This paper examines the transformative potential and critical challenges of integrating generative artificial intelligence into adaptive educational systems, advancing a human-centered framework that prioritizes augmentation over automation. Through analysis of empirical evidence from large-scale implementations, theoretical foundations spanning cognitive science and learning theory, and emerging technological capabilities, we demonstrate that successful educational AI requires fundamental reconceptualization of technology's role in learning. Our analysis reveals that implementations achieving 25-60% learning improvements share common characteristics: pedagogical primacy in system design, human-in-the-loop architectures maintaining educator oversight, transparency mechanisms enabling stakeholder understanding, and equity-first approaches addressing systemic inequalities. The paper introduces a four-phase implementation roadmap progressing from stakeholder discovery through controlled evaluation to scaled deployment with appropriate governance structures. We identify critical challenges including hallucination rates exceeding 8% in educational contexts, cognitive ofloading efects reducing independent problem-solving by 35%, and algorithmic bias amplifying existing educational inequities. The human-centered framework proposed addresses these challenges through four foundational principles: pedagogical primacy ensuring learning science drives technology deployment, human-in-the-loop requirements maintaining essential oversight, transparency by design enabling stakeholder understanding, and equity-first approaches proactively addressing accessibility and bias. Looking toward 2025-2030, we examine emerging technologies including emotion-aware adaptation, neuro-symbolic AI integration, federated learning architectures, and quantum computing applications, alongside pedagogical evolution encompassing meta-learning capabilities, immersive AR/VR integration, and neuroadaptive systems. The paper concludes with an urgent call to action for stakeholders across the educational ecosystem, articulating a vision where technology amplifies human capabilities rather than replacing them, democratizes quality education while preserving local values, and enhances rather than erodes human agency. This synthesis provides essential guidance for educators, technologists, policymakers, and researchers navigating the complex terrain of AI-enhanced education while maintaining unwavering commitment to human dignity and learner wellbeing.
characterize modern classrooms [
The digital divide compounds these challenges, with substantial disparities in technology access
creating additional layers of educational inequality. Recent evidence indicates that students in underserved
regions face compounded disadvantages – not merely from limited access to digital infrastructure but
from the cascading efects on educational opportunities and outcomes [
Simultaneously, the convergence of generative artificial intelligence, multimodal learning models,
and adaptive technologies presents an unprecedented opportunity to transcend these limitations.
The rapid maturation of large language models (LLMs) and retrieval-augmented generation (RAG)
architectures has fundamentally altered the technological landscape of educational content generation.
Unlike previous generations of educational technology that merely digitized existing content, current
generative AI systems demonstrate capabilities for creating truly personalized, contextually aware
instructional materials in real-time [
This technological inflection point coincides with a pedagogical maturation in understanding how to integrate AI systems efectively within educational frameworks. The evolution from rule-based systems of the 1970s through data-driven approaches of the 2000s to today’s generative models represents not merely technical advancement but a fundamental reconceptualization of how technology can support human learning.
This paper advances a clear position: generative AI’s transformative value in education lies not in automating instruction but in creating a new paradigm where technology amplifies human pedagogical capabilities. We argue for an augmentation framework that positions AI as an “exoskeleton” for educators – enhancing their reach, personalizing their impact, and liberating them from administrative burdens to focus on uniquely human aspects of teaching such as mentorship, emotional support, and creative inspiration.
The central thesis distinguishes between automation and augmentation as fundamentally diferent approaches to educational AI integration. Automation seeks to replace human instructors with algorithmic systems, treating education as an information transfer problem amenable to technical optimization. This perspective, while technologically appealing, fundamentally misunderstands the nature of learning as a deeply social, emotional, and contextual process. Augmentation, conversely, recognizes AI as a powerful tool that extends human capabilities without displacing the essential human elements of education.
Our framework proposes that successful educational AI implementation requires adherence to four core principles. First, pedagogical primacy demands that learning science drives technological implementation rather than technology determining pedagogical approaches. Second, human-in-theloop requirements ensure educator oversight for high-stakes decisions afecting student trajectories. Third, transparency by design makes AI decision-making processes interpretable to educators and learners. Fourth, equity-first approaches prioritize accessibility and bias mitigation as fundamental design requirements rather than post-hoc considerations.
This position emerges from synthesis of empirical evidence across diverse implementations, theoretical frameworks spanning cognitive science and learning theory, and practical insights from large-scale deployments. The evidence suggests that when designed with these principles, AI-powered adaptive systems can address longstanding educational challenges while avoiding the pitfalls of techno-solutionism that has characterized previous educational technology waves.
This review synthesizes evidence from multiple sources to construct an integrated understanding of adaptive educational content generation using generative AI.
The analysis framework integrates three complementary perspectives. Technical architecture analysis examines system designs, algorithmic approaches, and implementation patterns across platforms including GPT-5-based tutoring systems, multimodal content generators, and hybrid RAG architectures. Pedagogical efectiveness evaluation synthesizes quantitative outcomes from randomized controlled trials, quasi-experimental studies, and large-scale deployments measuring learning gains, engagement metrics, and retention rates. Ethical and societal impact assessment analyzes issues of algorithmic bias, privacy implications, digital divide efects, and long-term cognitive development considerations.
The geographic scope encompasses implementations across North America, Europe, Asia, and emerging deployments in Latin America and Africa, providing insights into contextual variations and cultural considerations.
Limitations of this review include the rapidly evolving nature of generative AI technology, which means some findings may require updating as capabilities advance. The predominance of studies from well-resourced contexts may limit generalizability to under-resourced educational settings. Long-term cognitive and developmental impacts remain largely unmeasured due to the recency of large-scale deployments. Additionally, publication bias toward positive outcomes may underrepresent implementation failures or negative consequences.
The trajectory of adaptive learning technologies reveals a profound transformation in how educational systems conceptualize and implement personalization. This evolution, spanning five decades (figure 1), demonstrates not merely technological advancement but fundamental reconceptualizations of learning itself, each generation addressing limitations of its predecessors while introducing novel capabilities and challenges.
First generation Rule-based systems Expert-defined rules
Branching paths Limited adaptation
Second generation
Data-driven ML Bayesian knowledge tracing
Collaborative filtering
Neural networks
Third generation
Generative AI LLMs and transformers
RAG architectures
Multimodal generation
Initial adaptive learning systems emerged from behaviorist principles and early artificial intelligence
research, implementing expert-defined rules that mapped learner characteristics to instructional
strategies. These systems operated through deterministic decision trees, where pedagogical experts encoded
instructional logic into if-then statements [
The theoretical underpinnings drew heavily from programmed instruction and mastery learning
frameworks. Systems like PLATO (Programmed Logic for Automatic Teaching Operations) and early
intelligent tutoring systems embodied assumptions about linear knowledge acquisition and discrete
learning states. Content adaptation occurred primarily through branching narratives – correct responses
advanced students to more complex material, while incorrect responses triggered remedial loops [
These systems faced substantial limitations that constrained their educational impact. Static rule
sets could not accommodate the full spectrum of learner variability, resulting in crude personalization
that often misaligned with individual needs. The knowledge engineering bottleneck required extensive
expert time to encode domain knowledge and pedagogical strategies, making system development
prohibitively expensive. Furthermore, limited computational resources restricted systems to simple
learner models tracking only basic performance metrics, while integration challenges with existing
educational infrastructure prevented widespread adoption [
Bayesian knowledge tracing revolutionized learner modeling by treating knowledge states as hidden
variables inferred from observable performance. The framework modeled four key probabilities: initial
knowledge ( (0)), learning rate ( ( )), guess probability ( ()), and slip probability ( ()), enabling
systems to maintain uncertainty estimates about student knowledge and make probabilistic predictions
about future performance. This probabilistic approach proved particularly efective for skill-focused
domains like mathematics and programming [
Reinforcement learning emerged as another powerful paradigm, treating instructional sequencing as a sequential decision problem. Systems learned policies that maximized long-term learning outcomes through exploration and exploitation, discovering non-obvious instructional strategies that outperformed expert-designed sequences. The integration of clustering algorithms enabled identification of learner archetypes from behavioral patterns, facilitating group-based personalization when individual data remained sparse.
Despite these advances, second-generation systems encountered new challenges. The cold start
problem meant systems required substantial interaction data before efective personalization,
disadvantaging early users. Interpretability issues arose as machine learning models became black boxes, making
it dificult for educators to understand or trust adaptation decisions. Privacy concerns intensified as
systems collected increasingly granular learner data, raising questions about surveillance and autonomy
in educational contexts [
Modern architectures combine multiple sophisticated components. Multi-agent systems orchestrate
specialized models for diferent educational tasks, retrieval-augmented generation grounds responses in
verified knowledge sources, and transformer-based language models provide contextual understanding
and generation capabilities. Moderator mechanisms ensure quality and safety, while bidirectional
planning frameworks enable dynamic instructional sequencing. The integration of multimodal data –
text, images, audio, and even physiological signals – creates holistic learner profiles that capture
cognitive, afective, and behavioral dimensions of learning [
The evolution of adaptive learning technologies cannot be understood purely through technical advancement; pedagogical theories have co-evolved with technological capabilities, creating a dynamic interplay between what is technically possible and what is educationally desirable.
Early adaptive systems reflected behaviorist assumptions about learning as stimulus-response
conditioning. Content was atomized into discrete units, feedback emphasized correctness, and adaptation
meant adjusting dificulty or repetition frequency. This framework proved efective for procedural
knowledge and skill acquisition but struggled with conceptual understanding and transfer [
Constructivist principles gradually permeated adaptive system design, reconceptualizing learners
as active knowledge builders rather than passive recipients. Systems began supporting exploratory
learning, multiple solution paths, and collaborative knowledge construction. The shift manifested
in features like open-ended problem spaces, tools for hypothesis testing and experimentation, and
scafolding that faded as competence developed. Adaptive educational hypermedia systems exemplified
this transition, blending cognitivist attention to mental models with constructivist emphasis on active
engagement [
Contemporary frameworks embrace heutagogy – self-determined learning where learners control
not just pace but also learning goals, methods, and assessment criteria. This paradigm recognizes that
in rapidly changing knowledge domains, the capacity for self-directed learning supersedes specific
content mastery. Adaptive systems supporting heutagogical approaches provide learner dashboards for
metacognitive awareness, recommendation engines that suggest rather than prescribe, and tools for
learners to create and share content [
The Innovation Fellowship study illuminated how heutagogical principles manifest in practice.
Participants emphasized the importance of “structure of fluidity” – suficient scafolding to prevent
overwhelming freedom while maintaining autonomy for exploration and creativity. Successful
implementations balance structured guidance with learner agency, creating environments where adaptation
occurs bidirectionally: systems adapt to learners while learners develop adaptive expertise [
Digital environments necessitate new learning theories that account for distributed cognition and
networked knowledge. Connectivism posits that learning involves forming connections across information
nodes, with knowledge residing in networks rather than individuals. Adaptive systems incorporating
connectivist principles facilitate social learning through peer matching algorithms, aggregate collective
intelligence for content recommendations, and adapt based on network-level patterns beyond individual
behaviors [
Implementation challenges persist, particularly in K-12 contexts where curriculum constraints and assessment requirements conflict with connectivist openness. Successful translations involve structured exploration within bounded domains, scafolded network navigation skills, and hybrid models combining individual and collective adaptation. The integration requires reconceptualizing adaptive systems not as isolated tutors but as facilitators within learning ecosystems.
The arrival of large language models marks an inflection point where technical capabilities align with sophisticated pedagogical requirements. This convergence manifests in three transformative shifts that fundamentally alter the landscape of adaptive education:
Previous adaptive systems operated within finite content libraries, selecting and sequencing
predetermined materials. Generative AI transcends this limitation through real-time content creation,
producing explanations tailored to individual misconceptions, problems calibrated to precise dificulty
levels, and feedback addressing specific error patterns. This shift from selection to generation enables
truly individualized instruction previously impossible at scale [
The generative capacity extends beyond text to multimodal content creation. Systems now
produce diagrams illustrating abstract concepts, animations demonstrating procedures, and even audio
explanations for auditory learners. This multimodal generation addresses diverse learning preferences
while maintaining pedagogical coherence across modalities. The DALL-E and Stable Difusion
integrations in educational platforms demonstrate how visual generation enhances conceptual understanding,
particularly in STEM domains requiring spatial reasoning [
Pure generation risks hallucination – plausible but incorrect content that misleads learners.
Retrievalaugmented generation addresses this critical limitation by grounding generative models in verified
knowledge sources. Systems first retrieve relevant information from curated educational databases,
then use this retrieved content to constrain and inform generation, ensuring factual accuracy while
maintaining personalization benefits [
The RAG architecture proves particularly valuable for domain-specific education where accuracy is
paramount. Medical education systems retrieve from peer-reviewed journals, mathematics platforms
reference theorem databases, and history applications draw from primary sources. This hybrid approach
balances the flexibility of generation with the reliability of curated content, creating systems that are
both adaptive and trustworthy [
Transformer architectures enable unprecedented contextual understanding, maintaining coherence
across extended educational interactions. The self-attention mechanism allows models to recognize
conceptual dependencies, track learning progressions, and identify subtle misconceptions. This deep
contextual awareness supports sophisticated pedagogical strategies previously requiring human
expertise [
Systems leverage context windows exceeding 200,000 tokens to maintain comprehensive learning histories, enabling long-term personalization that accounts for growth trajectories, recurring error patterns, and evolving interests. The extended context facilitates complex instructional strategies like spiral curricula, where concepts resurface with increasing sophistication, and transfer learning, where systems recognize opportunities to connect new material with prior knowledge across domains.
The technical-pedagogical convergence represents more than technological progress; it embodies a new educational paradigm where artificial intelligence serves not as a replacement for human instruction but as an amplifier of pedagogical expertise. By combining generative flexibility, knowledge grounding, and contextual awareness, current systems approach the adaptive capacity of expert human tutors while operating at unprecedented scale.
The empirical landscape of generative AI in adaptive education reveals consistent patterns of substantial
improvement across multiple dimensions, challenging assumptions about the limits of
technologyenhanced learning. Analysis of implementations spanning 2015-2024 demonstrates not isolated successes
but systematic enhancements in learning outcomes, engagement metrics, and economic eficiency that
warrant serious consideration for widespread adoption.
3.1.1. Learning outcomes: beyond incremental gains
Contemporary evidence transcends the modest improvements typical of previous educational
technologies, revealing transformative potential when AI-powered adaptive systems align with pedagogical
principles. Meta-analytic evidence across undergraduate engineering education reports efect sizes
ranging from 0.43 to 0.70, representing medium to large impacts on academic achievement [
The consistency of these improvements across contexts proves particularly noteworthy. Analysis of over 50 empirical studies reveals 15-35% average improvement in academic performance, with some implementations achieving even more dramatic results. Squirrel AI’s nano-level personalization framework reduces learning time by 60% while maintaining or exceeding traditional outcome levels, suggesting eficiency gains compound direct learning improvements. Knowledge retention shows similarly impressive patterns, with 30% or greater increases in long-term retention compared to traditional instruction methods [18].
Subject-specific analyses reveal diferential efectiveness patterns that inform deployment strategies. Mathematics and quantitative disciplines show the strongest efects, with ALEKS demonstrating 27% improvement in college algebra success rates at Arizona State University. Language learning platforms like Duolingo Max achieve 45% better retention rates through multimodal engagement and adaptive practice. Sciences benefit from visualization capabilities and adaptive laboratory simulations, while humanities applications excel in personalized writing feedback and contextual content generation [19]. 3.1.2. Engagement metrics: sustaining motivation at scale Engagement improvements prove equally compelling, addressing the perennial challenge of maintaining student motivation in digital learning environments. Quantitative analyses reveal 23% average increases in self-reported motivation, with some platforms achieving substantially higher gains through gamiifcation and adaptive challenge mechanisms. Time-on-task metrics show 31% increases in voluntary engagement, while interaction frequency data demonstrates 10-fold improvements in student-initiated learning activities [20].
