficit classifications, and reorienting teacher attention toward data dashboards and performance signals. The transformative force of AI in these classrooms, therefore, derives not from predictive accuracy or adaptivity sophistication, but from the authority granted to algorithmic outputs within institutional decision-making processes. The findings align with Díaz and Nussbaum ( 2024 ), who argue that AI's impact on education lies more in teaching skills than in technical ability. Their analysis shows that when AI lacks a teaching foundation, even basic systems can worsen existing practices. They automate curriculum progression and performance tracking, thereby altering teaching priorities and reinforcing negative classifications in institutional decision-making. 6.2. The Self-Regulation Gap: How Algorithmic Design Amplifies Advantage The finding that current AI tools privilege students with existing academic capital aligns with research demonstrating that AI-driven personalization benefits learners who can interpret feedback, persist independently, and navigate abstract learning pathways. Rather than compensating for unequal preparation, personalization often amplified existing differences by accelerating the progress of already-advantaged students while relegating others to repetitive or simplified tasks (Williamson, 2017 ; Yaseen et al., 2025 ). From an equity perspective, this suggests that personalization redistributes responsibility for learning onto students themselves, privileging those whose dispositions align with dominant academic norms. Mathematics teachers’ experiences thus challenge the assumption that personalization is inherently equitable, highlighting the need for AI systems that explicitly account for socially patterned differences in learning conditions. In line with Yaseen et al. ( 2025 ), the findings suggest that adaptive and AI-supported learning environments tend to benefit students with higher digital literacy and self-regulatory abilities, whilst providing limited support to those without these skills, thereby reinforcing existing educational inequalities rather than reducing them. 6.3. Beyond the Digital Divide: From Resource Access to Algorithmic Determinism While previous research on educational technology has established the Matthew Effect—where students with prior advantages benefit disproportionately from digital tools—the findings suggest that AI exacerbates this dynamic through mechanisms distinct from older, static technologies. Unlike traditional Computer-Assisted Instruction (CAI), which functioned as a passive repository of content, the AI systems observed here exercised algorithmic agency that actively reinforced achievement gaps. This occurred primarily through two AI-specific mechanisms: The Remedial Loop (Algorithmic Containment) : High-performing students were accelerated by the algorithm, receiving validation and novel content. In contrast, lower-performing students were frequently trapped in adaptive logic, resulting in repetitive remedial loops. Unlike a static textbook, where a student can turn a page and attempt a harder problem, the opaque control logic of the AI prevented these students from accessing grade-level content, effectively codifying their low ability status into the software’s architecture. Predictive Labeling (The Self-Fulfilling Prophecy) : The risk dashboards generated by these systems shifted the pedagogical gaze. Mathematics teachers reported that s tudents' red/green coding influenced their expectations before instruction began. Unlike a grade book, this records past performance; these AI predictive models project future failures. This creates a barrier to algorithmic determinism, where the machine’s statistical inference of a student’s potential overrides the teacher’s belief in the student's capacity for growth. The findings align with previous research showing that AI tools can inadvertently reinforce inequities by advantaging students with prior advantages while constraining access for lower-performing disadvantaged students (Owan et al., 2023 ). From an equity-centered view, significant change entails inclusive AI design that lessens bias and reorganizes learning opportunities (Oyetade & Zuva, 2025 ). Without such attention, algorithmic personalization risks amplifying existing structural disparities rather than fostering equitable mathematics learning (Al-Chaer, 2026 ). 6.4. Personalization, Pedagogical Mediation, and Student Isolation The observation that personalization manifests as solitary confinement in under-resourced classrooms confirms the study’s theoretical concern: that individualization often serves to strip away the relational scaffolding necessary for deep mathematical learning. In under-resourced classrooms, AI systems were often used to manage large class sizes or compensate for staffing shortages, reducing rather than enhancing relational engagement. Students who struggled academically were especially likely to disengage, guess through tasks, or withdraw silently, reinforcing patterns of marginalization. This finding underscores the importance of pedagogical mediation in AI-supported learning environments. Without sustained teacher interaction, personalization risks becoming a mechanism for individualization rather than inclusion, undermining the relational foundations of effective mathematics instruction. Despite the broader challenges regarding equity and conceptual depth, it is important to acknowledge where these tools succeeded. Teachers consistently reported that current AI systems were highly effective for developing procedural fluency and foundational numeracy. For students who specifically needed repetition to master arithmetic operations or formulaic application, the immediate feedback loops provided a safe, non-judgmental space for practice. The AI's efficiency in these narrow domains successfully unburdened teachers from manual grading of rote tasks, theoretically freeing up time for higher-order instruction. The critical failure, however, was not in the tool’s ability to drill, but in the systemic over-reliance on this function as a proxy for comprehensive teaching. Consistent with Saleem et al. ( 2025 ) and Zagami ( 2026 ), these findings suggest that while AI-enabled personalization is often promoted as transformative for teaching, its impact really depends on the conditions at the institution. In constrained classrooms, personalization often manifests as individual task completion with little interaction, which may increase disengagement among already affected learners rather than encouraging their participation. 6.5. Automation, Efficiency, and the Limits of Measurable Understanding While the efficiency gains identified in the findings (e.g., automated grading) reduced administrative labor, this study illuminates a critical tension: when efficiency is decoupled from pedagogical purpose—measured through task completion, speed, or data visibility—it begins to substitute for learning rather than support it. In such cases, what is optimized is not understanding, but manageability. In high-stakes accountability contexts, AI tools were frequently repurposed for test preparation and remediation, reinforcing procedural approaches to mathematics that research has long shown to disadvantage marginalized learners. Teachers expressed concern that while AI made classrooms more manageable, it did not support the kinds of mathematical engagement necessary for equitable learning (Vasquez et al., 2024 ). Efficiency contributes to educational transformation only when it is deliberately subordinated to pedagogical goals such as conceptual understanding, relational engagement, and teacher judgment, rather than treated as a proxy for educational quality. This finding reinforces critiques of datafication in education, suggesting that efficiency gains should not be conflated with pedagogical improvement. In mathematics education, what is most valuable for learning is often least amenable to automation. These observations align with those of Hwang ( 2022 ) and Dabingaya ( 2022 ), who showed that AI-driven personalized learning environments can improve student achievement by tailoring instruction to individual learning speeds. Owan et al. ( 2023 ) reported that while AI improves procedural efficiency and enables low-stakes practice, its value for meaningful learning depends on the teacher's involvement. However, the increased efficiency in grading, monitoring, and automation does not automatically lead to a better understanding of concepts or fair participation in mathematics. Automated systems often focus on correctness and speed instead of reasoning, explanation, and the development of mathematical understanding. 6.6. Mathematics teachers as Constrained Human-in-the-Loop Mediators Across all findings, teachers emerged as critical mediators navigating the tensions between AI systems, institutional constraints, and equity-oriented commitments. Teachers actively interpreted, adapted, and sometimes resisted AI tools to support their students. However, their agency was consistently shaped—and often limited—by resource scarcity, prescriptive curricula, and accountability pressures. Rather than positioning inequitable outcomes as failures of implementation, the findings suggest that responsibility for equity is frequently displaced onto teachers, even when AI systems are poorly aligned with the realities of disadvantaged classrooms. This displacement obscures the roles of design, governance, and upstream policy decisions. Recognizing teachers as constrained human-in-the-loop actors underscores the need for AI systems developed in dialogue with educators and evaluated in the contexts in which they are used. Consistent with Țîru et al. ( 2025 ), the findings show that teachers often adopt critical, adaptive, or resistant views toward AI tools, yet their ability to act is limited by institutional pressures, resource constraints, and misaligned technology. As a result, the responsibility for equitable learning outcomes is shifted from AI systems to teachers. 6.7. Reframing Equity in AI-Driven Mathematics Education Claims about educational transformation through AI rest implicitly on assumptions about what counts as improvement and for whom such improvement matters. In this study, equity is not treated as an optional normative preference or a secondary social goal, but as a necessary condition for meaningful educational transformation. Transformative change in mathematics education cannot be said to occur if new systems systematically advantage learners who already possess academic, institutional, or cultural resources while rendering others more visible, more monitored, and more constrained. In unequal educational systems, technologies that improve outcomes for some while stabilizing or deepening disadvantage for others constitute optimization rather than transformation. From this perspective, equity functions as a diagnostic lens for evaluating whether AI-driven change alters the underlying distribution of learning opportunity or merely reconfigures its management. Without explicit attention to equity, AI systems risk being celebrated as transformative on the basis of efficiency gains or aggregate performance improvements, even as they reproduce stratified learning trajectories. Treating equity as foundational, therefore, shifts the evaluative question from whether AI works to how it reorganizes opportunity across learners and contexts. In line with Oyetade and Zuva ( 2025 ), the findings indicate that claims of AI-driven transformation should be assessed with a focus on equity, ensuring that technologies do not maximize outcomes for already advantaged students but rather actively reallocate learning opportunities and enhance teacher capacity to support all students. 7. Future Directions: Towards Equity-Centred AI in Mathematics To move beyond the limitations observed in this study, the design and implementation of educational AI must undergo three fundamental shifts: 7. 1. From Opaque Prediction to Explainable Diagnostics. Currently, the predictive nature of AI acts as a black box that labels students. Future development should prioritize Explainable AI (XAI). Instead of simply flagging a student as at risk (a label that lowers teacher expectations), systems should provide granular, transparent data on why a misconception is occurring (e.g., Student consistently misapplies the distributive property). This shifts the AI’s role from a content gatekeeper to a diagnostic partner that empowers the teacher to intervene precisely. 7. 2. From Isolation to Human-in-the-Loop Orchestration. The model of personalized learning as a solitary student facing a screen must be abandoned in favor of pedagogical orchestration. Future directions should focus on AI tools designed to prompt teacher-student interaction rather than replace it. For example, rather than silently routing a struggling student to a remedial video, the system should signal the teacher to facilitate a small-group intervention. The metric of success for EdTech should not be time on system, but the quality of off-screen mathematical discourse the system inspires. 7. 3. Designing for the Margins, Not the Mean. Finally, developers must move beyond designing for the ideal self-regulated learner. AI systems must include scaffolding explicitly designed for students with lower self-regulation and higher anxiety. This means moving away from frustration-inducing remedial loops and incorporating metacognitive prompts that help students plan their learning strategies. Equity cannot be an afterthought; it must be a core design parameter of the algorithm itself. In this sense, equity and efficiency are not competing values, but operate in a hierarchical relationship: efficiency can enable transformation only when equity defines its direction and limits. 