www.ijcrsee.com

135

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

Review Article

Received: January 02, 2026.

Revised: March 29, 2026.

Accepted: April 23, 2026.

UDC:

004.738.5:339.138

37.015.3:159.953

10.23947/2334-8496-2026-14-1-135-144

Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Corresponding author: zspalevic@singidunum.ac.rs

Abstract: This paper examines how Artificial Intelligence in Education (AIED) is reshaping teaching and learning, drawing

on a systematic literature review alongside policy analysis to explore practical applications, the theories behind them, and their gov-

ernance consequences. Adopting the EU AI Act’s risk-based lens, we investigate the ways in which regulatory demands—ranging

from transparency and data stewardship to human oversight and provider accountability—influence how AIED tools are built and

taken up in practice. We group current uses into four areas—adaptive learning, intelligent assessment, learner profiling, and emerg-

ing tools—and read them through the prism of well-known learning theories such as constructivism. The analysis underscores that

while these technologies hold real promise, several prominent use cases—automated grading and learner profiling, for instance—

fall squarely within the EU AI Act’s higher-obligation categories, which means equity, explainability, and genuine human control

are not optional but essential for public trust. On the basis of these findings, we put forward concrete, compliance-oriented recom-

mendations aimed at helping educators, institutions, and policymakers deploy AI responsibly across varied educational settings.

Keywords: Artificial Intelligence, Education, Adaptive Learning, EU Artificial Intelligence Act, Transparency, Policy

and Governance.

Gabrijela Dimić

, Žaklina Spalević

, Milutin Nešić

, Ratko Ivković

, David Sotiroski

Čedomir Vasić

& Dragan Vučković

Academy of Technical and Art Applied Studies, Belgrade, Republic of Serbia,

e-mail: gabriela.dimic@viser.edu.rs, milutin.nesic@viser.edu.rs

Singidunum University Belgrade, Faculty of Tourism and Hospitality Management, Republic of Serbia,

e-mail: zspalevic@singidunum.ac.rs

MB University, Faculty of Business and Law, Belgrade, Serbia, email: rivkovic@mbuniverzitet.edu.rs, cedomir.vasic@ppf.edu.rs

Qinshift, Skopje, Republic of North Macedonia, e-mail: sotiroski_d@yahoo.com

Academy of Applied Studies Polytechnic, Beograd, Serbia, e-mail: dvuckovic@politehnika.edu.rs

Education in the Age of AI: Adaptive Systems, Assessment, and

Responsible Governance

Introduction

In the last ten years or so, advances in artificial intelligence (AI) have profoundly changed the

way educators and researchers think about what educational technology can achieve. Whether through

personalized tutoring platforms or systems that grade student work automatically, AI now sits at the heart

of education’s digital shift. Yet this growing ubiquity also brings thorny questions about ethical boundaries,

built-in biases, learner privacy, and fairness (U.S. Department of Education, 2023). The U.S. Department

of Education (2023) characterizes AI as a form of pattern-driven automation: machines go beyond mere

data collection to recognize regularities and act on them within educational settings.

When AI takes on tasks that were once the sole province of educators—deciding what a student

should study next, or flagging someone as at risk—it introduces responsibilities that schools have not

previously had to manage. What makes AI significant in education is not just what it can do technically, but

how it is beginning to reshape the very acts of learning, teaching, and assessment. “AI in education can

only grow at the speed of trust,” the report highlights (U.S. Department of Education, 2023).

For that reason, any deployment needs to be anchored in transparency and shared accountability,

www.ijcrsee.com

136

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

with teachers, students, families, and policymakers all having a seat at the table. We ground our discus-

sion in the European Union’s AI Act and its risk-based logic, which treats applications like automated

assessment and learner profiling as areas where stronger safeguards—greater transparency, tighter data

governance, and real human oversight—are non-negotiable.

Our review of the literature identifies four broad clusters of AI use in education: adaptive learning

and personalized tutoring; intelligent assessment and management; learner profiling and outcome predic-

tion; and novel AI-driven products (Wang et al., 2024).