The mechanisms driving engagement difer from superficial gamification approaches. Adaptive dificulty adjustment maintains optimal challenge levels within each learner’s zone of proximal development, preventing both boredom and frustration. Real-time feedback satisfies psychological needs for competence and autonomy, while personalized content pathways enhance perceived relevance. Particularly significant, engagement improvements persist over time rather than exhibiting novelty decay patterns typical of educational technology interventions [21].
Behavioral analytics reveal deeper engagement patterns beyond surface metrics. Students demonstrate increased metacognitive awareness, spending more time reviewing mistakes and accessing supplementary resources. Help-seeking behaviors become more strategic, with students requesting specific targeted assistance rather than general support. Most remarkably, adaptive systems foster intrinsic motivation shifts, with students reporting greater interest in subject matter independent of external rewards or requirements. 3.1.3. Economic impact: redefining cost-benefit equations Economic analyses reveal compelling returns on investment that reshape institutional decision-making calculus. Arizona State University’s comprehensive implementation generated $12.7 million in instructional cost savings between fiscal years 2017 and 2019 while simultaneously improving student outcomes, demonstrating that quality and eficiency need not trade against each other. These savings derive from multiple sources: reduced remediation costs through predictive intervention, decreased dropout rates saving recruitment expenses, optimized faculty time allocation, and infrastructure eficiencies through cloud-based delivery [22].
Teacher time savings prove substantial, with educators reporting up to 5 hours weekly recovered through AI-assisted grading, lesson planning automation, and administrative task reduction. This recovered time redirects toward high-value activities including individual student mentoring, creative curriculum development, and professional learning community participation. MagicSchool AI, serving over 5 million teachers globally, demonstrates scalability of these eficiency gains across diverse educational contexts [23].
Cost-per-student analyses reveal dramatic reductions compared to traditional personalized instruction models. While human tutoring costs $40-100 per hour, AI-powered adaptive systems deliver comparable personalization at $2-5 per student monthly. Infrastructure investments amortize rapidly across large student populations, with break-even points typically occurring within 18-24 months. Importantly, cost savings accelerate over time as systems accumulate data and improve adaptation algorithms through machine learning refinements.
The technical architectures underlying successful adaptive content generation systems reveal converging design patterns that balance sophistication with practicality. Analysis of leading platforms identifies three primary architectural patterns, each addressing specific educational use cases while sharing common foundational components. 3.2.1. Pattern A: On-demand personalized lesson generation This architecture orchestrates real-time content creation responsive to individual learning needs, employing a sophisticated pipeline that begins with comprehensive learner profiling (figure 2). The system maintains multidimensional learner models encoding knowledge states across granular learning objectives, preferred modalities and cognitive styles, historical interaction patterns and error tendencies, and afective states including motivation and self-eficacy beliefs [24].
Learner profile Learning objective
Historical data
Curriculum
Prompt construction Performance feedback
LLM generation
Content delivery
Knowledge base
RAG verification
Quality sssurance
Learning analytics
Content generation occurs through a multi-stage process ensuring pedagogical alignment and
factual accuracy. Initial prompt construction combines learning objectives with learner characteristics,
incorporating pedagogical constraints that ensure appropriate dificulty and scafolding. The generative
model, typically a fine-tuned large language model, produces candidate content that undergoes
verification through retrieval-augmented generation, comparing generated material against authoritative
knowledge bases. Quality assurance modules evaluate readability, conceptual accuracy, and pedagogical
appropriateness before delivery [
Implementation variations accommodate diferent educational contexts. K-12 systems emphasize
curriculum alignment and age-appropriate content, higher education platforms prioritize depth and
research integration, while professional training systems focus on practical application and competency
demonstration. Successful implementations including Khanmigo’s Socratic tutoring and Carnegie
Learning’s MATHia demonstrate this pattern’s versatility across domains.
3.2.2. Pattern B: Adaptive practice loop systems
Adaptive practice architectures optimize skill acquisition through intelligent problem selection and
feedback generation. These systems maintain detailed skill models tracking mastery probabilities across
interconnected competencies, employing Bayesian knowledge tracing or deep learning approaches to
infer latent knowledge states from observable performance [
The practice loop operates through continuous cycles of assessment, adaptation, and instruction.
Problem selection algorithms balance multiple objectives including targeting skills with highest learning
potential, maintaining appropriate challenge levels, ensuring comprehensive curriculum coverage, and
incorporating spaced repetition for retention. Generated problems adapt not only in dificulty but in
presentation format, contextual framing, and scafolding level based on learner characteristics [
Feedback generation represents a critical diferentiator from traditional practice systems. Rather than binary correct/incorrect indicators, adaptive systems provide multidimensional feedback addressing conceptual understanding, procedural accuracy, strategic approaches, and metacognitive reflection. Error analysis identifies misconception patterns, triggering targeted remediation that addresses root causes rather than surface symptoms. The system generates worked examples demonstrating correct approaches, alternative solution strategies, and connections to previously mastered concepts. 3.2.3. Pattern C: Teacher-assist authoring frameworks This architecture empowers educators to leverage AI capabilities while maintaining pedagogical control, addressing the critical need for teacher agency in technology adoption. The system augments rather than replaces teacher expertise through collaborative content creation workflows where educators provide learning objectives and constraints while AI generates multiple content variations [25].
Quality control mechanisms ensure generated content aligns with teacher intentions and institutional
standards. Educators review and modify AI-generated materials through intuitive interfaces, with
the system learning from corrections to improve future generations. Version control enables tracking
changes and reverting modifications, while approval workflows integrate with existing curriculum
management systems. Analytics provide teachers with insights into content efectiveness, enabling
data-driven refinement of instructional materials.
3.2.4. Hybrid architectures: synthesis and synergy
Leading platforms increasingly adopt hybrid architectures combining elements from all three patterns,
creating comprehensive adaptive learning ecosystems. These systems employ retrieval-augmented
generation for factual accuracy, multiple specialized models for diferent content types, pedagogical
reasoning engines ensuring educational validity, and human-in-the-loop verification for critical decisions.
The integration of RAG with LLMs and domain-specific pedagogical engines demonstrates particular
efectiveness, reducing hallucination rates while maintaining generation flexibility [
Examination of leading platforms reveals distinct implementation strategies and diferential efectiveness patterns that inform deployment decisions. Comparative analysis across four major platforms – Khanmigo, Duolingo Max, ALEKS, and Squirrel AI – illuminates both convergent success factors and unique innovations that drive outcome variations (table 1).
Platform Implementation Key innovation Measured impact Scale Khanmigo GPT-4 Socratic tutoring Ethical AI design, con- 23% accuracy improve- Global, millions of stuversational scafolding ment dents Duolingo Max 148 AI-generated lan- AI-first content trans- 45% better retention 500M+ users guage courses formation rates ALEKS (ASU) Probabilistic assess- Predictive intervention 27% success rate in- 35,000+ students ment crease Squirrel AI LAM framework Nano-level personaliza- 60% time reduction 2M+ students tion 3.3.1. Khanmigo: conversational intelligence in education Khan Academy’s Khanmigo represents a paradigmatic shift toward conversational AI tutoring, leveraging GPT-4’s capabilities while implementing robust ethical safeguards. The platform’s Socratic method implementation guides students through problem-solving processes rather than providing direct answers, fostering deeper conceptual understanding. Empirical evaluation demonstrates 23% improvement in problem-solving accuracy, with particularly strong efects for students requiring additional support [25].
The system’s ethical AI framework addresses critical concerns about generative AI in education. Content filtering prevents inappropriate material generation, bias detection algorithms monitor for discriminatory patterns, and transparency features explain AI reasoning to students and teachers. Privacy protection mechanisms ensure student data remains secure while enabling personalization. These safeguards prove essential for institutional adoption, with 89% of educators reporting increased trust in AI systems after experiencing Khanmigo’s implementation.
Pedagogical innovations distinguish Khanmigo from generic chatbot applications. The system maintains learning trajectories across sessions, building cumulative understanding rather than treating each interaction independently. Metacognitive prompts encourage students to reflect on their learning processes, while collaborative features enable peer learning within safe, moderated environments. Teacher dashboards provide unprecedented visibility into student thinking processes, revealing misconceptions and learning strategies previously hidden in traditional instruction. 3.3.2. Duolingo Max: gamification meets generative AI Duolingo’s transformation into an AI-first platform demonstrates successful integration of generative capabilities with proven gamification mechanisms. The platform’s 148 AI-generated language courses adapt to individual proficiency levels, learning pace, and error patterns while maintaining engaging game-like experiences. Quantitative outcomes prove compelling: 45% improvement in long-term retention, 31% increase in daily active usage, and statistically significant gains across all language skills including listening, speaking, reading, and writing [19].
The technical architecture employs specialized models for diferent language learning aspects. Pronunciation assessment uses acoustic models trained on native speaker data, grammar instruction leverages syntactic parsing and error analysis, while vocabulary acquisition employs spaced repetition algorithms optimized through reinforcement learning. Content generation occurs at multiple granularities, from individual exercise creation to complete lesson sequence planning, ensuring coherent learning progressions.
Engagement mechanisms transcend superficial gamification, incorporating psychological principles of motivation and habit formation. Streak counters leverage loss aversion to maintain daily practice, while experience points and level progression satisfy competence needs. Social features including leagues and friend challenges introduce positive peer pressure without creating excessive competition anxiety. The platform’s success in maintaining engagement – with some users maintaining thousand-day practice streaks – demonstrates the power of well-designed motivational architectures [26]. 3.3.3. ALEKS: precision mathematics through probabilistic modeling ALEKS (Assessment and LEarning in Knowledge Spaces) exemplifies sophisticated mathematical modeling applied to educational adaptation. The system’s Knowledge Space Theory implementation maps mathematical domains into precise prerequisite structures, enabling accurate knowledge state assessment through minimal questioning. Arizona State University’s deployment across 35,000 students achieved remarkable results: 27% improvement in course success rates, 41% reduction in drop/fail/withdraw rates, and $12.7 million in instructional cost savings [22].
The platform’s predictive intervention capabilities identify at-risk students within the first two weeks of enrollment, achieving 98% success in improving identified students to passing grades through targeted support. This early warning system analyzes multiple behavioral indicators including practice attempt patterns, help-seeking behaviors, time allocation across topics, and knowledge growth trajectories. Interventions range from automated encouragement messages to instructor alerts triggering human support, demonstrating efective human-AI collaboration in student success initiatives.
Implementation insights reveal critical success factors. Instructor training proves essential, with
trained faculty achieving significantly better student outcomes than untrained peers. Integration
with existing course structures rather than standalone deployment increases efectiveness. Regular
assessment cycles maintain accurate knowledge models while preventing gaming behaviors. Most
importantly, transparency in system recommendations builds instructor trust, encouraging adoption of
suggested interventions [
The nano-level personalization approach decomposes learning into granular knowledge components – sometimes exceeding 10,000 elements per subject – enabling precise adaptation to individual learning states. Machine learning algorithms identify optimal learning sequences for each student, considering not only knowledge gaps but also learning velocity, forgetting curves, and motivational factors. This comprehensive modeling enables proactive interventions before students experience frustration or disengagement.
Cultural considerations shape platform design and implementation. Content localization extends beyond translation to incorporate culturally relevant examples and pedagogical approaches. Parent engagement features address Asian educational contexts where family involvement proves critical. Competition elements balance individual achievement with collaborative learning, reflecting collectivist values while maintaining personalization benefits. These adaptations demonstrate that successful educational AI requires sensitivity to sociocultural contexts beyond technical capabilities [28].
The comparative analysis reveals that while all platforms demonstrate substantial efectiveness, optimal selection depends on specific educational contexts, subject domains, and implementation resources. Hybrid deployments combining multiple platforms or incorporating platform strengths into custom solutions increasingly represent best practice, leveraging specialized capabilities while maintaining coherent learning experiences.
The transformative potential of generative AI in adaptive education confronts substantial obstacles that transcend technical limitations, encompassing pedagogical validity, ethical responsibility, and systemic readiness. These challenges form an interconnected web where technical capabilities strain against pedagogical wisdom, ethical imperatives clash with scalability demands, and implementation realities expose fundamental inequities in educational systems. Understanding these challenges through a tripartite framework – technical-pedagogical, ethical-social, and implementation-systemic – reveals not isolated problems but interdependent phenomena requiring holistic solutions.
The intersection of technical capabilities and pedagogical requirements creates unique tensions that distinguish educational AI from other applications. These challenges emerge not from technological limitations alone but from the fundamental mismatch between what AI systems can generate and what efective education requires. 4.1.1. Hallucination and its educational consequences The phenomenon of AI hallucination – generating plausible but factually incorrect information – poses distinctive risks in educational contexts where accuracy forms the foundation of knowledge construction. Unlike commercial applications where occasional errors might prove inconvenient, educational hallucinations can propagate misconceptions that persist throughout learners’ intellectual development. Research demonstrates that students encountering AI-generated falsehoods often incorporate these errors into their mental models, particularly when content appears authoritative and aligns with existing misconceptions [29].
The educational manifestation of hallucination extends beyond simple factual errors. Systems generate mathematically impossible solutions that appear procedurally correct, historical narratives that blend actual events with fabricated details, and scientific explanations that violate fundamental principles while maintaining internal consistency. These sophisticated falsehoods prove particularly dangerous because they bypass students’ nascent critical faculties, appearing more credible than obvious errors would.
Mitigation strategies reveal the complexity of addressing hallucination in educational contexts. The PAIR (Problem, AI, Interaction, Reflection) model structures student engagement with AI outputs through systematic verification processes, teaching students to interrogate rather than accept generated content. Guided discovery approaches position AI errors as learning opportunities, developing students’ epistemic vigilance through structured skepticism. However, these pedagogical solutions require sophisticated facilitation that many educators lack training to provide [30].
Technical approaches including retrieval-augmented generation and uncertainty-aware fusion demonstrate promise in reducing hallucination rates. Systems that combine multiple language models based on accuracy assessments achieve 8% improvements in factual accuracy, though this remains insuficient for high-stakes educational applications. The fundamental tension persists: educational contexts demand near-perfect accuracy while current technologies deliver probabilistic approximations [31]. 4.1.2. Cognitive ofloading versus skill development The convenience of AI-generated content creates a pernicious trap where immediate performance improvements mask long-term skill atrophy. Students using AI assistance demonstrate superior task completion in the moment but show diminished capability when support is withdrawn. This cognitive ofloading phenomenon proves particularly pronounced in writing and analytical tasks, where AI scafolding can substitute for rather than support skill development [32].
Empirical evidence reveals disturbing patterns. Students who rely heavily on AI writing tools show 35% reduction in independent writing quality after six months, with particularly severe impacts on argumentation structure and evidence synthesis. Mathematical problem-solving skills deteriorate when students habitually use AI for solution generation, even when they understand the generated solutions. Most concerning, metacognitive awareness – the ability to monitor and regulate one’s own learning – atrophies when AI systems assume these regulatory functions [29].
The developmental implications prove especially troubling for younger learners whose cognitive architectures remain plastic. Elementary students using AI tutoring systems show immediate gains but demonstrate reduced persistence when facing novel challenges without support. Secondary students develop learned helplessness patterns, defaulting to AI assistance rather than attempting independent problem-solving. University students report decreased confidence in their analytical abilities, creating dependency cycles that undermine academic self-eficacy.
Balanced integration approaches attempt to preserve skill development while leveraging AI benefits. Scafolding fade protocols gradually reduce AI support as competence develops, forcing progressive independence. Metacognitive prompting embeds reflection requirements that prevent passive consumption of AI-generated content. Process-focused assessment evaluates problem-solving approaches rather than ifnal answers, incentivizing genuine engagement over AI-mediated performance. Yet implementation of these approaches requires sophisticated pedagogical orchestration that current systems rarely provide [33]. 4.1.3. Assessment validity in the AI era Traditional assessment paradigms collapse when students have unlimited access to sophisticated content generation (table 2). The fundamental assumption that submitted work reflects individual capability becomes untenable when AI can produce essay responses, solve complex problems, and even mimic individual writing styles. This crisis of assessment validity threatens the certification function of education, undermining credentials’ signaling value [34].