8. Conclusion This study set out to examine how AI-driven mathematics education is experienced and enacted in contexts of socio-economic disadvantage, foregrounding mathematics teachers’ perspectives to interrogate dominant narratives of personalization, efficiency, and transformation in AI–supported education. The findings demonstrate that AI transforms mathematics education not solely through instructional intelligence, but also through its capacity to reorganize time, attention, accountability, and pedagogical responsibility within already unequal systems. Across the analysis, three interrelated dynamics emerged. First, AI-driven personalization tended to benefit students who already possessed the dispositions, confidence, and self-regulatory skills required to navigate autonomous learning environments. Rather than compensating for unequal preparation, adaptive systems often accelerated existing hierarchies of achievement. Second, personalization without sustained pedagogical and relational support often led to student isolation, particularly among learners who needed the most guidance. Third, automation increased efficiency in assessment and monitoring but did little to support conceptual understanding or equitable participation in mathematics learning. Together, these dynamics suggest that the transformative potential of AI in education is sharply limited when equity is treated as a downstream outcome rather than a foundational design principle. By centering mathematics teachers working in disadvantaged contexts, this study contributes a grounded account of how AI systems operate in everyday classroom practice. Mathematics teachers emerged not as passive implementers of technology, but as constrained mediators attempting to reconcile equity-oriented commitments with tools designed primarily for efficiency, scalability, and measurability. Importantly, inequitable outcomes were not attributable to teacher resistance or poor implementation, but to misalignments between AI system design and the social and pedagogical realities of marginalized educational settings. These findings carry implications for AI-in-education research, policy, and design. For researchers, the study underscores the need to move beyond system-centered evaluations toward analyses that attend to context, pedagogy, and power. Claims about transformation should be assessed not only in terms of technical performance or short-term learning gains, but in relation to how AI reshapes learning opportunities, teacher agency, and student participation over time. For policymakers, the results highlight the risks of deploying AI as a cost-effective substitute for instructional support in under-resourced schools and underscore the need for governance frameworks that foreground equity, transparency, and accountability. For designers and developers, the findings suggest that equity-oriented AI must support relational pedagogy, make assumptions about learning explicit, and expand—rather than constrain—mathematics teachers’ professional judgment. More broadly, this study argues that the central question for AI in education is not whether it can personalize learning, but under what conditions, and for whom, such personalization contributes to meaningful educational opportunity. Without deliberate, equity-centered intervention, AI risks rendering longstanding inequalities more efficient, more automated, and less visible. Conversely, when guided by critical attention to context, pedagogy, and justice, AI has the potential to support—not replace—the human relationships at the core of equitable mathematics education. Contexts of disadvantage should therefore be understood not as outliers in AI-in-education research, but as sites where the structural consequences of algorithmic design become particularly clear. In conclusion, transforming education through AI requires more than sophisticated algorithms or scalable platforms. It requires a reorientation of AI development and deployment toward the social conditions of learning, mathematics teachers' professional expertise, and the diverse ways students engage with mathematics. Only by treating equity as a starting point rather than an afterthought can AI contribute to genuine educational transformation. This study relied on teacher reporting of student engagement; future research should triangulate these findings with direct student feedback and platform interaction logs. Declarations Contributions: B C is the sole contributor to this article’s conceptualisation, data generation, analysis, and write-up. Funding: The author did not receive support from any organisation for the submitted work Competing interests: The author declares no competing interests Availability of data and material: Any requests for data should be addressed to the author, who will share in accordance with what is allowed by the signed consent forms Clinical trial number : Not applicable Ethics approval and Consent to participate Ethical clearance for this study was applied for and approved by the CPUT Health and Wellness Sciences (Protocol Number: CPUT/HWS-REC 2025/ S1 ). I confirm that all methods were carried out in accordance with the University of Johannesburg’s ethical guidelines and regulations. Informed Consent: Written informed consent was obtained from Undergraduate students before participation in the study. Consent to publish: The Author consents to the article being published. References Al-Chaer, E.D. (2026). Mind the Algorithm: Charting a Responsible Course for AI in Higher Education. In: Badran, A., Baydoun, E., Hillman, S., Mesmar, J. (eds) Higher Education in the Arab World. Springer, Cham. https://doi.org/10.1007/978-3-031-99068-7_3 Au, W. (2016). Meritocracy 2.0: High-stakes, standardized testing as a racial project of neoliberal multiculturalism. Educational Policy , 30 (1), 39–62. Azevedo, R., Taub, M., & Mudrick, N. V. (2018). Understanding and reasoning about real-time cognitive, affective, and metacognitive processes to foster self-regulation with advanced learning technologies. In D. H. Schunk & J. A. Greene (Eds.), Handbook of self-regulation of learning and performance (2nd ed., pp. 256–270). Routledge. Benjamin, R. (2019). Race after technology: Abolitionist tools for the new Jim Code. Polity Biesta, G. (2015). The beautiful risk of education . Paradigm Publishers. Boaler, J. (2016). Mathematical mindsets: Unleashing students' potential through creative math, inspiring messages and innovative teaching. Jossey-Bass. Bourdieu, P., & Passeron, J. C. (1990). Reproduction in education, society and culture . Sage Publications. Cobb, P., Gresalfi, M., & Hodge, L. L. (2009). An interpretive scheme for analyzing the identities that students develop in mathematics classrooms. Journal for Research in Mathematics Education , 40 (1), 40–68. Dabingaya, M. (2022). Analyzing the effectiveness of AI-powered adaptive learning platforms in mathematics education. Interdisciplinary Journal Papier Human Review, 3 (1), 1–7. https://doi.org/10.47667/ijphr.v3i1.226. Darling-Hammond, L. (2010). The flat world and education: How America's commitment to equity will determine our future. Teachers College Press. Díaz, B., & Nussbaum, M. (2024). Artificial intelligence for teaching and learning in schools: The need for pedagogical intelligence . Computers & Education, 217 (105071). https://doi.org/10.1016/j.compedu.2024.105071 Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. St. Martin's Press. Holmes, W., Bialik, M., & Fadel, C. (2019). Artificial intelligence in education: Promises and implications for teaching and learning. Center for Curriculum Redesign. Hwang, S. (2022). Examining the effects of artificial intelligence on Elementary Students’ mathematics achievement: A meta-analysis. Sustainability, 14 (20), 13185. https://doi.org/10.3390/su142013185. Kizilcec, R. F., Pérez-Sanagustín, M., & Maldonado, J. J. (2017). Self-regulated learning strategies predict learner behavior and goal attainment in Massive Open Online Courses. Computers & Education , 104 , 18–33. Knox, J. (2020). Artificial intelligence and education in China. Learning, Media and Technology , 45 (3), 298–311. Luckin, R., Holmes, W., Griffiths, M., & Forcier, L. B. (2016). Intelligence unleashed: An argument for AI in education . Pearson. Nasir, N. S., Cabana, C., Shreve, B., Woodbury, E., & Louie, N. (2014). Mathematics for equity: A framework for successful practice . Teachers College Press. O ’ Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. Crown. Owan, V. J., Abang, K. B., Idika, D. O., Etta, E. O., & Bassey, B. A. (2023). Exploring the potential of artificial intelligence tools in educational measurement and assessment. Eurasia Journal of Mathematics, Science and Technology Education, 19 (8), em2307. https://doi.org/10.29333/ejmste/13428 Oyetade, K., & Zuva, T. (2025). Advancing equitable education with inclusive AI to mitigate bias and enhance teacher literacy. Educational Process: International Journal, 14, (e2025087). https://doi.org/10.22521/edupij.2025.14.87 Priestley, M., Biesta, G., & Robinson, S. (2015). Teacher agency: An ecological approach . Bloomsbury Publishing. Saleem, S., Aziz, M. U., Iqbal, M. J., & Abbas, S. (2025). AI in education: Personalized learning systems and their impact on student performance and engagement. The Critical Review of Social Sciences Studies, 3 (1), 2445–2459. https://thecrsss.com/index.php/Journal/about Selwyn, N. (2016). Is technology good for education? Polity. Selwyn, N., Hillman, T., Eynon, R., Ferreira, G., Knox, J., Macgilchrist, F., & Sancho-Gil, J. M. (2020). What's next for Ed-Tech? Critical hopes and concerns for the 2020s. Learning, Media and Technology , 45 (1), 1–6. Sfard, A. (2008). Thinking as communicating: Human development, the growth of discourses, and mathematizing . Cambridge University Press. Skovsmose, O. (2011). An invitation to critical mathematics education. Sense Publishers. Țîru, L. G., Gherheș, V., Stoicov, I., & Stanici, M. (2025). Not Ready for AI? Exploring Teachers’ Negative Attitudes Toward Artificial Intelligence. Societies, 15 (12), 337. https://doi.org/10.3390/soc15120337 van Dijk, J. (2020). The digital divide . Polity. Vasquez, E., Basham, J. D., Jimenez, B., & Marino, M. T. (2024). Ethical considerations for educators leveraging artificial intelligence. In CIDDL (Eds.), Inclusive intelligence: The impact of AI on education for all learners (pp. 92–98). Center for Innovation, Design, and Digital Learning. Williamson, B. (2017). Big data in education: The digital future of learning, policy and practice. Sage. Williamson, B. and Eynon, R. (2020) Historical Threads, Missing Links, and Future Directions in AI in Education. Learning, Media and Technology, 45 , 223-235. https://doi.org/10.1080/17439884.2020.1798995 Yaseen, H., Mohammad, A. S., Ashal, N., Abusaimeh, H., Ali, A., & Sharabati, A.-A. A. (2025). The Impact of Adaptive Learning Technologies, Personalized Feedback, and Interactive AI Tools on Student Engagement: The Moderating Role of Digital Literacy. Sustainability, 17 (3), 1133. https://doi.org/10.3390/su17031133 Zagami, J. (2026). Addressing the Dark Side of Differentiation: Bias and Micro-Streaming in Artificial Intelligence Facilitated Lesson Planning. Information , 17 (1), 12. https://doi.org/10.3390/info17010012 Additional Declarations No competing interests reported. Supplementary Files AppendixA.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8679398","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":592411447,"identity":"ce70cfe9-5462-4a05-879c-435c95085023","order_by":0,"name":"Brantina Chirinda","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyklEQVRIiWNgGAWjYBACAzjJ3sDATLyWAyCS5wBJWkCERAKRWsylDz97/KHgjjz/zDeGnwsqbBj427sT8Gqx7EszNzhg8Mxwxu0cY+kZZ9IYJM6c3YDfYWcYzCQOGBxm3CCdYyDN23aYwUAil5AW9m8gLfYbJM8Y/yZSCw/YlsQNEjxmxNli2cNTJnHG4HAy0Btl1jxn0ngI+sWch32bRMWfw7b97Yc33+apsJHjb+/FrwUJcIDjiIdY5SDA/oAU1aNgFIyCUTCCAABik0VKM2LeRAAAAABJRU5ErkJggg==","orcid":"","institution":"University of the Witwatersrand","correspondingAuthor":true,"prefix":"","firstName":"Brantina","middleName":"","lastName":"Chirinda","suffix":""}],"badges":[],"createdAt":"2026-01-23 12:54:24","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8679398/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8679398/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108081054,"identity":"08849262-3485-4c43-8789-756657bd1a67","added_by":"auto","created_at":"2026-04-29 07:41:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":384979,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8679398/v1/c0723238-5691-4daa-b007-a46e8c26f83c.pdf"},{"id":102832460,"identity":"b21f66ee-91bc-43db-b327-d52b4dbc150a","added_by":"auto","created_at":"2026-02-17 10:17:29","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":17929,"visible":true,"origin":"","legend":"","description":"","filename":"AppendixA.docx","url":"https://assets-eu.researchsquare.com/files/rs-8679398/v1/3444b99acba394ec834aac83.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"The role of Artificial Intelligence in amplifying educational inequalities in resource constrained mathematics classrooms","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAI is increasingly positioned as a transformative force in education, with growing influence over how teaching, learning, and assessment are organized. Across policy documents, commercial platforms, and academic research, AI-driven systems are promoted as solutions to longstanding educational challenges, including learner heterogeneity, teacher workload, and persistent achievement gaps. Crucially, this study adopts a functional rather than a strictly technical definition of Artificial Intelligence. While computer science distinguishes between dynamic machine learning algorithms and static, rule-based branching (often found in older educational software), this distinction is frequently invisible to the end-user. For a student isolated behind a screen or a teacher managing a data dashboard, the phenomenology of automation remains the same: an external system has assumed authority over the pacing, sequencing, and assessment of knowledge. Whether the underlying code is a neural network or a complex decision tree, the resulting pedagogical dynamic is one of \u0026lsquo;infrastructural control\u0026rsquo; (Williamson, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Therefore, the sanalyze these technologies not by their computational complexity, but by their capacity to automate instructional decisions and govern classroom activity. Proponents of educational technology frequently position algorithmic solutions\u0026mdash;including adaptive courseware and automated assessment\u0026mdash;as essential infrastructure for achieving equitable and efficient mathematics education (Holmes et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Luckin et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2016\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThese claims resonate strongly within contemporary reform agendas that emphasize data-driven decision-making, scalability, and accountability. Within this discourse, personalization is often treated as both a technical and a moral achievement: by tailoring instruction to individual learners, AI systems are assumed to better support diverse needs than traditional classroom practices. As a result, AI is increasingly presented not merely as an instructional aid but as a catalyst for transforming education systems toward greater inclusivity and effectiveness.\u003c/p\u003e \u003cp\u003eYet alongside this optimism, a growing body of scholarship cautions that the educational effects of AI cannot be understood independently of the social, institutional, and pedagogical contexts in which these technologies are deployed. Research in critical AI studies and sociology of education demonstrates that technological systems are never neutral; they encode assumptions about learners, learning, and success, and tend to reflect the priorities of the institutions and markets that produce them (Selwyn, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Williamson, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Benjamin, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). In education, these assumptions often align with dominant norms of self-regulation, autonomy, and measurable performance\u0026mdash;norms that are unevenly distributed across socio-economic, linguistic, and cultural contexts.