The bulk of existing research concentrates on higher education, leaning on theoretical lenses like

constructivism, learning-styles frameworks, and cognitive models (Wang et al., 2024). Scholarly atten-

tion to AI in education surged after 2017, a trend closely tied to the rise of generative AI tools and the

rapid pivot to remote learning prompted by COVID-19. By 2022, researchers had scrutinized over 2,200

publications in an effort to chart the intellectual terrain of this field—work that reveals an increasingly mul-

tidisciplinary landscape bridging computer science, education studies, and information-systems research

(Wang et al., 2024). At the same time, the U.S. Department of Education (2023) flags several pressing

dangers: biased training data, erosion of student privacy, automation that outpaces human oversight, and

AI models built without adequate theoretical foundations (U.S. Department of Education, 2023).

If left unchecked, AI risks deepening the inequalities already present in education, which is precise-

ly why current guidelines stress keeping humans in the loop—making sure that people, not algorithms,

hold the final say in decisions that affect learners (U.S. Department of Education, 2023). Among the most

exciting possibilities is formative assessment: AI can deliver feedback as students work, giving them guid-

ance that would be impossible for a single teacher to provide at scale.

Automated essay scoring is a good case in point—it can lighten a teacher’s workload considerably,

yet if deployed carelessly it can also undermine the pedagogical goals it is supposed to serve. There is

also a growing consensus that AI should not merely be something students encounter passively; they

need to understand how it works, where it falls short, and how to engage with it critically (U.S. Department

of Education, 2023). This makes it essential that educational AI systems be transparent in their reasoning,

open to being overridden, and subject to audit—because a wrong algorithmic decision in a school setting

can leave marks that last well beyond a single semester (U.S. Department of Education, 2023).

Getting there will require researchers to build AI tools that are rooted in educational theory and

adapted to local contexts, while educators, policymakers, technology firms, and academics work together

on shared standards that keep AI use safe, fair, and genuinely helpful. Put simply, education needs its

own regulatory guardrails—ones that protect both students and teachers without stifling the innovation

that makes AI worth pursuing in the first place (U.S. Department of Education, 2023).

To situate this analysis within current regulatory developments, we briefly outline the EU Artificial Intel-

ligence Act (Regulation (EU) 2024/1689) (European Union, Regulation 1689/2024). The Act establishes a risk-

based framework, prohibiting certain practices—including the use of systems to infer emotions in educational

institutions—and classifying several educational uses as high-risk (e.g., AI for admissions/access decisions,

evaluation of learning outcomes, placement/streaming, and test proctoring/monitoring).

High-risk systems must implement risk management, data and data-governance measures, trans-

parency to deployers, human oversight, and accuracy/robustness/cybersecurity; the Act also introduces

transparency duties for AI interactions and synthetic content.

For AIED, this implies that learner profiling, automated assessment, and exam proctoring typically

fall under high-risk obligations and must be designed and deployed with compliance-by-design and mean-

ingful human control.

This paper integrates insights from a systematic review of over 2,200 peer-reviewed publications

on AI in education, drawing primarily from indexed databases (e.g., Scopus, Web of Science, ERIC).

Inclusion criteria focused on works published post-2017 addressing empirical applications, theoretical

models, and governance frameworks. Thematic coding enabled the extraction of four main domains of

application. The review is complemented by an interpretive policy analysis grounded in legal and ethical

instruments (e.g., EU AI Act, UNESCO Guidance).

Given the rapid pace of technological advancement and the growing adoption of AI tools in educational

contexts, this manuscript integrates theoretical foundations, current applications, and policy recommenda

tions aimed at ensuring the responsible, equitable, and pedagogically grounded use of AI in education.

By synthesizing findings from recent research and international policy guidelines, it offers a solid

www.ijcrsee.com

137

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

foundation for future decision-making in educational technology and serves as a resource for teachers,

researchers, and policymakers seeking to integrate AI meaningfully, ethically, and effectively into practice.

Foundations and conceptual framing of AI in education

AI is now woven into almost every corner of digital life, yet its importance for education runs deeper than

hardware and code—it raises fundamental questions about how we construct knowledge, how we ought to

teach, and how we can fairly judge what students have learned.

This section lays out the theoretical groundwork for AIED, sketches the field’s intellectual architecture,

and traces the main lines of thinking that inform how practitioners and scholars make sense of AI in schools

and universities today.