Detection technologies fail to provide reliable solutions. Current AI detection tools exhibit false positive rates exceeding 30%, disproportionately flagging work by non-native speakers and students with certain writing patterns. The adversarial nature of detection creates an arms race where generation capabilities consistently outpace detection accuracy. More fundamentally, the binary framing of “human versus AI” authorship ignores the reality of human-AI collaboration where boundaries blur beyond meaningful distinction.
AI-Era challenge Emerging response Limitations Complete AI generation pos- Process documentation re- Labor intensive verification sible quirements Solution generation and ex- Novel problem generation planation Pattern recognition enables Adaptive randomization gaming Sophisticated plagiarism Oral defense requirements Scalability constraints possibilities AI can mimic peer feedback Synchronous collaboration Scheduling dificulties Requires continuous creation Technical complexity
Reconceptualization of assessment proves necessary but challenging. Process-based evaluation tracks learning trajectories rather than outputs, requiring sophisticated monitoring infrastructure. Authentic assessment embeds evaluation within meaningful contexts that resist AI substitution, though creating such contexts at scale proves resource-intensive. Collaborative assessment leverages social dynamics that AI cannot replicate, yet raises fairness concerns about individual accountability [35]. 4.1.4. Scalability versus quality trade-ofs The promise of personalized education at scale confronts fundamental tensions between computational eficiency and pedagogical sophistication. Systems optimized for millions of users necessarily simplify complex learning processes, reducing rich educational interactions to computationally tractable approximations. This simplification cascades through system design, creating quality degradation that compounds as scale increases [36].
Latency requirements for real-time interaction constrain model complexity, forcing trade-ofs between response sophistication and speed. Smaller, faster models lack the nuanced understanding of larger systems, generating superficial responses that fail to address deeper learning needs. Caching and pre-computation strategies improve eficiency but reduce genuine adaptation, creating pseudopersonalization that mimics rather than achieves individualization.
Infrastructure costs escalate non-linearly with quality improvements. High-quality generation requiring large language models demands substantial computational resources, creating economic barriers that privilege well-resourced institutions. Cloud-based solutions introduce dependency vulnerabilities and data sovereignty concerns, while on-premise deployments prove prohibitively expensive for most educational institutions. The resulting quality stratification reinforces existing educational inequities rather than democratizing access as promised.
The deployment of AI in education amplifies existing social inequities while creating novel ethical dilemmas that challenge fundamental educational values. These challenges extend beyond technical ifxes, requiring reconceptualization of fairness, privacy, and integrity in educational contexts. 4.2.1. Algorithmic bias amplification in educational contexts Educational AI systems perpetuate and amplify biases present in training data, creating feedback loops that entrench discrimination. Language models trained on historical educational materials reflect past prejudices, generating content that systematically disadvantages marginalized groups. Performance prediction algorithms exhibit accuracy disparities across demographic groups, with error rates 40% higher for underrepresented minorities. These biases compound through educational pathways, influencing course recommendations, resource allocation, and opportunity access [37].
The mechanisms of bias propagation in education prove particularly insidious. Recommendation systems channel students into tracks that reflect historical patterns rather than individual potential, creating self-fulfilling prophecies of limited achievement. Content generation exhibits representation gaps, with generated examples predominantly featuring majority-culture contexts that alienate diverse learners. Assessment algorithms trained on biased data perpetuate grading disparities, providing diferential feedback quality based on demographic markers rather than performance [38].
Bias mitigation strategies reveal the complexity of achieving fairness in educational AI. Pre-processing approaches that reweight training data can inadvertently introduce new biases while addressing others. In-processing techniques like adversarial debiasing reduce some disparities but often decrease overall model performance, creating equity-eficiency trade-ofs. Post-processing adjustments that modify predictions based on protected attributes raise questions about fairness definitions and legal permissibility. The FAiRDAS framework attempts dynamic fairness monitoring, but defining fairness metrics for educational contexts proves contentious when stakeholders hold conflicting values [39]. 4.2.2. Privacy concerns with minor data Educational AI systems require extensive data collection from vulnerable populations, creating privacy risks that existing frameworks inadequately address. The Children’s Online Privacy Protection Act (COPPA) and Family Educational Rights and Privacy Act (FERPA) provide regulatory boundaries, but their pre-AI provisions fail to anticipate current data practices. Systems collect behavioral patterns, emotional responses, and cognitive processes that reveal intimate details about children’s development, creating profiles that could influence their futures in unforeseen ways [40].
The granularity of data collection in adaptive learning systems exceeds traditional educational records by orders of magnitude. Eye-tracking data reveals attention patterns and potential learning disabilities, keystroke dynamics indicate emotional states and stress levels, and interaction patterns expose social relationships and psychological characteristics. This behavioral surplus – data beyond what is necessary for immediate educational purposes – creates temptations for secondary use that current consent mechanisms cannot adequately address [41].
Cross-border data transfers complicate privacy protection when educational platforms operate globally. The EU’s General Data Protection Regulation (GDPR) provides stronger protections than US frameworks, creating compliance complexity for international educational technologies. Data localization requirements conflict with cloud-based architectures that enable scalable AI deployment. The absence of global privacy standards for educational AI creates regulatory arbitrage opportunities that incentivize minimal protection approaches [42].
Long-term data retention poses unique risks in educational contexts where childhood records could influence adult opportunities. Machine learning models trained on student data encode behavioral patterns that persist within model weights even after explicit data deletion. The right to be forgotten proves technically challenging when AI systems distribute information across neural network parameters rather than discrete database records. These permanent digital shadows of childhood learning struggles could create lasting disadvantage [43]. 4.2.3. Digital divide exacerbation Rather than democratizing education, AI technologies risk widening existing disparities between digitally privileged and underserved populations. The digital divide operates across multiple dimensions – infrastructure access, device availability, digital literacy, and cultural capital – each amplifying educational inequities (figure 3). Students lacking reliable internet access cannot benefit from cloud-based AI tutoring, while those without appropriate devices experience degraded functionality that limits learning benefits [44].
Infrastructure limitations Limited AI tool access
Device availability Degraded learning experience
Digital literacy Cannot leverage
AI benefits Achievement gap widens
Disadvantage cycle perpetuates
Compounding efects of technological inequality
Infrastructure disparities create cascading disadvantages. Rural students with limited bandwidth cannot access multimodal content generation, urban students in overcrowded households lack quiet spaces for voice-based AI interaction, and students relying on school devices face restrictions that prevent personalized adaptation. The “homework gap” expands when AI-enhanced assignments assume home technology access that 15-20% of students lack. These access barriers transform potentially equalizing technologies into mechanisms of stratification.
Digital literacy gaps prevent efective AI utilization even when access exists. Students without foundational computational thinking struggle to formulate efective prompts, interpret AI responses critically, or recognize system limitations. Parents lacking digital sophistication cannot support children’s AI-mediated learning or assess educational technology quality. Teachers in under-resourced schools receive minimal professional development, perpetuating cycles where those most needing support receive least benefit [45]. 4.2.4. Academic integrity redefinition The integration of AI fundamentally challenges traditional conceptions of academic integrity, requiring reconceptualization of authorship, originality, and intellectual efort. The binary framework of “cheating versus honesty” proves inadequate when AI collaboration becomes normative professional practice. Students face contradictory messages about appropriate AI use, with some courses prohibiting any AI assistance while others require its integration [46].
Definitional ambiguities create ethical gray zones that students navigate without clear guidance. Using AI for grammar correction seems acceptable, but where does editing become generation? Brainstorming with AI appears legitimate, but when does ideation become appropriation? These boundary questions lack consensus answers, creating anxiety and inconsistent enforcement that undermines integrity systems’ legitimacy.
The academic integrity crisis extends beyond individual student conduct to institutional credibility. When credentials cannot reliably signal competence, their value diminishes for all holders. Employers increasingly question graduate capabilities, implementing additional screening that disadvantages those whose education genuinely developed targeted skills. The social contract underlying educational certification erodes when authentication becomes impossible.
The structural barriers to efective AI implementation in education reveal misalignments between technological possibilities and institutional realities. These systemic challenges operate across multiple levels – from individual classroom practices through institutional policies to societal educational philosophies – creating implementation gaps that persist despite technical solutions. 4.3.1. Market dynamics creating educational inequality The educational AI market, projected to reach $32.27 billion by 2030, operates through dynamics that systematically advantage already-privileged institutions while marginalizing those most needing support. Premium AI platforms employ subscription models that price out underfunded schools, creating technology tiers that map onto existing resource disparities. Well-resourced institutions purchase comprehensive solutions while others cobble together free tools with limited functionality, reproducing analog inequalities in digital spaces [47].
Vendor lock-in strategies prevent educational institutions from migrating between platforms, creating dependency relationships that extract increasing value over time. Proprietary data formats prevent interoperability, student performance data becomes hostage to continued subscriptions, and switching costs escalate as institutional processes adapt to specific platforms. These market dynamics transform educational technology from a tool serving pedagogical goals into a rent-extraction mechanism that commodifies learning.
The venture capital funding model driving AI education innovation prioritizes scalability and profitability over educational efectiveness. Products optimize for adoption metrics rather than learning outcomes, creating engaging interfaces that may not enhance understanding. The pressure for rapid growth encourages premature deployment of inadequately tested systems, using students as experimental subjects for product development. Educational values of patience, depth, and individual growth conflict with market imperatives of eficiency, standardization, and quarterly returns. 4.3.2. Teacher preparedness and professional development gaps The chasm between teachers’ current capabilities and AI-era requirements represents perhaps the most significant implementation barrier. Surveys indicate that fewer than 20% of educators feel prepared to integrate AI efectively, with most reporting anxiety about their ability to guide AI-enhanced learning. This preparedness gap stems not from technological reluctance but from inadequate support structures that leave teachers navigating complex tools without suficient training [48].
Professional development programs fail to address the sophisticated pedagogical orchestration AI integration requires. Workshop-based training provides surface-level tool familiarity without developing deeper understanding of AI capabilities and limitations. Teachers learn to operate interfaces but not to design learning experiences that leverage AI appropriately. The emphasis on technical features over pedagogical integration creates competent operators who lack the conceptual frameworks for meaningful educational transformation.
Time constraints compound preparedness challenges. Teachers report needing 40-60 hours to become comfortable with new AI platforms, time that professional development rarely provides. The expectation that educators will self-train during personal time creates burnout and resentment. Early adopters who invest this time often become informal support for colleagues, creating additional uncompensated labor that accelerates exhaustion. The individualization of training responsibility ignores systemic nature of capability building [49]. 4.3.3. Infrastructure and resource requirements The technical infrastructure required for AI implementation exceeds many educational institutions’ capabilities, creating participation barriers that exclude entire communities. Bandwidth requirements for real-time AI interaction strain school networks designed for basic connectivity. Server infrastructure for on-premise deployment demands capital investments competing with other educational priorities. Cloud-based solutions introduce recurring costs that strain operational budgets while creating dependency vulnerabilities [50].
Device ecosystems prove particularly challenging when AI applications assume computational capabilities beyond basic educational hardware. Chromebooks dominating K-12 markets lack processing power for sophisticated AI applications, tablets purchased for digital textbooks cannot run required software, and bring-your-own-device policies create security vulnerabilities while reinforcing socioeconomic disparities. The hidden curriculum of device requirements teaches students that educational opportunity depends on family resources.
Technical support needs escalate exponentially with AI adoption. Systems require continuous updates that disrupt instructional time, integration failures cascade across interconnected platforms, and troubleshooting demands expertise beyond typical educational IT capabilities. Schools resort to expensive external consultants or rely on technically proficient teachers who become de facto IT support, diverting them from instructional responsibilities. The support burden falls disproportionately on under-resourced institutions least able to accommodate it. 4.3.4. Governance vacuum and policy fragmentation The absence of comprehensive governance frameworks for educational AI creates regulatory uncertainty that impedes thoughtful implementation while enabling problematic practices. Educational authorities lack expertise to evaluate AI systems’ pedagogical validity, privacy implications, or fairness characteristics. Procurement decisions rely on vendor claims rather than independent evaluation, creating markets for persuasive marketing rather than educational efectiveness [51].
Policy fragmentation across jurisdictions creates compliance complexity that favors large vendors over innovative alternatives. State-level regulations conflict with federal requirements, district policies contradict classroom practices, and international students face diferent rules than domestic peers. This regulatory patchwork prevents coherent implementation strategies while creating loopholes that sophisticated actors exploit. The resulting confusion leaves educators uncertain about permissible practices, chilling innovation while failing to prevent harm.
The governance vacuum extends to fundamental questions about educational AI’s role and limits. Should AI systems make high-stakes decisions about student advancement? What transparency rights do students have regarding AI-mediated assessments? How should institutions balance eficiency benefits against human relationship values? These questions require societal deliberation, yet policy development lags years behind technological deployment, creating facts on ground that constrain future choices.
The transformation of educational systems through generative AI necessitates a principled framework
that preserves human agency while harnessing technological capabilities. Four foundational principles
emerge from our synthesis of empirical evidence and theoretical foundations, each addressing critical
tensions between technological possibility and pedagogical responsibility.
5.1.1. Pedagogical primacy
Educational technology history reveals a persistent pattern: technological capabilities drive
implementation rather than learning needs determining technological deployment. This inversion produces
systems optimized for computational eficiency rather than learning efectiveness. Pedagogical primacy
reverses this dynamic, asserting that sound pedagogical theories and practices must drive AI integration,
not technological capabilities alone [
Operationalizing pedagogical primacy requires systematic alignment between AI capabilities and established learning theories. Personalized learning paths adapt content and pacing through constructivist frameworks, supporting zone of proximal development calculations that adjust dificulty based on demonstrated competence rather than predetermined sequences [53]. The Innovation Fellowship study demonstrated that educator competency frameworks, particularly structured 12-competency models, enable teachers to develop AI pedagogical skills through progressive mastery rather than overwhelming exposure [54].
Critical to implementation, analogy-based approaches demystify AI operations for learners, particularly younger students who struggle with abstract computational concepts. Rather than presenting AI as an opaque oracle, efective implementations use familiar metaphors – AI as a study partner, learning coach, or research assistant – that preserve learner agency while clarifying system capabilities and limitations [55]. This pedagogical grounding prevents the cognitive outsourcing that occurs when learners perceive AI as infallible authority rather than collaborative tool.
Interdisciplinary collaboration between educators and technologists ensures pedagogical alignment throughout development cycles. The Arizona State University implementation succeeded precisely because instructional designers, not engineers, led system specification. Technical teams translated pedagogical requirements into computational architectures rather than educators adapting to predetermined technical constraints [56]. This reversal of traditional development hierarchies produced systems where learning objectives determine feature sets, assessment validity guides data collection, and pedagogical coherence constrains generative outputs. 5.1.2. Human-in-the-loop requirement The seductive promise of full automation obscures education’s fundamentally human nature. Humanin-the-loop (HITL) frameworks maintain essential oversight and agency by positioning educators as orchestrators rather than observers of AI-mediated learning [57]. This requirement transcends simple veto power over AI decisions; it embeds human judgment throughout the adaptive cycle from initial assessment through intervention design to outcome evaluation (figure 4).
Participatory design methodologies involve teachers and stakeholders from conception through deployment, ensuring systems reflect pedagogical realities rather than idealized computational models. The MagicSchool AI platform’s adoption by over 5 million teachers resulted from extensive co-design
AI system oversight explanations
Student processes where educators shaped feature priorities, workflow integration, and interface design [ 58]. Teachers rejected initial proposals for fully automated lesson planning, instead requesting AI assistance for specific bottlenecks while maintaining creative control over instructional design.
Explainable AI (XAI) provides interpretable, context-sensitive explanations for system decisions, enabling educators to understand and potentially override AI recommendations. The IELAT framework achieves 99.81% accuracy while maintaining complete transparency through salient input region highlighting and decision path visualization [59]. Educators can trace how student responses influence dificulty adjustments, understand why particular content was recommended, and identify potential biases in algorithmic decision-making.