\u003c/p\u003e \u003cp\u003eMathematics education provides a particularly salient site for examining these tensions. Long recognized as a gatekeeping subject, mathematics plays a central role in shaping access to advanced study, STEM careers, and broader social and economic opportunities (Boaler, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Despite decades of reform, patterns of mathematical achievement remain persistently stratified along socio-economic and racial lines, with students in disadvantaged contexts more likely to encounter procedural, remedial, and test-oriented instruction (Nasir et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). These structural inequities form the backdrop against which AI-driven mathematics education is currently being introduced.\u003c/p\u003e \u003cp\u003eWithin policy and industry narratives, AI is frequently framed as uniquely capable of addressing these challenges. Adaptive systems promise to identify knowledge gaps, adjust pacing, and provide immediate feedback, thereby compensating for limited instructional capacity and enabling individualized learning pathways. Such claims are especially compelling in under-resourced schools, where large class sizes, teacher shortages, and accountability pressures intensify the appeal of scalable technological interventions.\u003c/p\u003e \u003cp\u003eHowever, emerging research raises critical questions about whether AI-driven personalization functions equitably in practice. Studies suggest that AI-based learning environments often work most effectively for students who already possess strong self-regulatory skills, prior academic knowledge, and familiarity with dominant learning norms (Azevedo et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Kizilcec et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Conversely, students who face linguistic barriers, limited academic support, or fragmented access to technology may experience AI systems as confusing, isolating, or demotivating. Rather than compensating for inequality, personalization may redistribute responsibility for learning in ways that are socially patterned and uneven.\u003c/p\u003e \u003cp\u003eThese concerns point to a broader issue in AI-in-education research: the tendency to evaluate AI systems primarily in terms of efficiency, predictive accuracy, or short-term learning gains, while paying less attention to how these systems reshape pedagogical relationships, learner identities, and opportunities to learn. In mathematics education, this often manifests through an emphasis on what is easily measurable\u0026mdash;speed, accuracy, and task completion\u0026mdash;rather than on reasoning, sense-making, and collaborative problem-solving. Such emphases risk reinforcing pedagogical approaches that have historically marginalized students in disadvantaged contexts.\u003c/p\u003e \u003cp\u003eCrucially, the impacts of AI in education are mediated through teachers, who operate at the intersection of technology, curriculum, and institutional constraint. Teachers make day-to-day decisions about how AI tools are used, whom they support, and what forms of learning are prioritized. Yet teachers\u0026rsquo; perspectives\u0026mdash;particularly those working in disadvantaged contexts\u0026mdash;remain underrepresented in research on educational AI, which often foregrounds system design or learner analytics over lived classroom practice.\u003c/p\u003e \u003cp\u003eThis study addresses this gap by examining how AI-driven mathematics education is experienced and enacted in contexts of socio-economic disadvantage in South Africa, drawing on qualitative interviews with mathematics teachers across diverse institutional settings. Given the Global South's distinct socio-economic disparities and resource constraints, this context offers a critical vantage point for interrogating the often-universalized claims of efficacy attached to AI technologies developed in the Global North. Rather than asking whether AI can personalize learning in principle, the study investigates how personalization, automation, and data-driven decision-making operate in practice and what consequences they have for educational equity.\u003c/p\u003e \u003cp\u003eBy foregrounding mathematics teachers\u0026rsquo; perspectives, the study contributes to ongoing debates about the opportunities and challenges of AI in education in three ways. First, it challenges narratives of technological neutrality by showing how AI systems often align with and amplify existing inequalities. Second, it highlights the central role of mathematics teachers as constrained mediators of AI, whose professional judgment is shaped by accountability pressures, resource limitations, and institutional expectations. Third, it reframes equity not as an automatic outcome of AI adoption, but as a design, governance, and pedagogical challenge that must be addressed explicitly.\u003c/p\u003e \u003cp\u003eIn doing so, the study speaks directly to the aims of this special issue by offering a critical, empirically grounded account of how AI is transforming education\u0026mdash;not only through new technical capabilities, but through its interaction with longstanding social and institutional structures. The guiding research question is:\u003c/p\u003e \u003cp\u003eHow does the introduction of AI-driven mathematics education reshape teaching and learning in contexts of disadvantage, and what challenges does this transformation pose for educational equity from the perspective of teachers?\u003c/p\u003e"},{"header":"2. Literature Review","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. AI-Driven Transformation in Mathematics Education\u003c/h2\u003e \u003cp\u003eAI has emerged as a central component of contemporary educational reform agendas, with increasing influence over curriculum delivery, assessment, and learner support. In mathematics education, AI-driven systems\u0026mdash;such as adaptive learning platforms, intelligent tutoring systems, automated feedback tools, and predictive analytics\u0026mdash;are widely promoted for their capacity to scale instruction, offering a mechanism to deliver standardized mathematics content to large student populations without the resource constraints associated with traditional human-led teaching (Holmes et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). These systems are often positioned as responses to persistent challenges, including heterogeneous classrooms, teacher workload, and unequal access to high-quality instruction.\u003c/p\u003e \u003cp\u003eIn policy and commercial discourse, AI-enabled personalization is frequently framed as a means of improving both effectiveness and equity. By tailoring instruction to individual learners\u0026rsquo; performance data, AI systems are assumed to provide more precise and responsive support than traditional classroom practices (Luckin et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). This promise is particularly salient in under-resourced educational contexts, where structural constraints heighten the appeal of automated and scalable solutions.\u003c/p\u003e \u003cp\u003eHowever, existing research suggests that claims of transformation must be examined in relation to the historical and institutional conditions of mathematics education. Longstanding patterns of inequality\u0026mdash;shaped by tracking, assessment regimes, and deficit-oriented conceptions of ability\u0026mdash;continue to structure who benefits from instructional innovations (Boaler, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). As such, AI enters mathematics classrooms not as a neutral intervention, but as a sociotechnical system embedded within already stratified educational environments.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Personalization, Adaptivity, and Assumptions About Learning\u003c/h2\u003e \u003cp\u003eCentral to AI-in-education discourse is the concept of personalization. Adaptive systems aim to adjust content, pacing, and feedback based on learners\u0026rsquo; interactions, often using performance metrics such as accuracy, speed, and persistence. Within mathematics education, personalization is frequently equated with responsiveness to learner diversity and, by extension, with equity (Luckin et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Holmes et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eYet critics argue that such approaches rely on narrow and implicit assumptions about learning. AI systems typically operationalize learning as individual progression through predefined tasks, privileging self-regulation, metacognitive awareness, and independent navigation of learning pathways (Williamson, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Knox, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). These assumptions align more closely with the dispositions of students who already possess academic confidence, prior knowledge, and familiarity with dominant schooling norms.\u003c/p\u003e \u003cp\u003eEmpirical research supports this concern. Studies of adaptive and AI-supported learning environments consistently find that students with stronger prior achievement and learning strategies benefit more from personalization, while those who struggle academically are more likely to disengage or experience limited gains (Azevedo et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Kizilcec et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). In mathematics education, where conceptual gaps can quickly compound, personalization may therefore stabilize existing hierarchies rather than disrupt them.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Algorithmic Systems, Datafication, and Educational Inequality\u003c/h2\u003e \u003cp\u003eA growing body of critical AI scholarship emphasizes that algorithmic systems are shaped by the data, values, and institutional priorities embedded in their design (Benjamin, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Eubanks, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). In education, AI systems rely on data that reflect historical inequalities in access, opportunity, and support, raising concerns about the reproduction of deficit-based classifications through automated decision-making.\u003c/p\u003e \u003cp\u003ePredictive analytics and adaptive pathways can label students as \u0026ldquo;at risk,\u0026rdquo; \u0026ldquo;behind,\u0026rdquo; or \u0026ldquo;low performing\u0026rdquo; based on patterns that are socially patterned rather than purely cognitive (O\u0026rsquo;Neil, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Williamson \u0026amp; Eynon, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). In mathematics education, such classifications are particularly consequential, as early labeling can narrow curricular exposure and shape long-term trajectories.\u003c/p\u003e \u003cp\u003eResearch on digital inequality further underscores that educational outcomes are shaped less by access to technology than by differences in use, support, and institutional mediation (van Dijk, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Without deliberate attention to equity, AI systems may function as mechanisms for automating existing inequalities under the guise of objectivity and precision.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Personalization, Pedagogical Mediation, and Student Isolation\u003c/h2\u003e \u003cp\u003eWhile AI-driven personalization is often promoted as learner-centered, critics note that it frequently coincides with a reduction in human interaction. In practice, personalization may involve students working independently on screens for extended periods, with limited opportunities for dialogue, explanation, or collaborative problem-solving (Selwyn et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eResearch in mathematics education consistently highlights the importance of relational pedagogy\u0026mdash;teacher questioning, peer interaction, and collective sense-making\u0026mdash;for developing conceptual understanding and mathematical identity (Sfard, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Cobb et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). When AI systems replace rather than supplement these interactions, students who require the most support may experience learning as isolating or confusing.\u003c/p\u003e \u003cp\u003eStudies indicate that students facing academic difficulty are more likely to perceive automated feedback as opaque or demotivating, particularly when it lacks explanatory depth or opportunities for clarification (Eubanks, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). In disadvantaged contexts, where instructional support is already stretched, personalization without pedagogical mediation may intensify disengagement rather than foster inclusion.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5. Automation, Efficiency, and the Limits of Measurable Learning\u003c/h2\u003e \u003cp\u003eAnother dominant justification for AI in education is efficiency. Automated grading, progress monitoring, and analytics are promoted as ways to reduce teacher workload and enhance accountability (Luckin et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). While such efficiencies can streamline administrative processes, scholars caution that they also reshape what counts as legitimate learning.\u003c/p\u003e \u003cp\u003eAutomation tends to privilege what is easily measurable, narrowing instructional focus toward short-answer tasks, procedural fluency, and alignment with standardized assessments (Biesta, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Au, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). In mathematics education, this emphasis risks marginalizing reasoning, explanation, and exploratory problem-solving\u0026mdash;practices shown to be particularly important for equitable learning.\u003c/p\u003e \u003cp\u003ePredictive and monitoring systems may further contribute to rigid classifications of ability, especially when labels such as \u0026ldquo;at risk\u0026rdquo; become institutionalized within schools (O\u0026rsquo;Neil, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Benjamin, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). In high-pressure environments, efficiency-driven AI use may therefore support institutional management more than student understanding.