What Counts as AI in Educational Practice?

In the scholarly record, AI is typically portrayed as a broad family of computational systems that

exhibit capacities we associate with human cognition—learning from experience, reasoning through prob-

lems, perceiving their environment, and processing language. Alongside these conceptual descriptions sit

more operational definitions stressing that such systems pursue predefined objectives either through rule-

based logic or by learning patterns from data (U.S. Department of Education, 2023), (IEEE-USA Board of

Directors, 2017). For educators, these are not merely academic distinctions: AI now routinely steps into

roles that teachers used to fill on their own—selecting readings, adjusting the difficulty of a test, or even

assigning marks (Friedman et al., 2021).

Narrow vs. General Intelligence: Implications for Teaching and Learning

The AI tools that actually show up in classrooms today are overwhelmingly “narrow”—each one

built for a specific job, whether that means an expert system that scores quizzes, a neural network that

recommends reading material, or a learning-analytics agent that tracks how students are progressing

(Gartner, 2025). The idea of “general” AI—a system that could handle any intellectual task a human can—

opens up even weightier ethical territory for education, particularly around empathy and care. Today’s

AI cannot forge the emotional connections with students that lie at the heart of good, relationship-driven

teaching (The White House, 2022).

Models, Data, and Decisions: Why the Architecture Matters

At the technical core of every AI application is a model: a mathematical abstraction that takes in data

and generates outputs—a usage of the word “model” quite different from how educators speak of pedagogi-

cal or institutional models (Ivković, 2025). Large language models like GPT embody this principle, working by

predicting each successive word (or token) within enormous, high-dimensional parameter spaces. A key open

question for the field is whether these statistical constructs capture enough of what actually matters in real

classrooms. If a model’s training data skew toward certain populations or contexts, the resulting system can

discriminate in ways that hurt the very students it is meant to help. A robust framework for responsible AIED

therefore requires explicit understanding of model assumptions and limitations.

Learning Theories as Design Anchors for AIED

AI-powered learning environments tend to be most effective—and most defensible—when they are

rooted in well-tested learning theories. Constructivism is the framework cited most often, though research-

ers also draw on cognitive-load theory, item-response theory, and learning-styles accounts when designing

personalized content and assessments. Intelligent Tutoring Systems are a good illustration: they diagnose

individual learner needs and adjust pacing, feedback, and materials accordingly, going well beyond rote

content delivery to offer support that is genuinely informed by pedagogical theory.

Augmented Intelligence: Partnering Humans and Machines

Current thinking in the field does not cast AI as a replacement for teachers but rather as a collabo-

rator—a form of augmented intelligence that amplifies what human educators can do. In day-to-day terms,

www.ijcrsee.com

138

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

this looks like handing off repetitive chores—sending reminders, marking low-stakes quizzes—to AI, free-

ing teachers to spend more time on the creative, relational, and interpretive parts of their work. Critically,

the teacher must always be able to inspect, question, and overrule whatever the system produces. The

overarching idea is straightforward: AI ought to function as an ally in the learning process, not as an

opaque arbiter whose verdicts cannot be challenged.

Mapping aied in practice: A four-domain typology

AI has moved well past the stage of being an occasional classroom novelty; it is now built into

the routines and decision structures of education systems around the globe. Recent scholarship tends

to organize these applications into four clusters: adaptive learning and individualized tutoring; intelligent

assessment and classroom management; student profiling paired with predictive analytics; and newer

AI-driven products that push the boundaries of what educational technology can look like. This typology

is useful not only for comparing what each tool does, but also for thinking about the distinct pedagogical

possibilities and trade-offs each category brings.

Personalized Adaptation: Intelligent Tutors & Hypermedia

At its core, AI-driven tutoring is about meeting each student where they are—adjusting what they see,

how fast they move, and what feedback they receive. Intelligent Tutoring Systems (ITS) function much like

attentive mentors: they watch what a student does, infer what help is needed, and steer the learning path ac-

cordingly. ZOSMAT, a mathematics-focused system that pairs learning analytics with individualized navigation,

is one well-documented example. Common features of such systems include diagnostic quizzes that pinpoint

gaps, practice exercises calibrated to each learner’s level, and reading or activity suggestions that match

demonstrated competence. A complementary line of work, Adaptive Hypermedia Learning Systems (AHLS),

adapts presentation and sequencing to students’ learning preferences using methods like neural networks and

Bayesian classifiers—advancing the ideal of “education to the learner,” with self-paced progression matched

to ability and interest.