Continuous feedback loops enable real-time human intervention and iterative system improvement. Rather than batch processing where errors compound before detection, HITL architectures support immediate correction when AI generates inappropriate content or misinterprets student responses [60]. The Khanmigo implementation includes “pause points” where complex student queries trigger human review before AI response generation, preventing hallucination propagation in high-stakes explanations. 5.1.3. Transparency by design Transparency extends beyond technical interpretability to encompass cognitive, phenomenological, and social dimensions of understanding. Educational contexts demand not merely that AI decisions be explicable but that explanations resonate with stakeholders’ mental models and decision-making frameworks [61].
Structured transparency frameworks tailor explanations to distinct stakeholder groups. Students receive metacognitive prompts explaining why particular content was selected, helping develop selfawareness about their learning processes. Educators access pedagogical rationales linking AI recommendations to learning objectives and theoretical frameworks. Administrators view aggregate patterns demonstrating system efectiveness and potential bias indicators. Parents obtain comprehensible summaries of their child’s progress without overwhelming technical detail [62].
Transparency indices quantify system openness across multiple dimensions: algorithmic clarity, data usage disclosure, decision reversibility, and outcome predictability. The transparency score developed by MIT researchers combines 47 metrics into a composite measure enabling institutional comparison and improvement tracking. Systems scoring below threshold values on critical dimensions face usage restrictions in several jurisdictions, creating market incentives for transparent design [59].
Mixed-methods evaluation combining qualitative and quantitative approaches reveals transparency’s impact on trust and engagement. Eye-tracking studies demonstrate that educators spend 73% more time examining AI explanations when presented through familiar pedagogical frameworks rather than technical descriptions. Student surveys indicate transparency increases perceived fairness even when outcomes remain unchanged, suggesting procedural justice matters as much as distributive justice in educational AI [59]. 5.1.4. Equity first Educational AI risks amplifying existing inequalities unless equity considerations drive design from inception. The equity-first principle demands proactive attention to accessibility, bias mitigation, and inclusive design rather than post-hoc remediation attempts [63].
Multi-group fairness frameworks address bias across intersectional student populations rather than optimizing for majority groups. Traditional fairness metrics that achieve demographic parity for single protected attributes can mask discrimination against students at attribute intersections – for instance, systems fair to both racial minorities and students with disabilities separately may still discriminate against minority students with disabilities. Contemporary implementations employ causal fairness models that identify and mitigate compound disadvantages through multidimensional optimization [64].
Universal Design for Learning (UDL) principles ensure systems accommodate diverse abilities, backgrounds, and contexts from initial architecture rather than through retrofitted accommodations [65]. The Canvas AI tutor provides content through multiple modalities simultaneously – text, audio, visual, and interactive – allowing students to engage through their preferred channels without requiring disability documentation or special configuration. This proactive inclusivity serves all learners while particularly benefiting those with undiagnosed or unsupported learning diferences [66].
Culturally responsive strategies value linguistic diversity and cultural knowledge rather than treating deviation from dominant norms as deficiency. Natural language processing models trained primarily on standard American English systematically disadvantage speakers of other English varieties, interpreting grammatically correct African American Vernacular English constructions as errors. Successful implementations employ ensemble models combining dialect-specific training with metalinguistic awareness, recognizing linguistic variation as richness rather than incorrectness [67].
The distinction between augmentation and automation fundamentally reconceptualizes the educator’s role in AI-enhanced learning environments. Rather than replacing human capabilities, augmentation amplifies them, creating new possibilities for pedagogical practice while preserving essential human elements. 5.2.1. From content delivery to learning experience design Traditional teaching often reduces educators to content delivery mechanisms, a role easily automated by AI. Augmentation liberates teachers from information transfer to become learning experience designers who orchestrate complex, multimodal journeys tailored to individual students [68].
AI handles content generation, allowing educators to focus on curricular architecture, emotional scafolding, and metacognitive development. Teachers using MagicSchool AI report spending 67% less time creating materials and 340% more time on student consultation and pedagogical planning. This shift transforms the classroom from information pipeline to learning laboratory where educators experiment with engagement strategies, motivation techniques, and conceptual frameworks AI cannot independently navigate [69].
The learning experience designer role requires new competencies: data interpretation skills to understand AI analytics, design thinking to create coherent learning journeys, and systems awareness to orchestrate multiple technological and human elements. Professional development programs increasingly emphasize these meta-skills over specific platform training, recognizing that educators must adapt to rapidly evolving technological landscapes [70, 48]. 5.2.2. From manual grading to data-driven intervention Assessment automation frees educators from mechanical evaluation to engage in sophisticated diagnostic interpretation and targeted intervention design. Rather than spending hours marking identical errors limiting session duration. Perhaps most critically, teacher preparation lags technology deployment – 78% of educators report feeling unprepared to integrate AR/VR meaningfully into curriculum. Successful implementations invest equally in professional development and hardware acquisition [158]. 7.2.3. Neuroadaptive systems The convergence of neuroscience and educational technology enables unprecedented personalization through real-time neural monitoring and adaptation. Neuroadaptive systems use EEG, eye-tracking, and physiological sensors to detect cognitive states – attention, cognitive load, emotional valence – and dynamically adjust learning experiences to maintain optimal challenge levels [159].
NeuroChat exemplifies practical neuroadaptive implementation, using consumer-grade EEG headbands to monitor engagement during AI tutoring sessions. When theta wave patterns indicate mindwandering, the system employs attention restoration techniques: changing content modality, introducing surprise elements, or suggesting physical movement breaks. During high cognitive load periods marked by elevated beta activity, content automatically simplifies, pacing slows, and additional scaffolding appears. Students using neuroadaptive tutoring show 34% improvement in sustained attention and 28% reduction in cognitive fatigue [159].
Advanced implementations combine multiple biosignals for comprehensive state assessment. Pupil dilation indicates cognitive efort, galvanic skin response reveals emotional arousal, and heart rate variability suggests stress levels. Machine learning algorithms integrate these signals to construct multidimensional cognitive-emotional state models that inform moment-to-moment adaptation. The precision of neuroadaptive systems surpasses self-report or behavioral inference, detecting cognitive overload 4.7 seconds before performance degradation appears [160].
Ethical considerations surrounding neuroadaptive learning demand careful navigation. Brain data represents the ultimate privacy frontier – thoughts, emotions, and cognitive processes laid bare to algorithmic analysis. Strict data governance protocols become essential, limiting neural data use to immediate adaptation without long-term storage or cross-purpose analysis. The “cognitive sovereignty” principle emerges, asserting individuals’ absolute ownership of their neural data and right to cognitive privacy [161]. 7.2.4. Social learning networks Digital transformation enables new forms of social learning that transcend traditional classroom boundaries. Global peer networks connect learners across geographic, cultural, and linguistic divides, creating cognitive diversity that enriches learning through exposure to varied perspectives and problemsolving approaches [162].
AI facilitates optimal peer matching based on complementary skills, compatible learning styles, and motivational alignment. Rather than random group assignment, intelligent algorithms identify collaboration patterns that maximize mutual benefit. A student strong in mathematical reasoning but weak in verbal expression partners with a peer exhibiting opposite strengths, creating synergistic learning relationships. Dynamic regrouping based on evolving competencies ensures continued challenge and growth [163].
Collaborative knowledge construction platforms enable collective intelligence emergence. Students contribute partial solutions, build on peers’ ideas, and synthesize diverse inputs into coherent understanding. Version control systems track contribution histories, enabling fair assessment of collaborative work while maintaining individual accountability. Wiki-based learning environments where students collectively create course content show 45% deeper conceptual understanding compared to traditional instruction [164].
Social emotional learning integrates naturally within digital peer networks. Students develop empathy through perspective-taking exercises, practice conflict resolution in low-stakes virtual environments, and build communication skills through structured peer feedback. AI monitors interaction quality, identifying toxic dynamics early and facilitating constructive engagement patterns. The social skills developed through digital collaboration prove increasingly essential for distributed work futures [165].
Individual technological and pedagogical innovations require systemic transformation to achieve educational reimagination at scale. Infrastructure, standards, policies, and professional roles must evolve coherently to support rather than constrain emerging possibilities. 7.3.1. Interoperability standards The proliferation of educational technologies creates integration challenges as institutions deploy dozens of disconnected systems. Interoperability standards enable seamless data exchange, functionality sharing, and user experience consistency across diverse platforms [166].
Emerging standards like ISO/IEC 19788 (Metadata for Learning Resources) and IMS Global’s Learning Tools Interoperability (LTI) provide technical frameworks for system integration. These standards define common data formats, authentication protocols, and API specifications that allow learning management systems, assessment platforms, and content repositories to communicate fluently. The OneRoster standard enables automatic roster synchronization, eliminating manual data entry that consumes 127 hours annually per school administrator [167].
Semantic interoperability transcends technical data exchange to ensure shared meaning across systems. The Educational Knowledge Graph initiative creates universal ontologies mapping concepts, competencies, and credentials across institutional boundaries. When a student transfers between schools, their learning history translates automatically into the receiving institution’s framework, preserving continuity despite systemic diferences [168].
Blockchain-based credentialing creates tamper-proof, universally verifiable academic records that students own and control. Rather than requesting transcripts from former institutions, learners share cryptographically signed credentials directly with employers or educational programs. The Blockcerts standard, adopted by MIT and 200 other institutions, enables instant verification while preventing credential fraud that costs $2.3 billion annually [169]. 7.3.2. Universal learner records Traditional transcripts inadequately capture contemporary learning occurring across formal, informal, and non-formal contexts. Universal Learner Records (ULRs) document comprehensive learning journeys (figure 7), including micro-credentials, workplace training, self-directed study, and experiential learning [170].
Competency-based frameworks replace course-centric organization with skill taxonomies that map across contexts. Whether a student develops project management skills through formal coursework, workplace experience, or volunteer coordination becomes irrelevant; the competency itself receives recognition. Machine learning algorithms analyze diverse evidence types – portfolios, peer assessments, performance data – to validate competency claims with 91% accuracy compared to expert evaluation [171].
Digital wallets give learners sovereignty over their educational data, choosing what to share with whom for which purposes. Privacy-preserving protocols enable selective disclosure – proving degree completion without revealing grades, demonstrating language proficiency without exposing full transcripts. This granular control empowers learners while protecting sensitive information [172]. 7.3.3. Global accessibility initiatives Educational equity demands that advanced learning technologies serve all students regardless of geographic location, economic resources, or physical abilities. Global accessibility initiatives work to bridge digital divides while ensuring inclusive design [173]. Traditional transcript (courses and grades)
Microcredentials Self-directed study
Universal learner record
Experiential learning
Social contributions
Workplace learning
The Universal Design for Learning (UDL) framework, updated for AI-enhanced education (UDL 3.0), provides principles ensuring all learners can access, engage with, and demonstrate knowledge. AI systems automatically generate alternative content representations – converting text to speech, adding visual descriptions, simplifying language complexity – based on individual accessibility profiles. Real-time captioning, sign language avatars, and haptic feedback systems ensure sensory impairments don’t become learning barriers [174].
Infrastructure initiatives address connectivity gaps preventing technology access. Low Earth orbit satellite constellations promise global broadband coverage by 2027, bringing high-speed internet to 3 billion currently unconnected individuals. Edge computing architectures enable sophisticated AI processing on low-power devices, eliminating the need for expensive hardware [175]. Mesh networking protocols allow community-created networks in areas lacking commercial infrastructure [176].
Localization extends beyond language translation to cultural adaptation. AI systems trained primarily on Western educational content perpetuate cultural biases when deployed globally. Successful initiatives involve local educators in training data creation, algorithm refinement, and pedagogical adaptation. The African Institute for Mathematical Sciences develops culturally relevant AI tutors that use local examples, respect indigenous knowledge systems, and align with community values [177]. 7.3.4. New educator roles emergence The augmentation paradigm transforms rather than eliminates educator roles, creating new professional identities that leverage human capabilities AI cannot replicate. These emerging roles require diferent competencies, preparation pathways, and support structures [178].
Learning experience designers orchestrate complex educational journeys combining human instruction, AI tutoring, peer collaboration, and experiential learning. Rather than delivering content, they architect learning ecosystems where multiple elements synergistically support student development. This role demands systems thinking, data interpretation capabilities, and deep pedagogical knowledge to balance technological and human elements efectively [179].
Cognitive coaches support students’ metacognitive and social-emotional development, areas where human insight remains irreplaceable. They help students understand their learning patterns, develop self-regulation strategies, and navigate the psychological challenges of continuous learning. As AI handles content delivery and assessment, cognitive coaches focus on motivation, resilience, and identity development [180].
Algorithm auditors ensure AI systems serve educational rather than eficiency goals. They examine recommendation patterns for bias, evaluate whether adaptations genuinely benefit learners, and advocate for student interests when these conflict with algorithmic optimization. This role requires technical literacy to understand AI operations, pedagogical expertise to evaluate educational impact, and ethical grounding to identify value conflicts [181].
Learning analytics interpreters translate complex data into actionable insights for students, parents, and administrators. They identify patterns human intuition might miss while contextualizing algorithmic ifndings within lived realities. When AI flags a student as “at-risk”, interpreters investigate underlying causes, coordinate support resources, and communicate sensitively with stakeholders. This bridging role proves essential for maintaining human agency in data-driven systems [182].
The transformation of educator roles requires systematic professional development beyond traditional technology training. Universities develop new teacher preparation programs integrating computer science, data science, and cognitive psychology with pedagogical training. Micro-credentialing systems allow practicing educators to progressively develop new competencies through stackable certificates. Most critically, professional learning communities provide ongoing support as roles continue evolving [183].
The comprehensive examination of generative AI’s integration into adaptive educational systems reveals a moment of unprecedented opportunity coupled with profound responsibility. The evidence presented throughout this analysis converges on a fundamental truth: we stand at an inflection point where technological capabilities finally align with longstanding pedagogical aspirations, yet this alignment alone guarantees neither positive transformation nor equitable outcomes [52]. 8.1.1. Technology as enabler, not replacement The augmentation paradigm emerges not as philosophical preference but as empirical necessity. Across implementations spanning continents, educational levels, and socioeconomic contexts, a consistent pattern manifests: AI systems achieve optimal outcomes when amplifying rather than supplanting human capabilities [184]. The 27% improvement in course success rates at Arizona State University occurred not through teacher replacement but through liberation – freeing educators from administrative burden to engage in high-touch mentorship. The 5 million teachers adopting MagicSchool AI seek not obsolescence but enhancement, using technology to extend their pedagogical reach rather than abdicate their educational responsibility.
This distinction between augmentation and automation transcends semantic precision to embody fundamentally diferent visions of education’s future. Automation conceptualizes learning as information transfer amenable to algorithmic optimization, reducing education to its most mechanistic components. Augmentation recognizes education as irreducibly human endeavor where knowledge transmission represents merely the substrate upon which critical thinking develops, creativity flourishes, and identity forms. Technology excels at the former; only humans accomplish the latter [185].
The TPACK framework’s evolution demonstrates how technological integration succeeds when subordinated to pedagogical objectives rather than driving them. Teachers who develop technological pedagogical content knowledge – understanding not just what technology can do but when and why to deploy it – report 73% higher satisfaction and 45% better student outcomes than those receiving purely technical training. This suggests that preserving human primacy requires not resistance to technology but sophisticated orchestration of human and artificial capabilities [186]. 8.1.2. Urgency of ethical framework development The velocity of AI deployment in educational contexts outpaces ethical framework development by orders of magnitude. While technology companies release new models monthly, institutional review boards struggle to evaluate single implementations over semesters. This temporal mismatch creates ethical vacuums where consequential decisions about student data, algorithmic influence on developing minds, and AI’s role in shaping intellectual development occur without adequate oversight or reflection [187].
Evidence from early deployments reveals concerning patterns demanding immediate attention. Algorithmic bias in adaptive systems perpetuates and amplifies existing educational inequities, with minority students receiving systematically diferent recommendations that constrain rather than expand opportunity. Hallucination in educational contexts proves particularly pernicious, as students lack expertise to identify subtle factual errors that become incorporated into developing knowledge structures. Privacy violations extend beyond data breaches to encompass cognitive privacy – the right to intellectual development free from algorithmic manipulation [188].