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6. Teachers as Human-in-the-Loop Mediators of AI Systems\u003c/h2\u003e \u003cp\u003eIncreasingly, AI-in-education research recognizes that the effects of AI systems are mediated through teachers, who interpret, adapt, and contextualize technologies in practice. Teachers function as \u0026ldquo;human-in-the-loop\u0026rdquo; actors whose professional judgment shapes how AI is enacted in classrooms (Priestley et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eHowever, teacher agency is unevenly distributed. In disadvantaged contexts, teachers often operate under conditions of limited resources, prescriptive curricula, and high-stakes accountability, constraining their ability to integrate AI in equity-oriented ways (Darling-Hammond, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2010\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eBuilding on this literature, this study foregrounds mathematics teachers\u0026rsquo; perspectives from contexts of disadvantage to examine how AI is enacted in everyday practice. By explicitly linking AI-driven personalization, automation, and teacher mediation to questions of equity, the study contributes a critical, empirically grounded account of how AI is transforming mathematics education\u0026mdash;and for whom.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Theoretical Framework","content":"\u003cp\u003eWhile much of the literature frames AI in education through the lens of teacher acceptance or efficacy, this study adopts a critical sociotechnical perspective focused on pedagogical automation and labor redistribution. The study draws on the work of Selwyn (2019) and Williamson (2021) not merely to argue that technology is social, but to examine how AI-driven systems restructure the essential dynamics of teaching and learning. Central to this framework is the concept of redistributing learning responsibility. Rather than functioning simply as neutral aids, adaptive platforms frequently offload the labor of pacing, remediation, and engagement management from the teacher to the student-technology dyad. As noted later in the analysis, this shift presupposes a level of self-regulation and digital literacy that is unequally distributed among learners. Therefore, AI was aligned not as a tool that teachers use, but as an actor that automates pedagogical decision-making. This lens shows how the efficiency promised by AI often comes at the cost of deepening structural inequalities, as the burden of navigating these automated systems falls disproportionately on students in under-resourced environments who lack the scaffolding to manage this new responsibility.\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.1. AI as a Socio-technical System in Education\u003c/h2\u003e \u003cp\u003eThe study approaches AI not as a neutral pedagogical aid, but as a political artifact that creates new asymmetries of power in the classroom. From this perspective, algorithmic platforms do not simply support learning; they operationalize specific ideologies about knowledge and ability, effectively deciding whose ways of thinking are valued and whose are marginalized. AI systems are designed, trained, procured, and implemented within policy regimes that prioritize scalability, efficiency, accountability, and datafication (Williamson, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). These priorities shape how learning is modeled, what forms of knowledge are valued, and how success is defined within AI-driven environments.\u003c/p\u003e \u003cp\u003eFrom this perspective, the effects of AI cannot be attributed solely to algorithmic capability. Instead, they emerge through interactions between technical design, institutional constraints, and human actors. In mathematics education, AI systems often encode assumptions about learning as individualized, measurable, and self-directed\u0026mdash;assumptions that align unevenly with the realities of disadvantaged classrooms. Conceptualizing AI as sociotechnical enables the analysis to move beyond questions of effectiveness toward questions of power, governance, and equity. From this perspective, the educational significance of AI lies less in whether systems employ advanced machine learning, and more in how algorithmic automation, classification, and datafication reorganize pedagogical authority and learning opportunity.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Educational Inequality and the Redistribution of Learning Responsibility\u003c/h2\u003e \u003cp\u003eThe framework draws on sociological theories of educational inequality, particularly those that emphasize how schooling redistributes responsibility for success and failure in socially patterned ways. Mathematics education has long been identified as a site of social reproduction, where access to valued forms of knowledge is shaped by socio-economic status, language, race, and institutional sorting mechanisms (Bourdieu \u0026amp; Passeron, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e1990\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe study posits that when the responsibility for learning is delegated to an algorithm, the human elements of teaching\u0026mdash;care, motivation, and the diagnosis of root causes of error\u0026mdash;are often stripped away, leaving vulnerable students to face a deficit based solely on data metrics. AI-driven personalization plays a key role in this redistribution. By emphasizing self-paced progression, independent problem-solving, and continuous performance monitoring, AI systems shift responsibility for learning onto individual students. This shift advantages learners who possess forms of cultural capital aligned with dominant academic norms\u0026mdash;such as self-regulation, confidence, and familiarity with abstract mathematical representations\u0026mdash;while disadvantaging those who require relational, dialogic, or scaffolded support.\u003c/p\u003e \u003cp\u003eThis component of the framework directly informs the interpretation of the finding that AI works best for students who already know how to learn, positioning this outcome as a structural effect of personalization rather than a deficit of individual learners. By automating feedback and pacing, these systems risk mechanizing the Matthew Effect in education\u0026mdash;where those with prior advantages (connectivity, high self-regulation) gain the most, while those relying on the school for structure are further marginalized.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Critical Mathematics Education and the Nature of Mathematical Learning\u003c/h2\u003e \u003cp\u003eTo examine how AI reshapes what counts as legitimate mathematics learning, the framework draws on critical mathematics education. This tradition emphasizes that mathematics education is not neutral or purely technical, but involves normative decisions about which forms of reasoning, participation, and knowledge are valued (Skovsmose, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Frankenstein, 2012).\u003c/p\u003e \u003cp\u003eAI-driven systems frequently prioritize procedural correctness, efficiency, and task completion, aligning with what critical scholars describe as exercise paradigms of learning. Such paradigms limit opportunities for explanation, dialogue, and collective sense-making\u0026mdash;practices shown to be essential for developing conceptual understanding and mathematical identity, particularly for students in disadvantaged contexts.\u003c/p\u003e \u003cp\u003eWithin this framework, the finding that automation increases efficiency, not understanding, is interpreted as a consequence of how AI systems operationalize learning, rather than as an unintended side effect. The framework thus highlights tensions between AI-supported efficiency and equity-oriented pedagogy.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.4. Algorithmic Power, Classification, and Datafication\u003c/h2\u003e \u003cp\u003eThe framework also incorporates insights from critical data and algorithm studies, which examine how classification, prediction, and automation shape social outcomes. Educational AI systems rely on historical data that reflect existing inequalities in access, opportunity, and support (O\u0026rsquo;Neil, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Benjamin, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). As a result, adaptive pathways and predictive analytics may reproduce deficit-based narratives about students, particularly those from marginalized backgrounds.\u003c/p\u003e \u003cp\u003eIn mathematics education, algorithmic labeling\u0026mdash;such as identifying students as \u0026ldquo;at risk\u0026rdquo; or \u0026ldquo;low performing\u0026rdquo;\u0026mdash;can narrow learning opportunities and shape expectations in ways that become self-reinforcing. Datafication further privileges what is easily measurable, reinforcing curricular narrowing and accountability-driven pedagogies.\u003c/p\u003e \u003cp\u003eThis aspect of the framework informs the analysis of how automation and monitoring reshape teaching and learning, contributing to all three findings themes by explaining how inequality becomes embedded in algorithmic processes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.5. Teachers as Constrained Human-in-the-Loop Mediators\u003c/h2\u003e \u003cp\u003eA final and central component of the framework positions teachers as human-in-the-loop mediators between AI systems and students. Teachers interpret, adapt, and sometimes resist AI tools based on their professional judgment, pedagogical commitments, and contextual knowledge (Priestley et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). However, their capacity to do so is shaped by institutional conditions, including curriculum mandates, assessment regimes, resource constraints, and professional development opportunities.\u003c/p\u003e \u003cp\u003eIn disadvantaged contexts, teachers often face heightened accountability pressures and limited autonomy, constraining their ability to integrate AI in equity-oriented ways. This framework rejects deficit explanations that locate responsibility for inequitable outcomes in teachers\u0026rsquo; practices, instead emphasizing how upstream design and governance decisions shape what teachers can do in practice.\u003c/p\u003e \u003cp\u003eThis perspective is essential for understanding why personalization without support can lead to isolation, as teachers\u0026rsquo; ability to provide relational mediation is frequently undermined by the very systems designed to increase efficiency.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.6. Integrative Role of the Framework\u003c/h2\u003e \u003cp\u003eTogether, these theoretical perspectives provide a coherent lens for analyzing how AI-driven mathematics education operates within unequal educational systems. The framework links mathematics teachers\u0026rsquo; lived experiences to broader sociotechnical dynamics, enabling the study to move beyond surface-level descriptions of AI use toward a critical examination of how AI redistributes responsibility, reshapes pedagogy, and reconfigures equity.\u003c/p\u003e \u003cp\u003eBy aligning theories of sociotechnical systems, educational inequality, critical mathematics education, and algorithmic power, the framework supports a nuanced analysis of AI as both an opportunity and a risk. It also provides a foundation for rethinking how AI in education might be designed and governed to support\u0026mdash;not undermine\u0026mdash;equitable mathematics learning.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Methods","content":"\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Research Design\u003c/h2\u003e \u003cp\u003eThis study employed a qualitative research design to examine how AI\u0026ndash;driven mathematics education is enacted in contexts of socio-economic disadvantage. Qualitative methods were selected to capture mathematics teachers\u0026rsquo; situated experiences, professional judgments, and interpretations of AI systems in everyday classroom practice\u0026mdash;dimensions that are often obscured in system-centered or outcome-focused evaluations of educational AI.\u003c/p\u003e \u003cp\u003eGuided by a critical sociotechnical framework, the study conceptualized AI not as a neutral instructional tool, but as a system embedded within institutional constraints, pedagogical traditions, and historical patterns of educational inequality. This framework informed all stages of the research process, including participant selection, interview design, and analytic strategy.\u003c/p\u003e \u003cp\u003eWhile this study centers on teachers' perspectives rather than direct student data, mathematics teachers were positioned as critical expert witnesses in the learning process. Consequently, references to student 'confusion' or 'isolation' in the findings reflect the teachers' professional diagnosis of student affect based on longitudinal observation. Unlike platform analytics\u0026mdash;which capture only click-rates, time-on-task, and accuracy\u0026mdash;teachers possess the longitudinal and ecological context necessary to interpret why a student pauses or disengages. Teachers observe the affective and behavioral correlates of data points (e.g., the difference between a student pausing to think versus pausing due to disengagement). Consequently, this analysis treats teacher observation not merely as opinion, but as a form of situated professional judgment that illuminates the classroom dynamics often invisible to the algorithmic gaze.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Participants and Context\u003c/h2\u003e \u003cp\u003eParticipants were 26 mathematics teachers working in South Africa\u0026rsquo;s plublic schools characterized by socio-economic disadvantage, ensuring the analysis captured the realities of AI implementation outside of well-resourced, elite private institutions. South African school quintiles (Q1-Q5) classify public schools by community poverty, with Q1 being the poorest (non-fee paying) and Q5 the least poor (fee-paying). Mathematics teachers were recruited using purposive sampling to ensure variation across school levels, institutional contexts, and degrees of exposure to AI-driven instructional tools. All participants had direct experience using AI-based platforms for mathematics instruction, including adaptive learning systems, automated assessment tools, or data analytics dashboards. The platforms ranged from gamified interfaces requiring rapid-fire input to dashboard-centric systems presenting static problem sets.