Assessment and Orchestration: AI-Enabled Evaluation & LMS

Providing personalized attention to every student is hard enough in a single classroom; at institutional

scale, it becomes all but impossible without automated support—hence the growing role of AI in assessment

and classroom management. Intelligent assessment platforms handle tasks like scoring, competency map-

ping, and ongoing formative tracking—MI Write, for instance, evaluates student essays—while specialized

pronunciation trainers offer detailed, immediate feedback on spoken language. Alongside these tools, learning

management systems enhanced with AI consolidate student data, take over repetitive administrative work,

and flag issues that instructors might otherwise miss. Typical capabilities comprise collaborative learning sup-

port, exam and evaluation workflows, and resource management with content recommendations; AI’s role

within these platforms is increasingly pivotal for observing progress and orchestrating timely interventions.

Learner Trajectories: Profiling & Predictive Modeling

A third major area of AI application brings together educational data mining and learning analytics to

build profiles of individual learners and forecast their academic trajectories. Algorithms—Naive Bayes classi-

fiers, neural networks, decision trees, support vector machines, among others—sift through behavioral data to

estimate which students are academically at risk and how they are likely to perform. In practice, these systems

generate early alerts about potential dropouts, predict grades and completion rates, model patterns of student

behavior, and help advisors guide course selection. These predictive pipelines are already used to anticipate

course selection or satisfaction, enabling administrators to align programs with learner needs and to advance

precision education through proactive support.

Novel Interfaces: Robots, XR, and Generative Systems

Some of the most visible innovation in AIED comes from robotics, extended reality (VR and AR),

and generative AI. Educational robots—many of them chatbot-driven—take on roles as varied as co-

www.ijcrsee.com

139

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

teacher, study buddy, emotional-support companion, and telepresence proxy for students who cannot be

physically present. Published examples range from toy robots that help children practice Chinese idioms

to conversational bots that facilitate book-club discussions. Virtual and augmented reality add another di-

mension, placing students inside immersive simulations—virtual labs, historical reconstructions—that can

boost both engagement and experiential understanding. Meanwhile, generative tools such as ChatGPT

open up possibilities for essay drafting, conversational tutoring, and the collaborative creation of learning

materials—pushing AIED beyond its traditional focus on content delivery and assessment feedback.

Table 1 maps the principal AIED application domains onto the risk categories set out in the EU

Artificial Intelligence Act (2024), offering a bird’s-eye view of where each type of tool sits in the regulatory

landscape. Because the impact on learners and educational outcomes differs substantially from one do-

main to another—adaptive learning environments carry different risks than, say, generative AI chatbots—

so too do the levels of regulatory scrutiny that apply. The high-risk tier is especially pertinent for automated

grading, student profiling, and decisions about who gets access to educational opportunities; here the

Act requires stringent safeguards—transparency, auditability, human oversight, and verified data quality.

Reading the table from left to right gives practitioners a way to weigh both the pedagogical promise and

the compliance burden of different AI deployments in their own institutional context.

Table 1. Mapping AIED Domains Against EU AI Act Risk Categories and Regulatory Obligations

AIED Application

Domain

Examples of Tools /

Functions

Theoretical

Foundation

Risk Classification (EU

AI Act)

Regulatory Obligations

1. Adaptive

Learning and

Personalized

Tutoring

Intelligent Tutoring

Systems (ITS),

Adaptive Hypermedia

Systems (e.g.,

ZOSMAT)

Constructivism,

Cognitive

Learning Theories

Limited / High

Risk (depending

on their impact on

learner evaluation or

progression)

Transparency to users-

Human oversight- Risk

assessment (if affecting

learner access or

progression)

2. Intelligent

Assessment and

Learning Manag.