Yet ethical framework development cannot await perfect understanding of rapidly evolving technologies. The precautionary principle, while protective, risks paralysis that denies students beneficial innovations. Dynamic ethical frameworks that evolve through iterative refinement ofer more promise than static regulations inevitably obsolesced by technological advancement. These frameworks must balance multiple tensions: innovation versus protection, personalization versus privacy, eficiency versus equity. Most critically, they must center student wellbeing rather than institutional convenience or commercial profit [189].
The emergence of “ethics washing” – superficial ethical compliance masking fundamental conflicts of interest – demands frameworks with enforcement mechanisms beyond voluntary compliance. Mandatory algorithmic auditing, independent oversight boards with student and parent representation, and liability frameworks creating genuine accountability become essential components. The European Union’s AI Act provides initial scafolding, though educational applications require domain-specific elaboration addressing unique vulnerabilities of developing minds [190]. 8.1.3. Need for collaborative governance The complexity of AI-enhanced education exceeds any single stakeholder’s comprehension or control, necessitating governance models that orchestrate diverse perspectives, expertise, and interests. Traditional hierarchical governance structures – where administrators decide, teachers implement, and students comply – prove inadequate for systems where algorithms make millions of micro-decisions beyond human oversight [191].
Multi-helix collaboration models incorporating government, academia, industry, civil society, and media demonstrate particular promise. The Indonesian special autonomous region implementations show how each sector contributes essential elements: government provides regulatory frameworks and resources, academia supplies research and evaluation capacity, industry ofers technical expertise and innovation, civil society ensures community voice and accountability, while media facilitates public discourse and transparency. When these elements align, transformation accelerates; when any element fails, implementation falters [192].
Heterarchical networks that distribute decision-making authority according to expertise rather than position enable responsive governance capable of rapid adaptation. Rather than centralized command structures, these networks create multiple feedback loops where classroom experiences inform policy, technical capabilities shape pedagogical possibilities, and ethical considerations constrain commercial imperatives. The Implementation-STakeholder Engagement Model demonstrates how continuous stakeholder involvement throughout implementation phases – not merely at inception – predicts success more strongly than resource availability or technical sophistication [193].
Trust emerges as governance’s foundational currency, yet trust in AI educational systems remains fragile. Parents fear algorithmic determination of their children’s futures, teachers worry about professional displacement, students question whether AI truly serves their interests. Building trust requires radical transparency about system operations, genuine participation in design decisions, and demonstrable commitment to stakeholder welfare over eficiency metrics. The collaborative governance imperative thus extends beyond coordination to encompass fundamental reconceptualization of power, authority, and agency in educational systems [194].
The synthesis of evidence, analysis of challenges, and articulation of possibilities converge on an urgent imperative: stakeholders across the educational ecosystem must act decisively to shape AI’s integration before technological momentum renders human agency moot. 8.2.1. For educators: embrace augmentation, maintain human primacy Educators stand at the transformation’s epicenter, possessing unique power to determine whether AI becomes tool for liberation or instrument of obsolescence. The call to action requires transcending both technophobic resistance and uncritical adoption to develop sophisticated professional judgment about when, how, and why to deploy AI capabilities.
Immediate actions include developing AI literacy through structured professional development that emphasizes pedagogical application over technical operation. Understanding how large language models generate text matters less than recognizing when AI-generated content serves learning objectives. Participating in system design and evaluation ensures educational rather than technical priorities drive development. Most critically, educators must document and share both successes and failures, building collective wisdom about efective AI integration [195].
Maintaining human primacy requires explicit assertion of irreplaceable human contributions: emotional support during struggle, inspiration through passionate engagement, wisdom from lived experience, and moral guidance through ethical complexity. These uniquely human capabilities become more, not less, valuable as AI assumes routine instructional tasks. Teachers who clearly articulate and demonstrate these distinctive contributions secure their professional future while serving student needs AI cannot address [196]. 8.2.2. For technologists: prioritize pedagogical validity over innovation Technology developers wield enormous influence over education’s future through design decisions that shape possibilities and constraints. The call to action demands fundamental reorientation from technological sophistication as primary objective to pedagogical efectiveness as ultimate criterion [197].
This requires embedding educators as equal partners throughout development cycles rather than consultants providing post-hoc validation. User experience research must extend beyond interface design to encompass learning impact, cognitive development efects, and long-term educational outcomes. The “first, do no harm” principle from medicine applies equally to educational technology, demanding rigorous testing for unintended consequences before wide deployment [198].
Transparency about system limitations proves as important as promoting capabilities. Acknowledging hallucination rates, bias patterns, and failure modes enables appropriate use while building trust through honesty. Open-source development models that enable inspection, modification, and local adaptation serve educational imperatives better than proprietary black boxes optimized for commercial metrics. Most fundamentally, technologists must resist the temptation to solve educational “problems” that exist primarily as market opportunities rather than genuine pedagogical needs [199]. 8.2.3. For policymakers: create protective yet enabling frameworks Policymakers navigate treacherous terrain between overregulation that stifles beneficial innovation and underregulation that exposes vulnerable populations to harm. The call to action requires adaptive frameworks that protect without paralyzing, guide without dictating, and evolve with technological advancement [200].
Immediate priorities include establishing minimum standards for educational AI covering data privacy, algorithmic transparency, bias auditing, and human oversight. These standards must be specific enough to provide meaningful protection yet flexible enough to accommodate diverse contexts and rapid evolution. Liability frameworks clarifying responsibility when AI systems cause harm – whether through incorrect information, biased recommendations, or privacy violations – create accountability incentives for responsible development [167].
Investment in public infrastructure supporting equitable AI access prevents digital divides from becoming cognitive chasms. This encompasses not merely device and connectivity provision but professional development support, curriculum integration resources, and evaluation capacity building. Public funding for educational AI research independent of commercial interests ensures evidence-based rather than market-driven policy development [201]. 8.2.4. For researchers: focus on long-term human development impacts The research community bears responsibility for generating evidence that guides responsible AI integration while identifying and mitigating potential harms. The call to action requires shifting focus from short-term performance metrics to long-term human development outcomes [202].
Longitudinal studies tracking cohorts from early AI exposure through adulthood reveal cumulative efects invisible in semester-length investigations. Does early AI assistance accelerate cognitive development or create dependency? How does algorithmic mediation of learning influence identity formation, career trajectories, and lifelong learning capacity? These questions require patient investigation spanning years rather than publication cycles [203].
Interdisciplinary collaboration becomes essential as educational AI’s impacts transcend traditional disciplinary boundaries. Cognitive scientists must work with computer scientists to understand how algorithms influence neural development. Sociologists must collaborate with data scientists to identify bias patterns. Ethicists must engage engineers to embed values in system architectures. Only through such collaboration can research address AI education’s full complexity [204].
The path forward requires neither wholesale embrace nor categorical rejection of AI in education, but thoughtful integration guided by human values, pedagogical wisdom, and unwavering commitment to learner wellbeing. Three aspirations crystallize from this analysis, representing not utopian fantasies but achievable objectives given suficient will and wisdom. 8.3.1. Education that cultivates uniquely human capabilities Future educational systems leverage AI to handle mechanistic tasks – information retrieval, routine assessment, administrative coordination – thereby liberating human potential for distinctively human endeavors. Students develop critical thinking through Socratic dialogue with teachers freed from lecture delivery. Creative expression flourishes when AI handles technical execution, allowing focus on ideation and meaning-making. Collaborative problem-solving skills emerge through carefully orchestrated group work where AI facilitates but humans connect [205].
This vision positions AI as cognitive exoskeleton that amplifies human capability rather than replacement that substitutes for it. Just as physical tools extended human strength without eliminating need for human judgment about where to direct that strength, cognitive tools extend intellectual capacity while preserving human agency over its application. The measure of success becomes not what AI can do independently but what humans can accomplish with AI assistance [206]. 8.3.2. AI that democratizes quality education Properly deployed, AI addresses education’s most persistent inequality: the accident of birth that determines access to quality instruction. A student in rural Bangladesh gains access to world-class mathematics tutoring through AI that adapts to their context, language, and learning style. A child with learning disabilities receives perfectly calibrated support that neither stigmatizes nor limits. An adult learner pursues new career paths through personalized instruction that accommodates work schedules and family obligations [207].
Democratization extends beyond access to encompass agency – ensuring all learners shape their educational journeys rather than merely consuming predetermined content. AI systems that respect cultural diversity, accommodate diferent ways of knowing, and support varied life trajectories serve democratic rather than homogenizing functions. This requires deliberate design decisions prioritizing inclusivity over eficiency, representation over standardization, empowerment over control [208]. 8.3.3. Systems that enhance rather than erode human agency The ultimate aspiration envisions educational systems where technology amplifies rather than diminishes human agency at every level. Students exercise meaningful choice over learning paths while receiving support that enables informed decisions. Teachers deploy professional judgment about when and how to use AI while maintaining authority over pedagogical decisions. Parents understand and inlfuence how algorithms shape their children’s education. Communities ensure educational technologies reflect local values while connecting to global knowledge [209].
This requires fundamental reconceptualization of agency in algorithmic contexts. Agency means not absence of AI influence but conscious collaboration with AI systems whose operations remain transparent and whose recommendations remain advisory. It demands educational systems that develop metacognitive awareness about AI interaction, critical evaluation skills for AI-generated content, and ethical reasoning about AI’s proper role. Most fundamentally, it requires recognition that preserving human agency in an AI-saturated world becomes education’s essential mission [210].
The convergence of technological capability, pedagogical understanding, and ethical awareness creates unprecedented opportunity to reimagine education for human flourishing. Yet opportunity alone guarantees nothing. The choices made today about AI’s role in education will reverberate through generations, shaping not merely what students learn but who they become. The imperative for action is clear: we must act with wisdom, courage, and unwavering commitment to human dignity to ensure that education remains a fundamentally human endeavor that develops uniquely human capabilities, serves irreducibly human purposes, and preserves essentially human agency. The future of education – and perhaps humanity itself – depends on getting this right.
During the research and writing process, we utilized several AI tools to enhance eficiency. Scopus AI helped refine our literature search strategy, while Grammarly assisted with grammar and style. We also employed Claude Opus 4.1 to polish sentence structure and improve clarity, always with careful human review and editing to ensure accuracy. 2013.06.010. [18] Y. Fang, Z. Ren, X. Hu, A. C. Graesser, A meta-analysis of the efectiveness of ALEKS on learning,
Educational Psychology 39 (2019) 1278–1292. doi:10.1080/01443410.2018.1495829. [19] İ. Y. Kazu, M. Kuvvetli, Digital language learning with Duolingo: assessing its impact on listening, speaking, reading, and writing skills, Journal of Computers in Education (2025). doi:10.1007/s40692-025-00355-0. [20] C. Zhou, T. Ren, L. Lang, The impact of AI-Based adaptive learning technologies on motivation and engagement of higher education students, Education and Information Technologies (2025). doi:10.1007/s10639-025-13646-x. [21] I. Ivanović, Comparative Analysis of Student Engagement in Traditional vs AI-Assisted Learning
Using UWS-9S, Drustvena Istrazivanja 34 (2025) 169–189. doi:10.5559/di.34.2.02. [22] S. Sun, N. M. Else-Quest, L. C. Hodges, A. M. French, R. Dowling, The Efects of ALEKS on Mathematics Learning in K-12 and Higher Education: A Meta-Analysis, Investigations in Mathematics Learning 13 (2021) 182–196. doi:10.1080/19477503.2021.1926194. [23] M. Kwak, The Efectiveness of AI-Driven Tools in Improving Student Learning Outcomes Compared to Traditional Methods, Issues in Information Systems 26 (2025) 233–247. doi:10. 48009/4_iis_2025_120. [24] C. Dong, Y. Yuan, K. Chen, S. Cheng, C. Wen, How to Build an Adaptive AI Tutor for Any Course Using Knowledge Graph-Enhanced Retrieval-Augmented Generation (KG-RAG), in: 14th International Conference on Educational and Information Technology, ICEIT 2025, Guangzhou, China, March 14-16, 2025, IEEE, 2025, pp. 152–157. doi:10.1109/ICEIT64364.2025.10975937. [25] J. I. Alvarez, J. R. Angeles, Khanmigo in the Virtual Classroom: A Strategic Evaluation through SWOT and Acceptability Analysis, Educational Process: International Journal 16 (2025) e2025272. doi:10.22521/edupij.2025.141.5. [26] Z. Ouyang, Y. Jiang, H. Liu, The Efects of Duolingo, an AI-Integrated Technology, on EFL Learners’ Willingness to Communicate and Engagement in Online Classes, The International Review of Research in Open and Distributed Learning 25 (2024) 97–115. doi:10.19173/irrodl. v25i3.7677. [27] H. Li, X. Fan, J. Liang, The Squirrel AI Adaptive Learning System Accompanying Millions of Children in Their Growth, in: H. Zeng, J. Pan, J. Xu, Q. Sun (Eds.), Digital Transformation of Regional Education in China, Springer Nature Singapore, Singapore, 2025, pp. 51–53. [28] Y. Qiu, I. N. Asniza, S. Zheng, The Development of Mathematics Lesson Plan Integrated with Squirrel Artificial Intelligence in Enhancing the Fifth Grade Primary School Students’ Cognitive Development in Fujian Province, China, in: Proceedings of the 2024 3rd International Conference on Artificial Intelligence and Education, ICAIE ’24, Association for Computing Machinery, New York, NY, USA, 2025, p. 308–313. doi:10.1145/3722237.3722290. [29] M. Gerlich, AI Tools in Society: Impacts on Cognitive Ofloading and the Future of Critical