\u003c/p\u003e \u003cp\u003eWhile the commercial platforms utilized by participants were broadly marketed as AI, the study\u0026rsquo;s analysis distinguishes between the specific forms of algorithmic agency encountered in practice. The majority of tools relied primarily on rule-based branching and automation; these operated effectively as digital worksheets that routed students based on static if-then performance thresholds rather than dynamic machine-learning models. A smaller subset of systems employed predictive analytics, using opaque algorithms to generate risk labels or individualized pathways without transparency to users. This technical distinction is critical, as the study\u0026rsquo;s findings reveal that the limitations of the former (e.g., frustration with repetitive loops) and the opacity of the latter (e.g., fatalistic labeling) contributed distinctively to the inequities observed in the classroom. This distinction is not merely technical, but analytically central to the study. The transformative effects examined here do not arise primarily from the sophistication of machine learning models, but rather from algorithmic governance mechanisms embedded in these systems\u0026mdash;namely, automated instructional sequencing, data-driven learner classification, and continuous performance monitoring. Even relatively simple rule-based systems exert substantial sociotechnical power when they structure access to content, regulate pacing, and shape how teachers and institutions interpret student ability. Consequently, this study treats AI not as a question of computational complexity, but as a form of infrastructural control over pedagogical decision-making.\u003c/p\u003e \u003cp\u003eOf the 26 interviews, 17 were conducted virtually and 9 face-to-face. This mixed interview modality reflected both participants' availability and contextual constraints, enabling the inclusion of mathematics teachers across geographically dispersed settings. No substantive differences were observed in the depth or focus of responses across interview modes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e4.3. Data Collection\u003c/h2\u003e \u003cp\u003eData were collected through semi-structured interviews designed to elicit mathematics teachers\u0026rsquo; perspectives on the selection, implementation, and pedagogical consequences of AI-driven mathematics tools. Interview questions were explicitly informed by the study\u0026rsquo;s theoretical framework and treated its core components as sensitizing concepts rather than prescriptive categories.\u003c/p\u003e \u003cp\u003eInterview protocols explored five interrelated domains:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eInstitutional context and AI adoption\u003c/b\u003e, including how and why AI tools were introduced and the problems they were expected to address.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePedagogical use of AI\u003c/b\u003e, focusing on how AI systems shaped instructional practices and mathematical content.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePersonalization and differentiation\u003c/b\u003e, with particular attention to which students benefited most or least from AI-mediated instruction.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAutomation, data, and classification\u003c/b\u003e, including experiences with dashboards, analytics, and labels such as \u0026ldquo;at risk.\u0026rdquo;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eTeacher agency and mediation\u003c/b\u003e, examining mathematics teachers\u0026rsquo; capacity to adapt, resist, or supplement AI tools in practice.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003e Interviews typically lasted 45\u0026ndash;75 minutes and were audio-recorded with participants\u0026rsquo; consent. All interviews were transcribed verbatim prior to analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e4.4. Data Analysis\u003c/h2\u003e \u003cp\u003eData analysis followed a theoretically informed thematic analysis, where the study\u0026rsquo;s critical sociotechnical framework explicitly guided coding and theme development. Rather than treating theory as a post hoc interpretive lens, the framework provided sensitizing concepts that oriented analytic attention toward specific dimensions of AI-mediated practice, including the redistribution of learning responsibility, forms of pedagogical mediation, and the role of algorithmic systems in shaping opportunity.\u003c/p\u003e \u003cp\u003eThe analytic process proceeded in three iterative stages:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eFirst-Order Coding (Initial Coding)\u003c/strong\u003e \u003cp\u003eAnalysis began with an inductive focus on teachers\u0026rsquo; concrete descriptions of AI use, such as specific classroom practices, student behavioral responses (e.g.,\u0026ldquo;students work alone,\u0026rdquo; \u0026ldquo;system flags them as behind\u0026rdquo;), and institutional expectations.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eFocused Coding (Second-Order Categories)\u003c/strong\u003e \u003cp\u003eInitial codes were then clustered around higher-order categories aligned with the theoretical framework. For example, concrete observations of student autonomy were grouped under \u003cem\u003eRedistribution of Learning Responsibility\u003c/em\u003e, while descriptions of dashboards were analyzed through the lens of \u003cem\u003eDatafication and Classification\u003c/em\u003e.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eThematic Synthesis\u003c/b\u003e: Finally, these categories were integrated into three overarching themes that reflect broader socio-technical dynamics: (a) \u003cem\u003eAI works best for students who already know how to learn\u003c/em\u003e; (b) \u003cem\u003ePersonalization without support becomes isolation\u003c/em\u003e; and (c) \u003cem\u003eAutomation increases efficiency, not understanding\u003c/em\u003e. These themes were not treated as mutually exclusive but as interrelated manifestations of how AI systems reshape pedagogy and equity within unequal educational settings.\u003c/p\u003e \u003cp\u003eThroughout the analysis, codes were iteratively refined through constant comparison across interviews and contexts. Analytic memos were used to document decisions and trace connections between empirical patterns and theoretical constructs, ensuring the findings remained grounded in teachers' lived experiences while speaking directly to the study's sociotechnical focus.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e4.5. Mapping Framework Components to Methods and Analytic Codes\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTheoretical Framework\u0026ndash;Method\u0026ndash;Analysis Alignment\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTheoretical Framework Component\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHow It Informed Data Collection (Methods)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eKey Analytic Codes (Examples)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLinked Findings Theme\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAI as a Sociotechnical System\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInterview questions probed not only AI tools themselves, but also institutional contexts of adoption (policy mandates, accountability pressures, resource constraints). Teachers were asked how and why AI tools were introduced and what problems they were expected to solve.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eInstitutional pressure\u003c/em\u003e, \u003cem\u003eAI as management tool\u003c/em\u003e, \u003cem\u003epolicy-driven adoption\u003c/em\u003e, \u003cem\u003edata demands\u003c/em\u003e, \u003cem\u003estandardization\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCross-cutting across all themes\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eEducational Inequality \u0026amp; Redistribution of Learning Responsibility\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTeachers were asked which students benefited most from AI tools and which struggled, and why. Questions focused on learner autonomy, prior knowledge, and access to support.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eSelf-regulation required\u003c/em\u003e, \u003cem\u003eindependent learning burden\u003c/em\u003e, \u003cem\u003estudent responsibility\u003c/em\u003e, \u003cem\u003eprior knowledge gaps\u003c/em\u003e, \u003cem\u003econfidence divides\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAI works best for students who already know how to learn\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCritical Mathematics Education (Nature of Mathematical Learning)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInterview prompts explored how AI changed what mathematics was taught and how learning was structured (procedural vs conceptual, practice vs explanation).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eProcedural emphasis\u003c/em\u003e, \u003cem\u003eanswer-focused learning\u003c/em\u003e, \u003cem\u003elack of reasoning\u003c/em\u003e, \u003cem\u003erepetitive practice\u003c/em\u003e, \u003cem\u003enarrowed curriculum\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomation increases efficiency, not understanding\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAlgorithmic Power, Classification, and Datafication\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTeachers were asked about dashboards, analytics, and labels (e.g., \u0026ldquo;at risk\u0026rdquo;), and how these influenced instructional decisions and expectations.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eAlgorithmic labeling\u003c/em\u003e, \u003cem\u003edata-driven grouping\u003c/em\u003e, \u003cem\u003erigid classifications of ability\u003c/em\u003e, \u003cem\u003emonitoring and surveillance\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomation increases efficiency, not understanding (also informs Theme 1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTeachers as Human-in-the-Loop Mediators\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInterview questions explicitly addressed teacher agency: involvement in adoption, ability to adapt tools, time for intervention, and professional judgment.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eTeacher mediation\u003c/em\u003e, \u003cem\u003eworkarounds\u003c/em\u003e, \u003cem\u003eresistance/adaptation\u003c/em\u003e, \u003cem\u003etime constraints\u003c/em\u003e, \u003cem\u003ereduced relational support\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePersonalization without support becomes isolation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRelational Pedagogy vs Individualization\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTeachers were asked how AI affected classroom interaction, student talk, collaboration, and help-seeking behaviors.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eStudents working alone\u003c/em\u003e, \u003cem\u003ereduced dialogue\u003c/em\u003e, \u003cem\u003esilent disengagement\u003c/em\u003e, \u003cem\u003eloss of collective learning\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePersonalization without support becomes isolation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e4.6. Trustworthiness and Reflexivity\u003c/h2\u003e \u003cp\u003eSeveral strategies were employed to enhance the trustworthiness of the analysis. First, variation in participants\u0026rsquo; institutional contexts and teaching experiences supported analytic depth and comparative insight. Second, systematic memoing documented how theoretical assumptions informed coding decisions, supporting reflexivity and transparency. Third, attention was paid to disconfirming cases that complicated dominant patterns, particularly instances where AI tools appeared to support student engagement under specific conditions.\u003c/p\u003e \u003cp\u003eThroughout the study, the researcher remained attentive to their own positionality and professional experience in disadvantaged educational contexts, treating this background as an analytic resource while critically interrogating its influence on interpretation.\u003c/p\u003e \u003cp\u003eFuture research would benefit from triangulating teacher perspectives with additional data sources, such as student interviews, classroom observations, and platform interaction logs, to further illuminate how AI systems shape learning across analytical levels. Such triangulation is not required to validate teachers\u0026rsquo; accounts, but to extend understanding of how algorithmic processes, pedagogical mediation, and student experience intersect over time. This study, therefore, positions qualitative teacher inquiry as a foundational step in a broader research agenda on equity-centered AI in education\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e4.7. Ethical Considerations\u003c/h2\u003e \u003cp\u003e \u003cstrong\u003eEthical approval\u003c/strong\u003e \u003cp\u003e was obtained in accordance with institutional guidelines. All participants provided informed consent prior to participation. Pseudonyms were used, and identifying details were removed to protect confidentiality.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Findings","content":"\u003cp\u003eAnalysis of the teacher interviews revealed three interrelated themes that illuminate how AI-driven mathematics education is experienced and enacted in contexts of disadvantage. Across interviews, mathematics teachers did not reject AI outright; rather, they articulated nuanced accounts of how these systems interacted with existing inequalities, pedagogical constraints, and institutional priorities. The findings suggest that AI tools often function less as transformative interventions and more as amplifiers of pre-existing conditions.\u003c/p\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e5.1. AI Works Best for Students Who Already Know How to Learn\u003c/h2\u003e \u003cp\u003eConsistently, teachers reported that AI-based mathematics tools disproportionately benefit students who already possess strong learning strategies, prior knowledge, and confidence in mathematics. This pattern was closely linked to adaptivity logic embedded in AI systems, particularly performance-threshold routing, automated pacing, and feedback mechanisms that presupposed learner self-regulation and interpretive capacity. Mathematics teachers consistently reported that students who were organized, motivated, and academically successful were able to navigate adaptive platforms effectively, interpret feedback, and progress through content with minimal support.\u003c/p\u003e \u003cp\u003eAs one secondary mathematics teacher working in a low-income urban school explained:\u003c/p\u003e \u003cp\u003e \u003cem\u003e\u0026ldquo;The students who do well are the ones who already know how to sit down, focus, and figure things out on their own. For them, the program is like an extra tutor. But for the others, it just becomes another thing they don\u003c/em\u003e\u0026rsquo;\u003cem\u003et understand.\u0026rdquo;\u003c/em\u003e\u003c/p\u003e \u003cp\u003eParticipant mathematics teachers noted that many AI systems assume a level of self-regulation and metacognitive skill that cannot be taken for granted in disadvantaged contexts. Students who struggled with reading comprehension, language barriers, or gaps in foundational knowledge appeared to find AI feedback confusing or overwhelming. Rather than adapting meaningfully to these learners, the systems tended to recycle simpler tasks or flag students as \u0026ldquo;behind,\u0026rdquo; reinforcing deficit-oriented classifications.