Automated essay

scoring (e.g., MI Write),

AI- enhanced LMS

platforms, E-proctoring

tools

Assessment

theory,

Constructivist

Pedagogy

High Risk (especially

for automated grading,

test proctoring, and

evaluation systems)

Risk management

systems- High-quality

datasets- Meaningful

human control- Incident

reporting mechanisms

3. Learner

Profiling and

Predictive

Modeling

Dropout prediction,

performance

forecasting, behavioral

analytics

Learning

Analytics, Data-

Driven Instruction

High Risk (when

influencing educational

decisions through

profiling or prediction)

Bias testing and

documentation-

Auditability- Explainability

of decisions- Right to

human review

4. Innovative AI

Tools: Robotics,

XR, Gen. AI

ChatGPT, educational

robots, VR/AR for

immersive learning

Constructivism,

Experiential

Learning

Limited / Minimal Risk (if

not involved in decision-

making or profiling)

Disclosure of AI-generated

content- Notification of AI

interaction- Integration of

AI literacy in curricula

Governance and action agenda for trustworthy aied

As AI spreads through every level of education, the policy conversation has shifted from whether it

should be governed to how, precisely, to go about doing so. These are no longer speculative technologies;

they are tools that teachers and students encounter every day—automated grading, adaptive tutoring,

predictive analytics—and they carry genuine opportunities for better learning alongside real worries about

fairness, privacy, accountability, and the preservation of human agency. Building a coherent policy archi-

tecture that speaks directly to education’s needs is therefore urgent if we want AI to be deployed safely,

equitably, and to genuine effect.

Human-in-the-Loop and Pedagogical Sovereignty

The starting point for any sensible policy is that AI should strengthen human judgment, not supplant

it. Regulations should insist that AI systems used in schools be explainable, auditable, and capable of be-

ing overridden—so that teachers, students, and administrators can grasp what the system is recommend-

ing, check whether that recommendation makes sense, and set it aside when it does not. Preserving this

www.ijcrsee.com

140

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

kind of control is what keeps pedagogical sovereignty intact: educators must be free to step in whenever

an AI output clashes with what they know about their students or the realities of their classroom (U.S.

Department of Education, 2023).

Equity by Design: Bias, Fairness, Inclusion

Equity in digital education is not a nice-to-have; it is a baseline requirement. Models trained on data

that underrepresent certain groups risk replicating—and even amplifying—existing structural injustices.

Policy should therefore require concrete safeguards: routine checks on data quality, independent fairness

testing, and the active involvement of diverse stakeholders at every stage of design and evaluation, all

aimed at preventing AI from disproportionately harming the very students who most need support (The

White House, 2022).

Pedagogy-First Alignment of AI Tools

Educational AI ought to be grounded in pedagogical theory and designed to foster active, experience-

based, constructivist learning—not mere drill-and-recall. Close partnerships between tool developers, class

room teachers, and educational researchers are vital if the resulting products are to address genuine learn-

ing needs across different student populations. In other words, policy needs to keep AI firmly tethered to

well-founded pedagogical ideas and established learning science (European Commission, DG EAC, 2022).

Data Governance and Radical Transparency

Educational AI runs on vast quantities of sensitive information—behavioral traces, interaction logs,

grades, sometimes even socioeconomic background data—which means governance frameworks must

enshrine strong privacy safeguards (like FERPA compliance), along with anonymization protocols, strict

access controls, and disciplined rules for data sharing. Institutions must tell students and families plainly

what data are being gathered, why, and how those data feed into decisions—and obtaining informed

consent while shielding sensitive personal attributes is not optional (U.S. Department of Education, 2023).

Capacity Building and AI Literacy for Educators

Ironically, the people who deal with AI in classrooms most directly—teachers—are often the ones

with the least training and support for doing so effectively. Mandatory professional development and AI-

literacy initiatives should give teachers the skills to use these tools with confidence, read their outputs

critically, and step in when something goes wrong. Policymakers, for their part, need to back up those

programs with practical resources and everyday guidance on how to weave AI responsibly into regular

teaching (IEEE-USA Board of Directors, 2017).

Participatory Policy and Co-Design with Stakeholders

Governance becomes more credible when the people affected by it have a hand in shaping it.