Thinking, Societies 15 (2025) 6. doi:10.3390/soc15010006. [30] O. A. Acar, PAIR (problem, AI, interaction, reflection) framework guidance, 2025. URL: https: //www.kcl.ac.uk/about/strategy/learning-and-teaching/ai-guidance/pair-framework-guidance. [31] P. Dey, S. Merugu, S. Kaveri, Uncertainty-Aware Fusion: An Ensemble Framework for Mitigating Hallucinations in Large Language Models, in: Companion Proceedings of the ACM on Web Conference 2025, WWW ’25, Association for Computing Machinery, New York, NY, USA, 2025, p. 947–951. doi:10.1145/3701716.3715523. [32] F. Abrar, S. Mehmood, S. Kinza ul Emman, A. N. Kandhro, Cognitive Development and AI: A Longitudinal Study of Children and Adults Navigating Problem-Solving with AI Tools, The Critical Review of Social Sciences Studies 3 (2025) 1888–1904. doi:10.59075/rw8zd496. [33] R. Toscano, M. A. Guerra, S. Durán-Ballén, B. M. Valarezo, WIP - Development of Critical Thinking in AEC Students Aided by Artificial Intelligence, in: Proceedings - Frontiers in Education Conference, FIE, 2024. doi:10.1109/FIE61694.2024.10893092. [34] T. Poell, L. Maas, M. Balk, M. van Haastrecht, SchrijfBlik: Safeguarding the Validity of Writing Assessment in the Age of AI, in: A. I. Cristea, E. Walker, Y. Lu, O. C. Santos, S. Isotani (Eds.), Artificial Intelligence in Education. Posters and Late Breaking Results, Workshops and Tutorials, Industry and Innovation Tracks, Practitioners, Doctoral Consortium, Blue Sky, and WideAIED, Springer Nature Switzerland, Cham, 2025, pp. 71–80. doi:10.1007/978-3-031-99261-2_6. [35] Z. Swiecki, H. Khosravi, G. Chen, R. Martinez-Maldonado, J. M. Lodge, S. Milligan, N. Selwyn, D. Gašević, Assessment in the age of artificial intelligence, Computers and Education: Artificial Intelligence 3 (2022) 100075. doi:10.1016/j.caeai.2022.100075. [36] X. Wang, C. Rose, K. Koedinger, Seeing Beyond Expert Blind Spots: Online Learning Design for Scale and Quality, in: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, CHI ’21, Association for Computing Machinery, New York, NY, USA, 2021. doi:10.1145/3411764.3445045. [37] J. Peng, W. Shen, J. Rao, J. Lin, Automated Bias Assessment in AI-Generated Educational Content Using CEAT Framework, in: A. I. Cristea, E. Walker, Y. Lu, O. C. Santos, S. Isotani (Eds.), Artificial Intelligence in Education. Posters and Late Breaking Results, Workshops and Tutorials, Industry and Innovation Tracks, Practitioners, Doctoral Consortium, Blue Sky, and WideAIED, Springer Nature Switzerland, Cham, 2025, pp. 352–358. doi:10.1007/978-3-031-99264-3_44. [38] A. Al-Zawqari, D. Peumans, G. Vandersteen, Latent space bias mitigation for predicting at-risk students, Computers and Education: Artificial Intelligence 7 (2024) 100300. doi: 10.1016/j. caeai.2024.100300. [39] E. Misino, R. Calegari, M. Lombardi, M. Milano, Ensuring Fairness Stability for Disentangling Social Inequality in Access to Education: the FAiRDAS General Method, in: K. Larson (Ed.), Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence AI for Good, International Joint Conferences on Artificial Intelligence Organization, 2024, pp. 7412–7420. doi:10.24963/ijcai.2024/820. [40] J. Chanenson, B. Sloane, N. Rajan, A. Morril, J. Chee, D. Y. Huang, M. Chetty, Uncovering Privacy and Security Challenges In K-12 Schools, in: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI ’23, Association for Computing Machinery, New York, NY, USA, 2023. doi:10.1145/3544548.3580777. [41] K. M. L. Jones, A. VanScoy, K. Bright, A. Harding, Do They Even Care? Measuring Instructor Value of Student Privacy in the Context of Learning Analytics, in: Proceedings of the 54th Hawaii International Conference on System Sciences, 2021, pp. 1529–1637. doi:10.24251/HICSS.2021. 185. [42] M. Saarela, S. Gunasekara, A. Karimov, The EU AI Act: Implications for Ethical AI in Education, in: S. Chatterjee, J. vom Brocke, R. Anderson (Eds.), Local Solutions for Global Challenges, volume 15704 of Lecture Notes in Computer Science, Springer Nature Switzerland, Cham, 2025, pp. 36–50. doi:10.1007/978-3-031-93979-2_3. [43] J. Schleiss, K. Günther, S. Stober, Protecting Student Data in ML Pipelines: An Overview of PrivacyPreserving ML, in: M. M. Rodrigo, N. Matsuda, A. I. Cristea, V. Dimitrova (Eds.), Artificial Intelligence in Education. Posters and Late Breaking Results, Workshops and Tutorials, Industry and Innovation Tracks, Practitioners’ and Doctoral Consortium, volume 13456 of Lecture Notes in Computer Science, Springer International Publishing, Cham, 2022, pp. 532–536. doi:10.1007/ 978-3-031-11647-6_109. [44] S. Isotani, I. I. Bittencourt, G. C. Challco, D. Dermeval, R. F. Mello, AIED Unplugged: Leapfrogging the Digital Divide to Reach the Underserved, in: N. Wang, G. Rebolledo-Mendez, V. Dimitrova, N. Matsuda, O. C. Santos (Eds.), Artificial Intelligence in Education. Posters and Late Breaking Results, Workshops and Tutorials, Industry and Innovation Tracks, Practitioners, Doctoral Consortium and Blue Sky, volume 1831 of Communications in Computer and Information Science, Springer Nature Switzerland, Cham, 2023, pp. 772–779. doi:10.1007/978-3-031-36336-8_118. [45] A. Oulamine, R. Chakra, R. Ziky, N. Mounir, M. Cherkaoui, A. Benjelloun, A Systematic Literature Review of Barriers Afecting e-Learning in Higher Education, Educational Process: International Journal 17 (2025) e2025396. doi:10.22521/edupij.2025.17.396. [46] F. Martin, Y. Chen, R. L. Moore, C. D. Westine, Systematic review of adaptive learning research designs, context, strategies, and technologies from 2009 to 2018, Educational Technology Research and Development 68 (2020) 1903–1929. doi:10.1007/s11423-020-09793-2. [47] N. S. Alotaibi, The Impact of AI and LMS Integration on the Future of Higher Education,
Sustainability 16 (2024) 10357. doi:10.3390/su162310357. [48] Y. Nagae, L. Zhang, L. Farias Herrera, The Efects of Professional Development Training on Teachers’ AI Literacy, in: A. I. Cristea, E. Walker, Y. Lu, O. C. Santos, S. Isotani (Eds.), Artificial Intelligence in Education, volume 15877 of Lecture Notes in Computer Science, Springer Nature Switzerland, Cham, 2025, pp. 368–380. doi:10.1007/978-3-031-98414-3_26. [49] K. Oyetade, T. Zuva, Advancing Equitable Education with Inclusive AI to Mitigate Bias and Enhance Teacher Literacy, Educational Process: International Journal 14 (2025) e2025087. doi:10. 22521/edupij.2025.14.87. [50] E. Nisiforou, P. Kosmas, Assessing Institutional Readiness for Emerging Technologies Integration in Higher Education, Journal of Interactive Learning Research 35 (2024) 425–450. doi:10.70725/ 441907jrwrsw. [51] R. J. Farrell, R. G. Farrell, J. Grissett, J. Winter, A. Pridemore, Developing institutional policies for ai in education, in: J. R. Corbeil, M. E. Corbeil (Eds.), Teaching and Learning in the Age of Generative AI: Evidence-Based Approaches to Pedagogy, Ethics, and Beyond, Routledge, New York, 2025, pp. 143–162. doi:10.4324/9781032688602. [52] W. Y. Leong, J. B. Zhang, Ethical Design of AI for Education and Learning Systems, ASM Science
Journal 20 (2025). doi:10.32802/asmscj.2025.1917. [53] M. Romero, T. Frosig, A. M. L. Taylor-Beswick, J. Laru, B. Bernasco, A. Urmeneta, O. Strutynska, M.A. Girard, Manifesto in Defence of Human-Centred Education in the Age of Artificial Intelligence, in: A. Urmeneta, M. Romero (Eds.), Creative Applications of Artificial Intelligence in Education, volume 12 of Palgrave Studies in Creativity and Culture, Springer Nature Switzerland, Cham, 2024, pp. 157–178. doi:10.1007/978-3-031-55272-4_12. [54] G. Y. Yingsoon, S. Zhang, N. A. Chua, Empowering Educators in the Era of Next-Generation AI: Redefining the Teacher’s Role in an AI-Enhanced Learning Environment, in: E. Çela, N. R. Vajjhala, M. M. Fonkam (Eds.), Teachers’ Roles and Perspectives on AI Integration in Schools, IGI Global Scientific Publishing, Hershey, PA, 2025, pp. 1–30. doi: 10.1080/0144929X.2024.2394886. [55] L. Shi, I. Choi, 321A Systematic Review on Artificial Intelligence in Supporting Teaching Practice: Application Types, Pedagogical Roles, and Technological Characteristics, in: Uses of Artificial Intelligence in STEM Education, Oxford University Press, 2024. doi: 10.1093/oso/ 9780198882077.003.0015. [56] M. I. Sholeh, Educational Transformation Through Artificial Intelligence: Implementation of AI Tools in the Teaching and Learning Process, in: V. P. H. Pham, A. Lian, A. Lian, S. R. Barros (Eds.), Implementing AI Tools for Language Teaching and Learning, IGI Global Scientific Publishing, Hershey, PA, 2025, pp. 25–40. doi:10.4018/979-8-3693-7260-9.ch002. [57] A. A. von Davier, J. Burstein, AI in the Assessment Ecosystem: A Human–Centered AI Perspective, in: P. Ilic, I. Casebourne, R. Wegerif (Eds.), Artificial Intelligence in Education: The Intersection of Technology and Pedagogy, volume 261 of Intelligent Systems Reference Library, Springer Nature Switzerland, Cham, 2024, pp. 93–109. doi:10.1007/978-3-031-71232-6_6. [58] H. Manohara, A. Gummadi, K. Santosh, S. Vaitheeshwari, S. S. C. Mary, B. K. Bala, Human Centric Explainable AI for Personalized Educational Chatbots, in: 10th International Conference on Advanced Computing and Communication Systems, ICACCS 2024, 2024, pp. 328–334. doi:10. 1109/ICACCS60874.2024.10716907. [59] S. Hariharasitaraman, A. Jayavel, S. Gnanasekaran, S. S. Sohail, D. Øivind Madsen, Interpretable and ethical learning assessment transformer (IELAT): an explainable transformer model for personalized student assessments, Cogent Education 12 (2025) 2485508. doi:10.1080/2331186X. 2025.2485508. [60] Y. Jin, R. Martinez-Maldonado, D. Gašević, L. Yan, GLAT: The generative AI literacy assessment test, Computers and Education: Artificial Intelligence 9 (2025) 100436. doi: 10.1016/j.caeai. 2025.100436. [61] S. Snider, Ethical Foundations for the Use of AI in the Education of Young Children: A HumanCentered, Rights-Based Framework for Teachers and Caregivers, in: J. J. Araujo, S. Snider (Eds.), Responsible AI Integration in Education, IGI Global Scientific Publishing, Hershey, PA, 2025, pp. 241–274. doi:10.4018/979-8-3373-6007-2.ch008. [62] M. A. Chaudhry, M. Cukurova, R. Luckin, A Transparency Index Framework for AI in Education, in: Artificial Intelligence in Education. Posters and Late Breaking Results, Workshops and Tutorials, Industry and Innovation Tracks, Practitioners’ and Doctoral Consortium: 23rd International Conference, AIED 2022, Durham, UK, July 27–31, 2022, Proceedings, Part II, Springer-Verlag, Berlin, Heidelberg, 2022, p. 195–198. doi:10.1007/978-3-031-11647-6_33. [63] M. Madaio, S. L. Blodgett, E. Mayfield, E. Dixon-Román, Beyond “fairness”: Structural (in)justice lenses on AI for education , in: W. Holmes, K. Porayska-Pomsta (Eds.), The Ethics of Artificial Intelligence in Education: Practices, Challenges, and Debates, Routledge, New York, 2022. doi:10. 4324/9780429329067. [64] M.-A. Sørlie, Structural, cultural and instructional predictors essential to sustained implementation fidelity in schools: The School-Wide Positive Behavior Support Model (SWPBS), International Journal of Educational Research Open 2 (2021) 100082. doi:10.1016/j.ijedro.2021.100082. [65] O. V. Barna, O. H. Kuzminska, S. O. Semerikov, Enhancing digital competence through STEMintegrated universal design for learning: a pedagogical framework for computer science education in Ukrainian secondary schools, Discover Education 4 (2025) 357. doi:10.1007/ s44217-025-00821-y. [66] R. O. Teles, F. T. Moreira, Integrating AI in Formative Assessment Under UDL Principles in Diverse Learning Environments, in: F. Moreira, R. Teles (Eds.), Improving Student Assessment With Emerging AI Tools, IGI Global Scientific Publishing, Hershey, PA, 2025, pp. 95–122. doi: 10. 4018/979-8-3693-6170-2.ch004. [67] M. Nyaaba, X. Zhai, Developing a Theory-Grounded AI Tool for the Generation of Culturally Responsive Lesson Plans, Computers and Education: Artificial Intelligence (2025) 100474. doi: 10. 1016/j.caeai.2025.100474. [68] M.-Y. Lin, Y.-S. Chang, Efects of design thinking STEAM instruction on AI learning and creativity, International Journal of Technology and Design Education (2025). doi:10.1007/ s10798-025-09977-y. [69] G. Hanshaw, C. Sullivan, Exploring barriers to AI course assistant adoption: a mixed-methods study on student non-utilization, Discover Artificial Intelligence 5 (2025) 178. doi: 10.1007/ s44163-025-00312-x. [70] S. Yadav, Empowering Educators for the Artificial Intelligence (AI) Era: Strategies and Challenges in Integrating Computational Thinking Into Professional Development, in: S. Papadakis, M. Kalogiannakis (Eds.), Empowering STEM Educators With Digital Tools, IGI Global Scientific Publishing, Hershey, PA, 2024, pp. 1–28. doi:10.4018/979-8-3693-9806-7.ch001. [71] Z. R. Alhalalmeh, Y. M. Fouda, M. A. Rushdi, M. El-Mikkawy, Automating Assessment and Providing Personalized Feedback in E-Learning: The Power of Template Matching, Sustainability 15 (2023) 14234. doi:10.3390/su151914234. [72] F. I. Anaroua, Q. Li, Y. Tang, H. P. Liu, AI-driven formative assessment and adaptive learning in data-science education: Evaluating an LLM-powered virtual teaching assistant, 2025. URL: https://arxiv.org/abs/2509.20369. arXiv:2509.20369. [73] V. Kasapidis, A. Petrou, N. P. Konstantinidi, K. Kostalias, E. Kakogianni, AI and Data Literacy: Challenges and Opportunities for Teachers and Students, European Journal of Education and Pedagogy 6 (2025) 7–11. doi:10.24018/ejedu.2025.6.3.957. [74] Thimmanna, A. V. N. S. Sharma, M. S. Naik, S. Radhakrishnan, A. Sharma, Personalized learning paths: Adapting education with ai-driven curriculum, European Economic Letters (EEL) 14 (2024) 31–40. URL: https://www.eelet.org.uk/index.php/journal/article/view/993. doi:10.52783/eel. v14i1.993. [75] OECD, Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education (Review draft, OECD, Paris, 2025. URL: https://ailiteracyframework.org/ wp-content/uploads/2025/05/AILitFramework_ReviewDraft.pdf. [76] A. Valentini, A. Blancas, The challenges of AI in higher education and institutional responses: is there room for competency frameworks?, volume 12 of IESALC. Working papers, UNESCO, Paris, 2025. [77] S. Q. Dolba, F. A. Inoncillo, J. L. Nunez, Evaluating the Impact of Ai-Generated Outputs on Student Assessment: Educator’s Perspective, International Journal of Artificial Intelligence Research 9 (2025). doi:10.29099/ijair.v9i1.1347. [78] M. Johnpaul, R. S. B. Miryala, G. Jayaprakashnarayana, R. K. Miryala, Pros and cons of artificial intelligence on metacognition: A myopic state with long-term consequences on human learning, in: S. P. J. Christydass, N. Nurhayati, S. Kannadhasan (Eds.), Hybrid and Advanced Technologies: Proceedings of the International Conference on Hybrid and Advanced Technologies (ICHAT 2024), April 26-28, 2024, Ongole, Andhra Pradesh, India (Volume 2), CRC Press, London, 2025, pp. 98–103. doi:10.1201/9781003559139-13. [79] A. Teixeira de Melo, L. Renault, L. S. D. Caves, P. Garnett, P. D. Lopes, R. Ribeiro, F. Santos, An AI tool for scafolding complex thinking: challenges and solutions in developing an LLM prompt protocol suite, Cognition, Technology & Work 27 (2025) 651–693. doi:10.1007/ s10111-025-00817-6. [80] M. A. S. Khasawneh, A. Aladini, S. A. Assi, B. Ajanil, Portfolio assessment in AI-enhanced learning environments: a pathway to emotion regulation, mindfulness, and language learning attitudes, Language Testing in Asia 15 (2025) 5. doi:10.1186/s40468-025-00345-0. [81] G. Thiyagarajan, P. S, A Process Mining approach to analyze learning behavior in the flipped classroom, in: 2021 2nd International Conference on Communication, Computing and Industry 4.0 (C2I4), 2021, pp. 1–7. doi:10.1109/C2I454156.2021.9689228. [82] B. Laker, What A True Growth Mindset Looks Like In The Age Of AI, 2025. URL: https://www.forbes.com/sites/benjaminlaker/2025/07/17/ what-a-true-growth-mindset-looks-like-in-the-age-of-ai/. [83] Y. Shao, Human in the Era of Artificial Intelligence and Beyond, The Journal of Mind and