\u003c/p\u003e \u003cp\u003eSeveral teachers described a pattern in which AI tools appeared to confirm existing hierarchies in the classroom:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003eIt doesn\u0026rsquo;t really change who succeeds. It just gives more practice to the kids who were already going to get it.\u003c/em\u003e \u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis theme highlights how personalization, when detached from broader pedagogical and relational support, can reproduce rather than disrupt inequitable learning trajectories. In line with\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e5.2. Personalization as Transformative Individualization Under Constraint\u003c/h2\u003e \u003cp\u003eThis theme illustrates not a failure of AI-driven personalization to transform classroom practice, but the specific form that transformation takes under conditions of institutional constraint. While AI systems were frequently introduced under the banner of personalized learning, mathematics teachers emphasized that personalization often translated into students working alone for extended periods, with reduced opportunities for interaction and explanation. This was particularly pronounced in under-resourced classrooms, where AI tools were sometimes used to manage large class sizes or compensate for limited instructional capacity.\u003c/p\u003e \u003cp\u003eA primary school teacher described this tension as follows:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003eThey call it personalized, but what it really means is each child on a screen, working through questions by themselves. If they get stuck, the system gives them another question, not a conversation.\u003c/em\u003e \u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eMathematics teachers expressed concern that students who were already disengaged or struggling seemed to experience AI-mediated learning as isolating rather than supportive. Without consistent teacher intervention, these students were more likely to disengage silently, guess their way through tasks, or abandon the activity altogether. In contrast, higher-achieving students were more likely to seek clarification independently or use feedback productively.\u003c/p\u003e \u003cp\u003eImportantly, mathematics teachers framed this isolation not as a flaw of individual students, but as a structural consequence of how AI was implemented:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003eIn theory, I\u003c/em\u003e\u0026rsquo;\u003cem\u003em supposed to be circulating and supporting. In reality, I\u003c/em\u003e\u0026rsquo;\u003cem\u003em troubleshooting technology.\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis theme underscores that personalization, in the absence of relational and instructional support, can exacerbate signs of alienation\u0026mdash;particularly for students who already feel marginalized within mathematics classrooms. In this sense, AI does transform pedagogy\u0026mdash;shifting classrooms from collective, dialogic engagement toward individualized, screen-mediated participation\u0026mdash;but this transformation remains pedagogically regressive in contexts where relational support cannot be sustained.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e5.3. Automation Increases Efficiency, Not Understanding\u003c/h2\u003e \u003cp\u003eMathematics teachers widely acknowledged that AI tools increased efficiency in certain respects, particularly in grading, progress tracking, and administrative reporting. Importantly, teachers identified bounded but tangible pedagogical opportunities arising from these efficiencies. Automated assessment and instant feedback were widely valued for supporting procedural fluency, particularly in foundational numeracy and exam-oriented practice. Several teachers described how AI systems created a low-stakes environment for repetition, enabling students to practice without fear of public failure and reducing teachers\u0026rsquo; administrative burden for routine marking. In these cases, AI did not deepen conceptual understanding, but it reconfigured instructional time, creating potential\u0026mdash;albeit unrealized\u0026mdash;for teachers to redirect attention toward explanation, discussion, and targeted intervention. These opportunities were therefore not inherent to AI itself, but contingent on how efficiency gains were pedagogically reinvested.\u003c/p\u003e \u003cp\u003eA recurring concern was that AI systems prioritized speed and correctness over reasoning and explanation:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003eThe system knows if the answer is right or wrong, but it doesn\u0026rsquo;t\u003c/em\u003e\u0026rsquo;\u003cem\u003et know why the student is thinking that way. That part still falls on the teacher\u0026mdash;and we don\u003c/em\u003e\u0026rsquo;\u003cem\u003et always have time.\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn high-pressure environments shaped by standardized testing and accountability demands, several participant teachers reported using AI tools primarily for exam preparation or remediation. This often narrowed instructional focus to repetitive practice aligned with assessment formats, limiting opportunities for exploratory or conceptually rich mathematics.\u003c/p\u003e \u003cp\u003eTeachers also raised concerns about predictive features that labeled students as \u0026ldquo;at risk\u0026rdquo; based on performance data:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003eOnce a student is flagged, everything they do is seen through that lens. The system expects them to struggle, and so does the school.\u003c/em\u003e \u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eRather than supporting growth, automation sometimes reinforced static categorizations of competence, particularly among students from disadvantaged backgrounds. Teachers emphasized that while AI made certain processes more manageable, it did not replace the pedagogical work required to foster understanding, confidence, and mathematical identity.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e5.4. Cross-Cutting Observation: Mathematics Teachers as Mediators Under Constraint\u003c/h2\u003e \u003cp\u003eAcross all three themes, mathematics teachers positioned themselves as mediators attempting to reconcile equity-oriented commitments with tools that were not designed for their students\u0026rsquo; realities. While many teachers adapted or resisted aspects of AI systems to better support learners, their agency was constrained by limited time, resources, and institutional expectations.\u003c/p\u003e \u003cp\u003eAs one teacher summarized:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003eAI doesn\u0026rsquo;t\u003c/em\u003e\u0026rsquo;\u003cem\u003et decide what happens in my classroom\u0026mdash;but it definitely shapes it. And it shapes it more when you don\u003c/em\u003e\u0026rsquo;\u003cem\u003et have many other options.\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThese findings should not be read merely as a rejection of AI, but as a roadmap for more context-sensitive design. Teachers explicitly valued the technology when it operated as a low-stakes environment for repetition, allowing students to build foundational fluency without the social cost of public error. This suggests that the most equitable application of current AI in disadvantaged contexts may be its most modest: not as a substitute for teacher-led conceptual instruction, but as a dedicated tool for consolidating procedural knowledge. By designing systems that explicitly aim to support these bounded tasks\u0026mdash;automating routine feedback to liberate teacher time for relational intervention\u0026mdash;developers can realign AI tools with the actual needs of resource-constrained classrooms.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Discussion","content":"\u003cp\u003eThis study examined how AI-driven mathematics education is enacted in South African contexts of socio-economic disadvantage, foregrounding mathematics teachers\u0026rsquo; perspectives to interrogate dominant claims about personalization, efficiency, and equity. Data analysis identified three primary dynamics that challenge these claims: the tendency of current tools to disproportionately favor self-regulated learners, the pedagogical risk of student isolation inherent in personalized pathways, and the prioritization of procedural efficiency over conceptual understanding. By aligning these findings with the conceptual structure established in the literature review, this section situates the empirical results within broader debates in AI-in-education and critical mathematics education. Collectively, the data suggest that AI does not simply transform education through technical innovation; rather, it reshapes existing pedagogical and institutional dynamics, often reproducing inequality.\u003c/p\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e6.1. AI-Driven Transformation in Mathematics Education: Reform Without Redistribution\u003c/h2\u003e \u003cp\u003eAI is frequently framed as a transformative intervention capable of fundamentally reshaping teaching and learning. The findings of this study suggest that AI does transform mathematics education, but primarily by reconfiguring pedagogical relations, redistributing responsibility for learning, and reshaping what counts as legitimate mathematical activity. In disadvantaged contexts, these transformations occur within tightly constrained institutional environments, resulting in forms of change that prioritize efficiency, monitoring, and individualization rather than conceptual depth or relational support. The findings are in line with Ț\u0026icirc;ru et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), who argue that teachers' engagement with AI is influenced by institutional limits that lead to cautious, adaptive, or resistant practices, showing that AI-driven change tends to reshape teaching roles and responsibilities rather than deliver inherently progressive or equitable improvements in teaching and learning.\u003c/p\u003e \u003cp\u003eRather than enabling new pedagogical possibilities, AI systems in this study transformed existing practices by automating inequality-sensitive processes: accelerating the advancement of already-advantaged learners, individualizing struggle, and rendering underperformance more visible and administratively manageable. Transformation, therefore, did not take the form of innovation, but of intensification\u0026mdash;making longstanding patterns of stratification more efficient, scalable, and difficult to contest. This interpretation reinforces the literature's positioning of AI as a socio-technical system embedded within existing educational regimes, rather than as a neutral or autonomous driver of change.\u003c/p\u003e \u003cp\u003eImportantly, the findings demonstrate that low-level or rule-based AI systems can exert high-level socio-technical power. Even in the absence of advanced machine learning, these systems reshape educational practice by automating curricular pathways, stabilizing deficit classifications, and reorienting teacher attention toward data dashboards and performance signals. The transformative force of AI in these classrooms, therefore, derives not from predictive accuracy or adaptivity sophistication, but from the authority granted to algorithmic outputs within institutional decision-making processes. The findings align with D\u0026iacute;az and Nussbaum (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), who argue that AI's impact on education lies more in teaching skills than in technical ability. Their analysis shows that when AI lacks a teaching foundation, even basic systems can worsen existing practices. They automate curriculum progression and performance tracking, thereby altering teaching priorities and reinforcing negative classifications in institutional decision-making.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e6.2. The Self-Regulation Gap: How Algorithmic Design Amplifies Advantage\u003c/h2\u003e \u003cp\u003eThe finding that current AI tools privilege students with existing academic capital aligns with research demonstrating that AI-driven personalization benefits learners who can interpret feedback, persist independently, and navigate abstract learning pathways. Rather than compensating for unequal preparation, personalization often amplified existing differences by accelerating the progress of already-advantaged students while relegating others to repetitive or simplified tasks (Williamson, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Yaseen et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFrom an equity perspective, this suggests that personalization redistributes responsibility for learning onto students themselves, privileging those whose dispositions align with dominant academic norms. Mathematics teachers\u0026rsquo; experiences thus challenge the assumption that personalization is inherently equitable, highlighting the need for AI systems that explicitly account for socially patterned differences in learning conditions. In line with Yaseen et al. (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), the findings suggest that adaptive and AI-supported learning environments tend to benefit students with higher digital literacy and self-regulatory abilities, whilst providing limited support to those without these skills, thereby reinforcing existing educational inequalities rather than reducing them.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003e6.3. Beyond the Digital Divide: From Resource Access to Algorithmic Determinism\u003c/h2\u003e \u003cp\u003eWhile previous research on educational technology has established the Matthew Effect\u0026mdash;where students with prior advantages benefit disproportionately from digital tools\u0026mdash;the findings suggest that AI exacerbates this dynamic through mechanisms distinct from older, static technologies. Unlike traditional Computer-Assisted Instruction (CAI), which functioned as a passive repository of content, the AI systems observed here exercised algorithmic agency that actively reinforced achievement gaps. This occurred primarily through two AI-specific mechanisms:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eThe Remedial Loop (Algorithmic Containment)\u003c/b\u003e: High-performing students were accelerated by the algorithm, receiving validation and novel content. In contrast, lower-performing students were frequently trapped in adaptive logic, resulting in repetitive remedial loops. Unlike a static textbook, where a student can turn a page and attempt a harder problem, the opaque control logic of the AI prevented these students from accessing grade-level content, effectively codifying their low ability status into the software\u0026rsquo;s architecture.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePredictive Labeling (The Self-Fulfilling Prophecy)\u003c/b\u003e: The risk dashboards generated by these systems shifted the pedagogical gaze. Mathematics teachers reported that \u003cb\u003es\u003c/b\u003etudents' red/green coding influenced their expectations before instruction began. Unlike a grade book, this records past performance; these AI predictive models project future failures. This creates a barrier to algorithmic determinism, where the machine\u0026rsquo;s statistical inference of a student\u0026rsquo;s potential overrides the teacher\u0026rsquo;s belief in the student's capacity for growth. The findings align with previous research showing that AI tools can inadvertently reinforce inequities by advantaging students with prior advantages while constraining access for lower-performing disadvantaged students (Owan et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). From an equity-centered view, significant change entails inclusive AI design that lessens bias and reorganizes learning opportunities (Oyetade \u0026amp; Zuva, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Without such attention, algorithmic personalization risks amplifying existing structural disparities rather than fostering equitable mathematics learning (Al-Chaer, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2026\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec33\" class=\"Section2\"\u003e \u003ch2\u003e6.4. Personalization, Pedagogical Mediation, and Student Isolation\u003c/h2\u003e \u003cp\u003eThe observation that personalization manifests as solitary confinement in under-resourced classrooms confirms the study\u0026rsquo;s theoretical concern: that individualization often serves to strip away the relational scaffolding necessary for deep mathematical learning. In under-resourced classrooms, AI systems were often used to manage large class sizes or compensate for staffing shortages, reducing rather than enhancing relational engagement. Students who struggled academically were especially likely to disengage, guess through tasks, or withdraw silently, reinforcing patterns of marginalization. This finding underscores the importance of pedagogical mediation in AI-supported learning environments. Without sustained teacher interaction, personalization risks becoming a mechanism for individualization rather than inclusion, undermining the relational foundations of effective mathematics instruction.\u003c/p\u003e \u003cp\u003eDespite the broader challenges regarding equity and conceptual depth, it is important to acknowledge where these tools succeeded. Teachers consistently reported that current AI systems were highly effective for developing procedural fluency and foundational numeracy. For students who specifically needed repetition to master arithmetic operations or formulaic application, the immediate feedback loops provided a safe, non-judgmental space for practice. The AI's efficiency in these narrow domains successfully unburdened teachers from manual grading of rote tasks, theoretically freeing up time for higher-order instruction. The critical failure, however, was not in the tool\u0026rsquo;s ability to drill, but in the systemic over-reliance on this function as a proxy for comprehensive teaching. Consistent with Saleem et al. (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) and Zagami (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2026\u003c/span\u003e), these findings suggest that while AI-enabled personalization is often promoted as transformative for teaching, its impact really depends on the conditions at the institution. In constrained classrooms, personalization often manifests as individual task completion with little interaction, which may increase disengagement among already affected learners rather than encouraging their participation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003e6.5. Automation, Efficiency, and the Limits of Measurable Understanding\u003c/h2\u003e \u003cp\u003eWhile the efficiency gains identified in the findings (e.g., automated grading) reduced administrative labor, this study illuminates a critical tension: when efficiency is decoupled from pedagogical purpose\u0026mdash;measured through task completion, speed, or data visibility\u0026mdash;it begins to substitute for learning rather than support it. In such cases, what is optimized is not understanding, but manageability.\u003c/p\u003e \u003cp\u003eIn high-stakes accountability contexts, AI tools were frequently repurposed for test preparation and remediation, reinforcing procedural approaches to mathematics that research has long shown to disadvantage marginalized learners. Teachers expressed concern that while AI made classrooms more manageable, it did not support the kinds of mathematical engagement necessary for equitable learning (Vasquez et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Efficiency contributes to educational transformation only when it is deliberately subordinated to pedagogical goals such as conceptual understanding, relational engagement, and teacher judgment, rather than treated as a proxy for educational quality.\u003c/p\u003e \u003cp\u003eThis finding reinforces critiques of datafication in education, suggesting that efficiency gains should not be conflated with pedagogical improvement. In mathematics education, what is most valuable for learning is often least amenable to automation.\u003c/p\u003e \u003cp\u003eThese observations align with those of Hwang (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) and Dabingaya (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), who showed that AI-driven personalized learning environments can improve student achievement by tailoring instruction to individual learning speeds. Owan et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) reported that while AI improves procedural efficiency and enables low-stakes practice, its value for meaningful learning depends on the teacher's involvement. However, the increased efficiency in grading, monitoring, and automation does not automatically lead to a better understanding of concepts or fair participation in mathematics. Automated systems often focus on correctness and speed instead of reasoning, explanation, and the development of mathematical understanding.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec35\" class=\"Section2\"\u003e \u003ch2\u003e6.6. Mathematics teachers as Constrained Human-in-the-Loop Mediators\u003c/h2\u003e \u003cp\u003eAcross all findings, teachers emerged as critical mediators navigating the tensions between AI systems, institutional constraints, and equity-oriented commitments. Teachers actively interpreted, adapted, and sometimes resisted AI tools to support their students. However, their agency was consistently shaped\u0026mdash;and often limited\u0026mdash;by resource scarcity, prescriptive curricula, and accountability pressures.\u003c/p\u003e \u003cp\u003eRather than positioning inequitable outcomes as failures of implementation, the findings suggest that responsibility for equity is frequently displaced onto teachers, even when AI systems are poorly aligned with the realities of disadvantaged classrooms. This displacement obscures the roles of design, governance, and upstream policy decisions. Recognizing teachers as constrained human-in-the-loop actors underscores the need for AI systems developed in dialogue with educators and evaluated in the contexts in which they are used. Consistent with Ț\u0026icirc;ru et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), the findings show that teachers often adopt critical, adaptive, or resistant views toward AI tools, yet their ability to act is limited by institutional pressures, resource constraints, and misaligned technology. As a result, the responsibility for equitable learning outcomes is shifted from AI systems to teachers.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec36\" class=\"Section2\"\u003e \u003ch2\u003e6.7. Reframing Equity in AI-Driven Mathematics Education\u003c/h2\u003e \u003cp\u003eClaims about educational transformation through AI rest implicitly on assumptions about what counts as improvement and for whom such improvement matters. In this study, equity is not treated as an optional normative preference or a secondary social goal, but as a necessary condition for meaningful educational transformation. Transformative change in mathematics education cannot be said to occur if new systems systematically advantage learners who already possess academic, institutional, or cultural resources while rendering others more visible, more monitored, and more constrained. In unequal educational systems, technologies that improve outcomes for some while stabilizing or deepening disadvantage for others constitute optimization rather than transformation.\u003c/p\u003e \u003cp\u003eFrom this perspective, equity functions as a diagnostic lens for evaluating whether AI-driven change alters the underlying distribution of learning opportunity or merely reconfigures its management. Without explicit attention to equity, AI systems risk being celebrated as transformative on the basis of efficiency gains or aggregate performance improvements, even as they reproduce stratified learning trajectories. Treating equity as foundational, therefore, shifts the evaluative question from whether AI works to how it reorganizes opportunity across learners and contexts. In line with Oyetade and Zuva (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), the findings indicate that claims of AI-driven transformation should be assessed with a focus on equity, ensuring that technologies do not maximize outcomes for already advantaged students but rather actively reallocate learning opportunities and enhance teacher capacity to support all students.\u003c/p\u003e \u003c/div\u003e"},{"header":"7. Future Directions: Towards Equity-Centred AI in Mathematics","content":"\u003cp\u003eTo move beyond the limitations observed in this study, the design and implementation of educational AI must undergo three fundamental shifts:\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003cstrong\u003e7. 1. From Opaque Prediction to Explainable Diagnostics.\u003c/strong\u003e Currently, the predictive nature of AI acts as a black box that labels students. Future development should prioritize Explainable AI (XAI). Instead of simply flagging a student as at risk (a label that lowers teacher expectations), systems should provide granular, transparent data on why a misconception is occurring (e.g., Student consistently misapplies the distributive property). This shifts the AI\u0026rsquo;s role from a content gatekeeper to a diagnostic partner that empowers the teacher to intervene precisely.\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003cstrong\u003e7. 2. From Isolation to Human-in-the-Loop Orchestration.\u003c/strong\u003e The model of personalized learning as a solitary student facing a screen must be abandoned in favor of pedagogical orchestration. Future directions should focus on AI tools designed to prompt teacher-student interaction rather than replace it. For example, rather than silently routing a struggling student to a remedial video, the system should signal the teacher to facilitate a small-group intervention. The metric of success for EdTech should not be time on system, but the quality of off-screen mathematical discourse the system inspires.\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003cstrong\u003e7. 3. Designing for the Margins, Not the Mean.\u003c/strong\u003e Finally, developers must move beyond designing for the ideal self-regulated learner. AI systems must include scaffolding explicitly designed for students with lower self-regulation and higher anxiety. This means moving away from frustration-inducing remedial loops and incorporating metacognitive prompts that help students plan their learning strategies. Equity cannot be an afterthought; it must be a core design parameter of the algorithm itself. In this sense, equity and efficiency are not competing values, but operate in a hierarchical relationship: efficiency can enable transformation only when equity defines its direction and limits.\u003cbr\u003e\u003c/span\u003e\u003c/p\u003e"},{"header":"8. Conclusion","content":"\u003cp\u003eThis study set out to examine how AI-driven mathematics education is experienced and enacted in contexts of socio-economic disadvantage, foregrounding mathematics teachers\u0026rsquo; perspectives to interrogate dominant narratives of personalization, efficiency, and transformation in AI\u0026ndash;supported education. The findings demonstrate that AI transforms mathematics education not solely through instructional intelligence, but also through its capacity to reorganize time, attention, accountability, and pedagogical responsibility within already unequal systems.\u003c/p\u003e \u003cp\u003eAcross the analysis, three interrelated dynamics emerged. First, AI-driven personalization tended to benefit students who already possessed the dispositions, confidence, and self-regulatory skills required to navigate autonomous learning environments. Rather than compensating for unequal preparation, adaptive systems often accelerated existing hierarchies of achievement. Second, personalization without sustained pedagogical and relational support often led to student isolation, particularly among learners who needed the most guidance. Third, automation increased efficiency in assessment and monitoring but did little to support conceptual understanding or equitable participation in mathematics learning. Together, these dynamics suggest that the transformative potential of AI in education is sharply limited when equity is treated as a downstream outcome rather than a foundational design principle.\u003c/p\u003e \u003cp\u003eBy centering mathematics teachers working in disadvantaged contexts, this study contributes a grounded account of how AI systems operate in everyday classroom practice. Mathematics teachers emerged not as passive implementers of technology, but as constrained mediators attempting to reconcile equity-oriented commitments with tools designed primarily for efficiency, scalability, and measurability. Importantly, inequitable outcomes were not attributable to teacher resistance or poor implementation, but to misalignments between AI system design and the social and pedagogical realities of marginalized educational settings.\u003c/p\u003e \u003cp\u003eThese findings carry implications for AI-in-education research, policy, and design. For researchers, the study underscores the need to move beyond system-centered evaluations toward analyses that attend to context, pedagogy, and power. Claims about transformation should be assessed not only in terms of technical performance or short-term learning gains, but in relation to how AI reshapes learning opportunities, teacher agency, and student participation over time. For policymakers, the results highlight the risks of deploying AI as a cost-effective substitute for instructional support in under-resourced schools and underscore the need for governance frameworks that foreground equity, transparency, and accountability. For designers and developers, the findings suggest that equity-oriented AI must support relational pedagogy, make assumptions about learning explicit, and expand\u0026mdash;rather than constrain\u0026mdash;mathematics teachers\u0026rsquo; professional judgment.