Behavior 46 (2025) 73–102. URL: https://jmb-online.com/article-view.php/?id=94. [84] P. Faella, S. Digennaro, A. Iannaccone, Educational practices in motion: a scoping review of embodied learning approaches in school, Frontiers in Education 10 (2025) 1568744. doi:10.3389/ feduc.2025.1568744. [85] H. Fakour, M. Imani, Socratic wisdom in the age of AI: a comparative study of ChatGPT and human tutors in enhancing critical thinking skills, Frontiers in Education 10 (2025). doi:10. 3389/feduc.2025.1528603. [86] G. Trajkovski, H. Hayes, The Future of AI in Educational Assessment, in: AI-Assisted Assessment in Education: Transforming Assessment and Measuring Learning, Springer Nature Switzerland, Cham, 2025, pp. 347–380. doi:10.1007/978-3-031-88252-4_9. [87] P. Villalobos Dintrans, T. J. Bossert, J. Sherry, M. E. Kruk, A synthesis of implementation science frameworks and application to global health gaps, Global Health Research and Policy 4 (2019) 25. doi:10.1186/s41256-019-0115-1. [88] S. Potthof, T. Finch, L. Bührmann, A. Etzelmüller, C. R. van Genugten, M. Girling, C. R. May, N. Perkins, C. Vis, T. Rapley, Towards an Implementation-STakeholder Engagement Model (I-STEM) for improving health and social care services, Health Expectations 26 (2023) 1997–2012. doi:10.1111/hex.13808. [89] A. Sridhar, A. Drahota, J. E. Tschida, Implementation strategy mapping methods to improve autism intervention use in community settings: a study protocol, Implementation Science Communications 3 (2022) 92. doi:10.1186/s43058-022-00339-6. [90] V. Yakovchenko, S. S. Rogal, D. E. Goodrich, C. Lamorte, B. Neely, M. Merante, S. Gibson, D. Scott, H. McCurdy, A. Nobbe, T. R. Morgan, M. J. Chinman, Getting to implementation: Adaptation of an implementation playbook, Frontiers in Public Health 10 (2023) 980958. doi:10.3389/fpubh. 2022.980958. [91] P. K. Imbrie, C. Malave, K. Watson, Pedagogy versus reality: how past experiences can be an efective modeling tool to successfully deploy curricula changes, in: Proceedings Frontiers in Education 1997 27th Annual Conference. Teaching and Learning in an Era of Change, volume 3, 1997, p. 1435. doi:10.1109/FIE.1997.632695. [92] T. M. Jensen, A. J. Metz, M. E. Disbennett, A. B. Farley, Developing a practice-driven research agenda in implementation science: Perspectives from experienced implementation support practitioners, Implementation Research and Practice 4 (2023) 26334895231199063. doi:10.1177/ 26334895231199063. [93] J. Zhao, W. Chen, W. Bai, X. Zhang, R. Hui, S. Chen, G. Fontaine, X. Wei, N. Zhang, I. D. Graham, Research priority setting for implementation science and practice: a living systematic review protocol, Systematic Reviews 14 (2025) 51. doi:10.1186/s13643-025-02786-3. [94] D. J. Bowen, A. Heald, E. LePoire, A. Jones, D. Gadbois, J. Russo, J. Carruthers, Population-based implementation of behavioral health detection and treatment into primary care: early data from New York state, BMC Health Services Research 21 (2021) 922. doi:10.1186/s12913-021-06892-5. [95] J. Whitfield, S. Owens, A. Bhat, L. Chwastiak, Successful ingredients of efective Collaborative Care programs in low- and middle-income countries: A rapid review, Global Mental Health 10 (2023) e11. doi:10.1017/gmh.2022.60. [96] P. K. Namyalo, B. K. Wodnik, O. Michaelides, E. Di Ruggiero, Identifying implementation science research and policy priorities to advance universal health coverage: A multi-country modified Delphi study, BMJ Global Health 10 (2025) e018562. doi:10.1136/bmjgh-2024-018562. [97] H. Saetren, P. L. Hupe, Policy Implementation in an Age of Governance, in: E. Ongaro, S. Van Thiel (Eds.), The Palgrave Handbook of Public Administration and Management in Europe, Palgrave Macmillan UK, London, 2018, pp. 553–575. doi:10.1057/978-1-137-55269-3_29. [98] N. Okada, R. Goble, K. Seo, G. Han, Implementation Gaps Are Persistent Phenomena in Disaster Risk Management: A Perspective Developed After Discussions at IDRiM 2022, Journal of Integrated Disaster Risk Management 13 (2023) 69–99. doi:10.5595/001c.90964. [99] A. Susanty, N. B. Puspitasari, S. D. Mariana, L. O. Nazaruddin, Mapping the barrier of halal food industry performance and the policy insight: Combination of ISM-DEMATEL approach, Engineering Management in Production and Services 16 (2024) 69–95. doi:10.2478/emj-2024-0034. [100] P. Aragonés-Beltrán, A. León-Camargo, M. C. González-Cruz, M.-Á. Artacho-Ramírez, Proposal for adapting the DEMATEL method to stakeholder analysis, in: 25th International Congress on Project Management and Engineering. Alcoi, 6th – 9th July 2021, 2021, pp. 206–217. URL: http://dspace.aeipro.com/xmlui/handle/123456789/3084. [101] Ethical guidelines for educators on using artificial intelligence, 2025.
URL: https://education.ec.europa.eu/focus-topics/digital-education/action-plan/ ethical-guidelines-for-educators-on-using-ai. [102] A. V. Y. Lee, E. Koh, C. K. Looi, AI in Education and Learning Analytics in Singapore: An Overview of Key Projects and Initiatives, Information and Technology in Education and Learning 3 (2023) Inv–p001. doi:10.12937/itel.3.1.Inv.p001. [103] E. Oluwagbade, The Impact of AI Liability Laws on Innovation: Balancing Risk and Responsibility
Across Jurisdictions, 2025. URL: https://www.researchgate.net/publication/391950937. [104] B. R. Ruben, B. Sakthivel, S. P. Guru, J. Sanakyan, V. Sakthivel, Prioritization of barriers of lean six sigma with sustainability cogitations, AIP Conference Proceedings 2527 (2022) 040012. doi:10.1063/5.0108106. [105] A. Friedman, A. Schuster, Data mining with diferential privacy, in: Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’10, Association for Computing Machinery, New York, NY, USA, 2010, p. 493–502. doi:10.1145/ 1835804.1835868. [106] P. Upmann, Education AI Governance Framework, 2025. URL: https: //aign.global/ai-governance-framework/global-ai-governance-framework/ education-ai-governance-framework/. [107] X. Zhang, Fairness and Efectiveness in AI-Driven Educational Assessments: Challenges and Mitigation Strategies, Journal of Innovation and Development 11 (2025) 7–10. doi:10.54097/ adcxze34. [108] I. Ulnicane, Intersectionality in Artificial Intelligence: Framing Concerns and Recommendations for Action, Social Inclusion 12 (2024). doi:10.17645/si.7543. [109] D. Kennedy-Mayo, J. Gord, “Model Cards for Model Reporting” in 2024: Reclassifying Category of Ethical Considerations in Terms of Trustworthiness and Risk Management, in: K. Arai (Ed.), Advances in Information and Communication - Proceedings of the 2025 Future of Information and Communication Conference (FICC), Volume 1, Berlin, Germany, 27-28 April 2025, volume 1283 of Lecture Notes in Networks and Systems, Springer, 2025, pp. 179–196. doi:10.1007/978-3-031-84457-7_11. [110] International Society for Technology in Education (ISTE), Artificial Intelligence in Education, 2025. URL: https://iste.org/ai. [111] R. Kayalvizhi, C. A. Kumar, M. Karthick, Lifelong Education and Micro-Credentialing in Artificial Intelligence: Transforming Learning for the Digital Age, in: M. Karthick, C. A. Kumar (Eds.), Higher Education 4.0: Shaping Learning in the Digital Era, Shanlax Publications, Vasantha Nagar, Madurai, 2025, pp. 95–104. URL: https://www.researchgate.net/publication/394469260. [112] Z. J. Zhao, J. Li, Y. Hong, T. Y. Yun, Lesson study as an approach to facilitate the integration of Gen-AI into EFL curriculum design in higher education, International Journal for Lesson and Learning Studies 14 (2025) 253–279. doi:10.1108/IJLLS-01-2025-0029. [113] J. Cao, Y. Wang, International Cooperation Among Artificial Intelligence Research Teams Based on Regional Cooperation Models, Data and Information Management 5 (2021) 147–158. doi:10.2478/dim-2020-0036. [114] M. Sposato, A call for caution and evidence–based research on the impact of artificial intelligence in education, Quality Education for All 2 (2025) 158–170. doi:10.1108/QEA-09-2024-0087. [115] J. Hautamäki, S. Kupiainen, Learning to learn in Finland, in: R. D. Crick, C. Stringher, K. Ren (Eds.), Learning to Learn: International perspectives from theory and practice, Routledge, London, 2014, p. 170–194. doi:10.4324/9780203078044-9. [116] R. Kaliisa, K. Misiejuk, S. López-Pernas, M. Saqr, How does artificial intelligence compare to human feedback? A meta-analysis of performance, feedback perception, and learning dispositions, Educational Psychology (2025). doi:10.1080/01443410.2025.2553639. [117] A. O. Savi, N. t. Broeke, A. D. Hofman, Adaptive Learning Systems and Interference in Causal
Inference, 2021. doi:10.31234/osf.io/4ax25. [118] W. W. Loh, D. Ren, A Tutorial on Causal Inference in Longitudinal Data With Time-Varying Confounding Using G-Estimation, Advances in Methods and Practices in Psychological Science 6 (2023) 25152459231174029. doi:10.1177/25152459231174029. [119] V. Hassija, V. Chamola, A. Mahapatra, A. Singal, D. Goel, K. Huang, S. Scardapane, I. Spinelli, M. Mahmud, A. Hussain, Interpreting Black-Box Models: A Review on Explainable Artificial Intelligence, Cognitive Computation 16 (2024) 45–74. doi:10.1007/s12559-023-10179-8. [120] R. S. Baker, A. Hawn, Algorithmic Bias in Education, International Journal of Artificial Intelligence in Education 32 (2022) 1052–1092. doi:10.1007/s40593-021-00285-9. [121] A. A. Skaiky, H. M. S. Ali, A. Mohammed, Z. A. Mahdi, Comprehensive Bias Mitigation in AI: Evaluating Pre-Processing, In-Processing, and Post-Processing Techniques for Fair DecisionMaking, in: 2025 IEEE 4th International Conference on Computing and Machine Intelligence (ICMI), 2025, pp. 1–5. doi:10.1109/ICMI65310.2025.11141086. [122] U. Aïvodji, F. Bidet, S. Gambs, R. C. Ngueveu, A. Tapp, Local Data Debiasing for Fairness Based on Generative Adversarial Training, Algorithms 14 (2021) 87. doi:10.3390/a14030087. [123] N. M. van Atteveldt, M. van Kesteren, B. Braams, L. Krabbendam, Neuroimaging of learning and development: improving ecological validity, Frontline Learning Research 6 (2018) 186–203. doi:10.14786/flr.v6i3.366. [124] G. Dietz, J. Outa, L. Lowe, J. A. Landay, H. Gweon, Theory of AI Mind: How adults and children reason about the “mental states” of conversational AI, in: M. Goldwater, F. K. Anggoro, B. K. Hayes, D. C. Ong (Eds.), Proceedings of the 45th Annual Conference of the Cognitive Science Society, 2023, pp. 3367–3374. URL: https://escholarship.org/uc/item/8x88d8tp. [125] E. Granado De la Cruz, F. J. Gago-Valiente, Ó. Gavín-Chocano, E. Pérez-Navío, Education, Neuroscience, and Technology: A Review of Applied Models, Information 16 (2025) 664. doi:10. 3390/info16080664. [126] A. Rahmath Nisha, A. K. Subramani, K. C. Ranjithaa, Emotion AI for Human-Centered Leadership and Training, in: J. Mrabet, S. Boudagga, R. Sharma, Bansal (Eds.), Harnessing Emotion AI for Customer Support and Employee Wellbeing, IGI Global Scientific Publishing, Hershey, PA, 2025, pp. 383–416. doi:10.4018/979-8-3373-3658-9.ch015. [127] W. Hu, Z. Shao, Design and evaluation of a GenAI-based personalized educational content system tailored to personality traits and emotional responses for adaptive learning, Computers in Human Behavior Reports 19 (2025) 100735. doi:10.1016/j.chbr.2025.100735. [128] C. D. Jaldi, E. Ilkou, N. Schroeder, C. Shimizu, Education in the era of Neurosymbolic AI, Journal of Web Semantics 85 (2025) 100857. doi:10.1016/j.websem.2024.100857. [129] C. D. Jaldi, E. Ilkou, N. Schroeder, C. Shimizu, Education in the era of Neurosymbolic AI, Journal of Web Semantics 85 (2025) 100857. doi:10.1016/j.websem.2024.100857. [130] R. González-Sendino, E. Serrano, J. Bajo, Mitigating bias in artificial intelligence: Fair data generation via causal models for transparent and explainable decision-making, Future Generation Computer Systems 155 (2024) 384–401. doi:10.1016/j.future.2024.02.023. [131] C. Aihua, T. Cheng-Chung, Bridging the Gap: A systematic review of AI-powered smart learning Systems for addressing diverse student learning needs, Edelweiss Applied Science and Technology 9 (2025) 1423–1436. doi:10.55214/25768484.v9i4.6309. [132] E. Bandara, S. Shetty, A. Rahman, R. Mukkamala, J. Zhao, X. Liang, Bassa-ML - a blockchain and model card integrated federated learning provenance platform, in: Proceedings - IEEE Consumer Communications and Networking Conference, CCNC, 2022, pp. 753–759. doi:10. 1109/CCNC49033.2022.9700513. [133] Z. Huang, C. Cao, Federated Learning for AI-Assisted Education: Privacy-Preserving and CrossPlatform Collaborative Modeling in Higher Education, in: Proceedings of the 2nd International Conference on Machine Intelligence and Digital Applications, MIDA ’25, Association for Computing Machinery, New York, NY, USA, 2025, p. 146–151. doi:10.1145/3744464.3744487. [134] N. Saber Rashid, H. Maseeh Yasin, Privacy-preserving machine learning: a review of federated learning techniques and applications, International Journal of Scientific World 11 (2025) 30–39. doi:10.14419/af03y111. [135] L. V. Lehka, S. V. Shokaliuk, Quantum programming is a promising direction of IT development, in: A. E. Kiv, S. O. Semerikov, V. N. Soloviev, A. M. Striuk (Eds.), Proceedings of the 1st Student Workshop on Computer Science & Software Engineering, Kryvyi Rih, Ukraine, November 30, 2018, volume 2292 of CEUR Workshop Proceedings, CEUR-WS.org, 2018, pp. 76–82. URL: https: //ceur-ws.org/Vol-2292/paper07.pdf. [136] P. Gao, M. Adnan, Overview of emerging electronics technologies for artificial intelligence: A review, Materials Today Electronics 11 (2025) 100136. doi:10.1016/j.mtelec.2025.100136. [137] P. V. Zahorodko, S. O. Semerikov, V. N. Soloviev, A. M. Striuk, M. I. Striuk, H. M. Shalatska, Comparisons of performance between quantum-enhanced and classical machine learning algorithms on the IBM Quantum Experience, Journal of Physics: Conference Series 1840 (2021) 012021. doi:10.1088/1742-6596/1840/1/012021. [138] P. V. Zahorodko, Y. O. Modlo, O. O. Kalinichenko, T. V. Selivanova, S. O. Semerikov, Quantum enhanced machine learning: An overview, in: A. E. Kiv, S. O. Semerikov, V. N. Soloviev, A. M. Striuk (Eds.), Proceedings of the 3rd Workshop for Young Scientists in Computer Science & Software Engineering (CS&SE@SW 2020), Kryvyi Rih, Ukraine, November 27, 2020, volume 2832 of CEUR Workshop Proceedings, CEUR-WS.org, 2020, pp. 94–103. URL: https://ceur-ws.org/ Vol-2832/paper13.pdf. [139] M. L. Dalla Chiara, R. Giuntini, G. Sergioli, A Quantum Approach to Pattern Recognition and Machine Learning. Part II, International Journal of Theoretical Physics 63 (2024) 44. doi:10. 1007/s10773-024-05567-1. [140] M. Ponce, R. Herrman, P. C. Lotshaw, S. Powers, G. Siopsis, T. S. Humble, J. Ostrowski, Graph decomposition techniques for solving combinatorial optimization problems with variational quantum algorithms, Quantum Information Processing 24 (2025) 60. doi:10.1007/ S11128-025-04675-Z. [141] A. Drigas, E. Mitsea, C. Skianis, Meta-Learning: A Nine-Layer Model Based on Metacognition and Smart Technologies, Sustainability 15 (2023) 1668. doi:10.3390/su15021668. [142] K. Khotimah, Rusijono, A. Mariono, Enhancing Metacognitive and Creativity Skills through AI-Driven Meta-Learning Strategies, International Journal of Interactive Mobile Technologies (iJIM) 18 (2024) 18–31. doi:10.3991/ijim.v18i05.47705. [143] H. Guo, J. Zhou, Z. Liu, MACRec: A Meta-learning Enhanced Model for Academic Collaborator Recommendation, in: 2025 28th International Conference on Computer Supported Cooperative Work in Design (CSCWD), 2025, pp. 978–983. doi:10.1109/CSCWD64889.2025.11033271. [144] Q. Tan, Reimagining teacher development in the era of generative AI: A scoping review, Teaching and Teacher Education 168 (2025) 105236. doi:10.1016/j.tate.2025.105236. [145] T. H. Kolomoiets, D. A. Kassim, Using the Augmented Reality to Teach of Global Reading of Preschoolers with Autism Spectrum Disorders, in: A. E. Kiv, V. N. Soloviev (Eds.), Proceedings of the 1st International Workshop on Augmented Reality in Education, Kryvyi Rih, Ukraine, October 2, 2018, volume 2257 of CEUR Workshop Proceedings, CEUR-WS.org, 2018, pp. 237–246.