\u003c/p\u003e \u003cp\u003eMore broadly, this study argues that the central question for AI in education is not whether it can personalize learning, but under what conditions, and for whom, such personalization contributes to meaningful educational opportunity. Without deliberate, equity-centered intervention, AI risks rendering longstanding inequalities more efficient, more automated, and less visible. Conversely, when guided by critical attention to context, pedagogy, and justice, AI has the potential to support\u0026mdash;not replace\u0026mdash;the human relationships at the core of equitable mathematics education. Contexts of disadvantage should therefore be understood not as outliers in AI-in-education research, but as sites where the structural consequences of algorithmic design become particularly clear.\u003c/p\u003e \u003cp\u003eIn conclusion, transforming education through AI requires more than sophisticated algorithms or scalable platforms. It requires a reorientation of AI development and deployment toward the social conditions of learning, mathematics teachers' professional expertise, and the diverse ways students engage with mathematics. Only by treating equity as a starting point rather than an afterthought can AI contribute to genuine educational transformation. This study relied on teacher reporting of student engagement; future research should triangulate these findings with direct student feedback and platform interaction logs.\u003c/p\u003e "},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eContributions:\u0026nbsp;\u003c/strong\u003eB C is the sole contributor to this article\u0026rsquo;s conceptualisation, data generation, analysis, and write-up.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e The author did not receive support from any organisation for the submitted work\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u0026nbsp;\u003c/strong\u003eThe author declares no competing interests\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and material:\u0026nbsp;\u003c/strong\u003eAny requests for data should be addressed to the author, who will share in accordance with what is allowed by the signed consent forms\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number\u003c/strong\u003e: Not applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and Consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEthical clearance for this study was applied for and approved by the CPUT Health and Wellness Sciences (Protocol Number: CPUT/HWS-REC 2025/\u003cem\u003eS1\u003c/em\u003e). I confirm that all methods were carried out in accordance with the University of Johannesburg\u0026rsquo;s ethical guidelines and regulations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed Consent:\u0026nbsp;\u003c/strong\u003eWritten informed consent was obtained from Undergraduate students before participation in the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to publish:\u003c/strong\u003e The Author consents to the article being published.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAl-Chaer, E.D. (2026). Mind the Algorithm: Charting a Responsible Course for AI in Higher Education. In: Badran, A., Baydoun, E., Hillman, S., Mesmar, J. (eds) Higher Education in the Arab World. Springer, Cham. https://doi.org/10.1007/978-3-031-99068-7_3\u003c/li\u003e\n\u003cli\u003eAu, W. (2016). Meritocracy 2.0: High-stakes, standardized testing as a racial project of neoliberal multiculturalism. \u003cem\u003eEducational Policy\u003c/em\u003e, \u003cem\u003e30\u003c/em\u003e(1), 39\u0026ndash;62.\u003c/li\u003e\n\u003cli\u003eAzevedo, R., Taub, M., \u0026amp; Mudrick, N. V. (2018). Understanding and reasoning about real-time cognitive, affective, and metacognitive processes to foster self-regulation with advanced learning technologies. In D. H. Schunk \u0026amp; J. A. Greene (Eds.), \u003cem\u003eHandbook of self-regulation of learning and performance\u003c/em\u003e (2nd ed., pp. 256\u0026ndash;270). Routledge.\u003c/li\u003e\n\u003cli\u003eBenjamin, R. (2019).\u003cem\u003e Race after technology: Abolitionist tools for the new Jim Code.\u003c/em\u003e Polity\u003c/li\u003e\n\u003cli\u003eBiesta, G. (2015). \u003cem\u003eThe beautiful risk of education\u003c/em\u003e. Paradigm Publishers.\u003c/li\u003e\n\u003cli\u003eBoaler, J. (2016).\u003cem\u003e Mathematical mindsets: Unleashing students\u0026apos; potential through creative math, inspiring messages and innovative teaching. \u003c/em\u003eJossey-Bass.\u003c/li\u003e\n\u003cli\u003eBourdieu, P., \u0026amp; Passeron, J. C. (1990). \u003cem\u003eReproduction in education, society and culture\u003c/em\u003e. Sage Publications.\u003c/li\u003e\n\u003cli\u003eCobb, P., Gresalfi, M., \u0026amp; Hodge, L. L. (2009). An interpretive scheme for analyzing the identities that students develop in mathematics classrooms. \u003cem\u003eJournal for Research in Mathematics Education\u003c/em\u003e, \u003cem\u003e40\u003c/em\u003e(1), 40\u0026ndash;68.\u003c/li\u003e\n\u003cli\u003eDabingaya, M. (2022). Analyzing the effectiveness of AI-powered adaptive learning platforms in mathematics education. \u003cem\u003eInterdisciplinary Journal Papier Human Review, 3\u003c/em\u003e(1), 1\u0026ndash;7. https://doi.org/10.47667/ijphr.v3i1.226.\u003c/li\u003e\n\u003cli\u003eDarling-Hammond, L. (2010). \u003cem\u003eThe flat world and education: How America\u0026apos;s commitment to equity will determine our future. \u003c/em\u003eTeachers College Press. \u003c/li\u003e\n\u003cli\u003eD\u0026iacute;az, B., \u0026amp; Nussbaum, M. (2024). \u003cem\u003eArtificial intelligence for teaching and learning in schools: The need for pedagogical intelligence\u003c/em\u003e. \u003cem\u003eComputers \u0026amp; Education, 217\u003c/em\u003e(105071). https://doi.org/10.1016/j.compedu.2024.105071\u003c/li\u003e\n\u003cli\u003eEubanks, V. (2018). \u003cem\u003eAutomating inequality: How high-tech tools profile, police, and punish the poor.\u003c/em\u003e St. Martin\u0026apos;s Press.\u003c/li\u003e\n\u003cli\u003eHolmes, W., Bialik, M., \u0026amp; Fadel, C. (2019). \u003cem\u003eArtificial intelligence in education: Promises and implications for teaching and learning. \u003c/em\u003eCenter for Curriculum Redesign.\u003c/li\u003e\n\u003cli\u003eHwang, S. (2022). Examining the effects of artificial intelligence on Elementary Students\u0026rsquo; mathematics achievement: A meta-analysis.\u003cem\u003e Sustainability, 14\u003c/em\u003e(20), 13185. https://doi.org/10.3390/su142013185.\u003c/li\u003e\n\u003cli\u003eKizilcec, R. F., P\u0026eacute;rez-Sanagust\u0026iacute;n, M., \u0026amp; Maldonado, J. J. (2017). Self-regulated learning strategies predict learner behavior and goal attainment in Massive Open Online Courses. \u003cem\u003eComputers \u0026amp; Education\u003c/em\u003e, \u003cem\u003e104\u003c/em\u003e, 18\u0026ndash;33.\u003c/li\u003e\n\u003cli\u003eKnox, J. (2020). Artificial intelligence and education in China. \u003cem\u003eLearning, Media and Technology\u003c/em\u003e, \u003cem\u003e45\u003c/em\u003e(3), 298\u0026ndash;311.\u003c/li\u003e\n\u003cli\u003eLuckin, R., Holmes, W., Griffiths, M., \u0026amp; Forcier, L. B. (2016). \u003cem\u003eIntelligence unleashed: An argument for AI in education\u003c/em\u003e. Pearson.\u003c/li\u003e\n\u003cli\u003eNasir, N. S., Cabana, C., Shreve, B., Woodbury, E., \u0026amp; Louie, N. (2014). \u003cem\u003eMathematics for equity: A framework for successful practice\u003c/em\u003e. Teachers College Press.\u003c/li\u003e\n\u003cli\u003eO\u003cspan dir=\"RTL\"\u003e\u0026rsquo;\u003c/span\u003eNeil, C. (2016).\u003cem\u003e Weapons of math destruction: How big data increases inequality and threatens democracy.\u003c/em\u003e Crown.\u003c/li\u003e\n\u003cli\u003eOwan, V. J., Abang, K. B., Idika, D. O., Etta, E. O., \u0026amp; Bassey, B. A. (2023). Exploring the potential of artificial intelligence tools in educational measurement and assessment. \u003cem\u003eEurasia Journal of Mathematics, Science and Technology Education, 19\u003c/em\u003e(8), em2307. https://doi.org/10.29333/ejmste/13428\u003c/li\u003e\n\u003cli\u003eOyetade, K., \u0026amp; Zuva, T. (2025). Advancing equitable education with inclusive AI to mitigate bias and enhance teacher literacy. \u003cem\u003eEducational Process: International Journal, 14, \u003c/em\u003e(e2025087). https://doi.org/10.22521/edupij.2025.14.87\u003c/li\u003e\n\u003cli\u003ePriestley, M., Biesta, G., \u0026amp; Robinson, S. (2015). \u003cem\u003eTeacher agency: An ecological approach\u003c/em\u003e. Bloomsbury Publishing.\u003c/li\u003e\n\u003cli\u003eSaleem, S., Aziz, M. U., Iqbal, M. J., \u0026amp; Abbas, S. (2025). AI in education: Personalized learning systems and their impact on student performance and engagement. \u003cem\u003eThe Critical Review of Social Sciences Studies, 3\u003c/em\u003e(1), 2445\u0026ndash;2459. https://thecrsss.com/index.php/Journal/about\u003c/li\u003e\n\u003cli\u003eSelwyn, N. (2016). \u003cem\u003eIs technology good for education?\u003c/em\u003e Polity.\u003c/li\u003e\n\u003cli\u003eSelwyn, N., Hillman, T., Eynon, R., Ferreira, G., Knox, J., Macgilchrist, F., \u0026amp; Sancho-Gil, J. M. (2020). What\u0026apos;s next for Ed-Tech? Critical hopes and concerns for the 2020s. \u003cem\u003eLearning, Media and Technology\u003c/em\u003e, \u003cem\u003e45\u003c/em\u003e(1), 1\u0026ndash;6.\u003c/li\u003e\n\u003cli\u003eSfard, A. (2008). \u003cem\u003eThinking as communicating: Human development, the growth of discourses, and mathematizing\u003c/em\u003e. Cambridge University Press.\u003c/li\u003e\n\u003cli\u003eSkovsmose, O. (2011).\u003cem\u003e An invitation to critical mathematics education.\u003c/em\u003e Sense Publishers.\u003c/li\u003e\n\u003cli\u003eȚ\u0026icirc;ru, L. G., Gherheș, V., Stoicov, I., \u0026amp; Stanici, M. (2025). Not Ready for AI? Exploring Teachers\u0026rsquo; Negative Attitudes Toward Artificial Intelligence. \u003cem\u003eSocieties, 15\u003c/em\u003e(12), 337. https://doi.org/10.3390/soc15120337\u003c/li\u003e\n\u003cli\u003evan Dijk, J. (2020). \u003cem\u003eThe digital divide\u003c/em\u003e. Polity.\u003c/li\u003e\n\u003cli\u003eVasquez, E., Basham, J. D., Jimenez, B., \u0026amp; Marino, M. T. (2024). Ethical considerations for educators leveraging artificial intelligence. In CIDDL (Eds.), \u003cem\u003eInclusive intelligence: The impact of AI on education for all learners\u003c/em\u003e (pp. 92\u0026ndash;98). Center for Innovation, Design, and Digital Learning.\u003c/li\u003e\n\u003cli\u003eWilliamson, B. (2017).\u003cem\u003e Big data in education: The digital future of learning, policy and practice. \u003c/em\u003eSage.\u003c/li\u003e\n\u003cli\u003eWilliamson, B. and Eynon, R. (2020) Historical Threads, Missing Links, and Future Directions in AI in Education.\u003cem\u003e Learning, Media and Technology, 45\u003c/em\u003e, 223-235.\u003cbr\u003ehttps://doi.org/10.1080/17439884.2020.1798995\u003c/li\u003e\n\u003cli\u003eYaseen, H., Mohammad, A. S., Ashal, N., Abusaimeh, H., Ali, A., \u0026amp; Sharabati, A.-A. A. (2025). The Impact of Adaptive Learning Technologies, Personalized Feedback, and Interactive AI Tools on Student Engagement: The Moderating Role of Digital Literacy. \u003cem\u003eSustainability, 17\u003c/em\u003e(3), 1133. https://doi.org/10.3390/su17031133\u003c/li\u003e\n\u003cli\u003eZagami, J. (2026). Addressing the Dark Side of Differentiation: Bias and Micro-Streaming in Artificial Intelligence Facilitated Lesson Planning. \u003cem\u003eInformation\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(1), 12. https://doi.org/10.3390/info17010012\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Artificial Intelligence in Education (AIED), Algorithmic Equity, Teacher Agency, Educational Data Mining, Mathematics Education, Global South","lastPublishedDoi":"10.21203/rs.3.rs-8679398/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8679398/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eArtificial intelligence (AI) is increasingly positioned as a transformative force in education, with particular promise for addressing persistent challenges in mathematics education, such as low achievement, limited instructional capacity, and unequal access to quality teaching. AI-driven systems\u0026mdash;including adaptive learning platforms, intelligent tutoring systems, and automated assessment tools\u0026mdash;are widely promoted as mechanisms for personalizing instruction at scale and advancing educational equity. This study critically examines these claims by investigating how AI-mediated mathematics education is enacted in South Africa\u0026rsquo;s schools located in contexts of socio-economic disadvantage. Drawing on qualitative interviews with mathematics teachers working across diverse institutional settings, the study explores how AI tools are selected, implemented, and experienced in everyday classroom practice. The findings reveal that, rather than functioning as neutral or equalizing technologies, current AI systems often amplify existing inequalities by disproportionately benefiting students who already possess strong self-regulatory skills, institutional support, and prior mathematical confidence. Personalization without sustained pedagogical and relational support often leads to observed student isolation, while automation prioritizes efficiency and monitoring over conceptual understanding. By foregrounding mathematics teachers\u0026rsquo; perspectives, the study demonstrates that AI transforms educational practice primarily through the automation of pedagogical decision-making, the redistribution of learning responsibility, and the reconfiguration of what counts as legitimate mathematical activity. It argues that meaningful educational transformation through AI requires treating equity as a foundational design and governance condition, with efficiency serving as an instrumental outcome rather than a proxy for pedagogical improvement.\u003c/p\u003e","manuscriptTitle":"The role of Artificial Intelligence in amplifying educational inequalities in resource constrained mathematics classrooms","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-17 10:17:24","doi":"10.21203/rs.3.rs-8679398/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d7e8bc17-55c1-4ce4-bc9c-345367789197","owner":[],"postedDate":"February 17th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-29T07:40:41+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-17 10:17:24","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8679398","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8679398","identity":"rs-8679398","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}