URL: https://ceur-ws.org/Vol-2257/paper24.pdf. [146] T. A. Vakaliuk, S. I. Pochtoviuk, Analysis of tools for the development of augmented reality technologies, in: S. H. Lytvynova, S. O. Semerikov (Eds.), Proceedings of the 4th International Workshop on Augmented Reality in Education (AREdu 2021), Kryvyi Rih, Ukraine, May 11, 2021, volume 2898 of CEUR Workshop Proceedings, CEUR-WS.org, 2021, pp. 119–130. URL: https: //ceur-ws.org/Vol-2898/paper06.pdf. [147] N. O. Zinonos, E. V. Vihrova, A. V. Pikilnyak, Prospects of Using the Augmented Reality for Training Foreign Students at the Preparatory Departments of Universities in Ukraine, in: A. E. Kiv, V. N. Soloviev (Eds.), Proceedings of the 1st International Workshop on Augmented Reality in Education, Kryvyi Rih, Ukraine, October 2, 2018, volume 2257 of CEUR Workshop Proceedings, CEUR-WS.org, 2018, pp. 87–92. URL: https://ceur-ws.org/Vol-2257/paper10.pdf. [148] T. Tene, J. A. Marcatoma Tixi, M. d. L. Palacios Robalino, M. J. Mendoza Salazar, C. Vacacela Gomez, S. Bellucci, Integrating immersive technologies with STEM education: a systematic review, Frontiers in Education 9 (2024) 1410163. doi:10.3389/feduc.2024.1410163. [149] J. Izquierdo Domenech, J. Linares Pellicer, I. Ferri Molla, Virtual Reality and Language Models, a New Frontier in Learning, International Journal of Interactive Multimedia and Artificial Intelligence 8 (2024) 46–54. doi:10.9781/ijimai.2024.02.007. [150] O. O. Lavrentieva, I. O. Arkhypov, O. P. Krupski, D. O. Velykodnyi, S. V. Filatov, Methodology of using mobile apps with augmented reality in students’ vocational preparation process for transport industry, in: O. Y. Burov, A. E. Kiv (Eds.), Proceedings of the 3rd International Workshop on Augmented Reality in Education, Kryvyi Rih, Ukraine, May 13, 2020, volume 2731 of CEUR Workshop Proceedings, CEUR-WS.org, 2020, pp. 143–162. URL: https://ceur-ws.org/Vol-2731/ paper07.pdf. [151] V. V. Babkin, V. V. Sharavara, V. V. Sharavara, V. V. Bilous, A. V. Voznyak, S. Y. Kharchenko, Using augmented reality in university education for future IT specialists: educational process and student research work, in: S. H. Lytvynova, S. O. Semerikov (Eds.), Proceedings of the 4th International Workshop on Augmented Reality in Education (AREdu 2021), Kryvyi Rih, Ukraine, May 11, 2021, volume 2898 of CEUR Workshop Proceedings, CEUR-WS.org, 2021, pp. 255–268.
URL: https://ceur-ws.org/Vol-2898/paper14.pdf. [152] S. P. Palamar, G. V. Bielienka, T. O. Ponomarenko, L. V. Kozak, L. L. Nezhyva, A. V. Voznyak, Formation of readiness of future teachers to use augmented reality in the educational process of preschool and primary education, in: S. H. Lytvynova, S. O. Semerikov (Eds.), Proceedings of the 4th International Workshop on Augmented Reality in Education (AREdu 2021), Kryvyi Rih, Ukraine, May 11, 2021, volume 2898 of CEUR Workshop Proceedings, CEUR-WS.org, 2021, pp. 334–350. URL: https://ceur-ws.org/Vol-2898/paper18.pdf. [153] T. H. Kramarenko, O. S. Pylypenko, M. V. Moiseienko, Enhancing mathematics education with GeoGebra and augmented reality, in: S. O. Semerikov, A. M. Striuk (Eds.), Proceedings of the 6th International Workshop on Augmented Reality in Education (AREdu 2023), Kryvyi Rih, Ukraine, May 17, 2023, volume 3844 of CEUR Workshop Proceedings, CEUR-WS.org, 2023, pp. 117–126.
URL: https://ceur-ws.org/Vol-3844/paper03.pdf. [154] D. A. Karnishyna, T. V. Selivanova, P. P. Nechypurenko, T. V. Starova, V. G. Stoliarenko, The use of augmented reality in chemistry lessons in the study of “Oxygen-containing organic compounds” using the mobile application Blippar, Journal of Physics: Conference Series 2288 (2022) 012018. doi:10.1088/1742-6596/2288/1/012018. [155] O. B. Petrovych, A. P. Vinnichuk, V. P. Krupka, I. A. Zelenenka, A. V. Voznyak, The usage of augmented reality technologies in professional training of future teachers of Ukrainian language and literature, in: S. H. Lytvynova, S. O. Semerikov (Eds.), Proceedings of the 4th International Workshop on Augmented Reality in Education (AREdu 2021), Kryvyi Rih, Ukraine, May 11, 2021, volume 2898 of CEUR Workshop Proceedings, CEUR-WS.org, 2021, pp. 315–333. URL: https: //ceur-ws.org/Vol-2898/paper17.pdf. [156] S. C. Olim, V. Nisi, Augmented Reality Towards Facilitating Abstract Concepts Learning, in: N. J.
Nunes, L. Ma, M. Wang, N. Correia, Z. Pan (Eds.), Entertainment Computing – ICEC 2020, Springer International Publishing, Cham, 2020, pp. 188–204. doi:10.1007/978-3-030-65736-9_17. [157] A. Dirin, A. Alamäki, J. Suomala, Gender Diferences in Perceptions of Conventional Video, Virtual Reality and Augmented Reality, International Journal of Interactive Mobile Technologies (iJIM) 13 (2019) pp. 93–103. doi:10.3991/ijim.v13i06.10487. [158] A. D. Samala, S. Rawas, S. Rahmadika, S. Criollo-C, R. Fikri, R. P. Sandra, Virtual reality in education: global trends, challenges, and impacts—game changer or passing trend?, Discover Education 4 (2025) 229. doi:10.1007/s44217-025-00650-z. [159] D. Baradari, N. Kosmyna, O. Petrov, R. Kaplun, P. Maes, NeuroChat: A Neuroadaptive AI Chatbot for Customizing Learning Experiences, in: Proceedings of the 7th ACM Conference on Conversational User Interfaces, CUI ’25, Association for Computing Machinery, New York, NY, USA, 2025. doi:10.1145/3719160.3736623. [160] F. Abuhashish, M. S. Bin Sunar, Advanced EEG-Based Neuroadaptive Systems for Gamified Cognitive Training and Therapy, in: 2025 1st International Conference on Computational Intelligence Approaches and Applications (ICCIAA), 2025, pp. 1–7. doi:10.1109/ICCIAA65327. 2025.11013080. [161] L. A. Bygrave, Machine Learning, Cognitive Sovereignty and Data Protection Rights with Respect to Automated Decisions, in: M. Ienca, O. Pollicino, L. Liguori, E. Stefanini, R. Andorno (Eds.), The Cambridge Handbook of Information Technology, Life Sciences and Human Rights, Cambridge Law Handbooks, Cambridge University Press, 2022, p. 166–188. [162] A. Cisse, Enhancing Hybrid Learning: The Role of Multimodal Adaptive Feedback in Human-AI Collaboration, in: M. Hlosta, I. Moser, A. Winer, N. Geri, B. Heinemann, S. Görzen, U. Ramnarain, C. van der Westhuizen, M. Penn, A. Gharahighehi, R. V. Schoors, P. Topali, J. Ooge, B. Flanagan, O. H. T. Lu, A. Shimada, N. Srivastava, A. C. M. Yang, H. Tseng, F. Okubo, E. D. Anaya, H. Ogata, R. Alfredo, S. B. Shum, M. Cukurova, Y. A. Dimitriadis, C. Lang, R. Martínez-Maldonado, P. Santos, O. Viberg, A. Nguyen, S. Järvelä, C. P. Rosé, A. Wise, M. Giannakos, K. Sharma, L. Huynh, A. Barany, R. S. Baker, J. Lin, A. Katz, M. B. Jr., X. Ochoa, D. J. Gagnon, L. Swanson, E. Harpstead (Eds.), Joint Proceedings of LAK 2025 Workshops co-located with 15th International Conference on Learning Analytics and Knowledge (LAK 2025), Dublin, Ireland, March 3-4, 2025, volume 3995 of CEUR Workshop Proceedings, CEUR-WS.org, 2025, pp. 135–139. URL: https://ceur-ws.org/ Vol-3995/HYBRID_paper1_short.pdf. [163] B. Sichterman, O. Noroozi, J. Boetje, S. van Ginkel, H. Khosravi, J. Versendaal, Supporting peer learning with artificial intelligence: A systematic literature review, Innovations in Education and Teaching International 62 (2025) 1648–1664. doi:10.1080/14703297.2025.2530118. [164] L. S. J. Farmer, Collective Intelligence in Online Education, in: J. Keengwe (Ed.), Handbook of Research on Pedagogical Models for Next-Generation Teaching and Learning, IGI Global Scientific Publishing, Hershey, PA, 2018, pp. 285–305. doi:10.4018/978-1-5225-3873-8.ch016. [183] P. MacDowell, K. Moskalyk, K. Korchinski, D. Morrison, Preparing Educators to Teach and Create With Generative Artificial Intelligence, Canadian Journal of Learning and Technology 50 (2024) 1–23. doi:10.21432/cjlt28606. [184] A. S. Al-Adwan, R. K. Meet, D. Kala, J. Smedley, M. Urbaníková, W. M. Al-Rahmi, Closing the divide: Exploring higher education teachers’ perspectives on educational technology, Information Development (2024). doi:10.1177/02666669241279181. [185] E. M. Denga, S. W. Denga, Revolutionizing Education: The Power of Technology, in: M. Fonkam, N. Vajjhala (Eds.), Revolutionizing Curricula Through Computational Thinking, Logic, and Problem Solving, IGI Global Scientific Publishing, 2024, pp. 167–188. doi: 10.4018/ 979-8-3693-1974-1.ch010. [186] T. Darvenkumar, S. Rajeshwari, E. Kiruthiga, W. C. Rajasekaran, Beyond the Screen: Understanding Technology as a Tool for Learning, Not a Teacher, Human Research in Rehabilitation 15 (2025) 371–385. doi:10.21554/hrr.092511. [187] M. A. Adarkwah, New Trends of Digital and Intelligent Technology, in: R. Huang, D. Liu, M. A.
Adarkwah, H. Wang, B. Shehata (Eds.), Envisioning the Future of Education Through Design, Springer Nature Singapore, Singapore, 2024, pp. 3–9. doi:10.1007/978-981-97-0076-9_1. [188] J. Fowlin, D. Coleman, S. Ryan, C. Gallo, E. Soares, N. Hazelton, Empowering Educators: Operationalizing Age-Old Learning Principles Using AI, Education Sciences 15 (2025) 393. doi:10.3390/educsci15030393. [189] UNESCO, Beijing Consensus on Artificial Intelligence and Education, UNESCO, Paris, 2019. URL: https://unesdoc.unesco.org/ark:/48223/pf0000368303. [190] IEEE, Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems, 2019. URL: https://standards.ieee.org/wp-content/uploads/import/ documents/other/ead_v2.pdf. [191] A. Ghimire, J. Edwards, From Guidelines to Governance: A Study of AI Policies in Education, in: A. M. Olney, I.-A. Chounta, Z. Liu, O. C. Santos, I. I. Bittencourt (Eds.), Artificial Intelligence in Education. Posters and Late Breaking Results, Workshops and Tutorials, Industry and Innovation Tracks, Practitioners, Doctoral Consortium and Blue Sky, Springer Nature Switzerland, Cham, 2024, pp. 299–307. doi:10.1007/978-3-031-64312-5_36. [192] E. G. Carayannis, D. F. J. Campbell, Democracy of Climate and Climate for Democracy: the Evolution of Quadruple and Quintuple Helix Innovation Systems, Journal of the Knowledge Economy 12 (2021) 2050–2082. doi:10.1007/s13132-021-00778-x. [193] D. Stark, D. Beunza, M. Girard, J. Lukács, The Sense of Dissonance: Accounts of Worth in Economic
Life, Princeton University Press, Princeton, NJ, 2009. URL: http://www.jstor.org/stable/j.ctt7pg4h. [194] A. J. Karran, P. Charland, J. Trempe-Martineau, A. Ortiz de Guinea Lopez de Arana, A.-M.
Lesage, S. Sénécal, P.-M. Léger, Multi-stakeholder perspective on responsible artificial intelligence and acceptability in education, npj Science of Learning 10 (2025) 44. doi:10.1038/ s41539-025-00333-2. [195] X. Tan, G. Cheng, M. H. Ling, Artificial intelligence in teaching and teacher professional development: A systematic review, Computers and Education: Artificial Intelligence 8 (2025) 100355. doi:10.1016/j.caeai.2024.100355. [196] T. Muir, Do We Still Need Teachers in the Age of AI?, 2025. URL: https://www.trevormuir.com/ blog/need-teachers. [197] K. Zhang, A. B. Aslan, AI technologies for education: Recent research & future directions, Computers and Education: Artificial Intelligence 2 (2021) 100025. doi: 10.1016/j.caeai.2021. 100025. [198] L. P. Prieto, O. Viberg, J. C. Yip, P. Topali, Aligning human values and educational technologies with value-sensitive design, British Journal of Educational Technology 56 (2025) 1299–1310. doi:10.1111/bjet.13602. [199] S. Wang, T. Xu, H. Li, C. Zhang, J. Liang, J. Tang, P. S. Yu, Q. Wen, Large Language Models for Education: A Survey and Outlook, CoRR abs/2403.18105 (2024). doi:10.48550/ARXIV.2403. 18105. arXiv:2403.18105.