That means bringing students, parents, teachers, school leaders, and technology companies together to

co-develop guidelines—ensuring that the tools deployed in schools genuinely reflect the educational and

human values of the communities they serve. When communities are involved, the result is more likely to

respect local languages, cultures, and priorities—which in turn makes the technology more relevant and

its governance more just.

National Frameworks and EU Guidance for Education

Individual institutional policies are not enough; countries need dedicated regulatory frameworks for

education that set minimum technical and ethical standards, require impact assessments across different

student groups, evaluate how AI affects both staff and learners, and mandate ongoing monitoring of live

systems. As a reference point, the European Commission’s educator-focused Ethical guidelines on the

use of AI and data in teaching and learning offer a useful model for sectoral governance (Maslej et al.,

2023). Legal scholarship further underscores alignment with constitutional guarantees and international

norms on digital rights and the right to education.

Weaving these normative commitments directly into the fabric of policy does more than ensure

www.ijcrsee.com

141

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

legal compliance—it lends the whole enterprise a measure of ethical credibility that purely technical stand-

ards cannot provide.

The EU AI Act (2024) and education

This chapter interprets the EU Artificial Intelligence Act (2024), arguing that the Act’s risk-based archi-

tecture, graduated obligations, and governance innovations provide both a regulatory compass and a practi-

cal playbook for educational actors deploying adaptive systems and assessment technologies at scale; yet,

translating that framework into classroom-ready practice will require sustained alignment with international

pedagogical guidance, robust cybersecurity preparedness, and institution-level capacity building.

At the core of the AI Act is a four-tier risk taxonomy—unacceptable, high, limited, and minimal/no

risk—that calibrates obligations according to anticipated harms and the salience of fundamental rights in

context; crucially for education, the Act explicitly positions certain uses (e.g., scoring of exams, systems

that shape access to education) within the high-risk category, thereby elevating educational assessment

and learner-profiling beyond mere ed-tech functionality to the domain of regulated socio-technical sys-

tems with mandatory risk management, high-quality datasets, traceability/logging, human oversight, and

demonstrable levels of robustness, cybersecurity, and accuracy (European Commission, 2025), (Euro-

pean Parliament, 2025).

In parallel, the Act enumerates a closed list of prohibited practices, including emotion recognition

in educational institutions, social scoring, and untargeted scraping for facial recognition databases, which

directly constrains a swath of speculative “affective computing” and invasive monitoring tools sometimes

marketed to schools; for educators and administrators, this prohibition line clarifies not only what cannot

be procured or piloted, but also how to frame vendor due diligence and contract clauses ex ante so that

“innovation” initiatives do not drift into rights-eroding experimentation (European Commission, 2025).

The Act’s transparency layer further touches classroom realities: interactions with AI must be dis-

closed, AI-generated content must be identifiable, and deepfakes or other synthetic media intended to

inform the public require clear labeling—obligations that intersect with academic integrity policies, media

literacy curricula, and institutional rules on the acceptable use of generative models in teaching and as-

sessment (European Commission, 2025), (European Parliament, 2025). For general-purpose AI (GPAI)

models—the engines increasingly powering adaptive learning platforms, automated feedback tools, and

content generators—the AI Act introduces targeted obligations (e.g., transparency on training data via

public summaries, copyright compliance, risk assessment and mitigation for models with systemic risk),

complemented by Commission guidance and a Code of Practice to smooth compliance pathways while

the regime matures; these GPAI rules matter for schools precisely because they externalize upstream

model accountability, making it more feasible for deployers (schools, universities, ministries) to rely on

contractually enforceable representations about model safety, security, and data provenance (European

Commission, 2025).

The timeline also conditions sectoral planning: the Act entered into force on 1 August 2024, with

prohibitions (and AI-literacy duties) applicable from 2 February 2025, GPAI obligations applicable from 2

August 2025, full applicability slated for 2 August 2026, and extended deadlines (e.g., certain high-risk

systems embedded in regulated products) out to 2 August 2027—sequencing that allows ministries and

institutions to schedule gap assessments, budget cycles, and teacher professional development in step

with legal milestones (European Commission, 2025), (European Parliament, 2025). Read normatively, the

AI Act reframes adaptive systems from a narrow narrative of hyper-personalization and efficiency into a

governed pedagogy of adaptation: personalisation remains legitimate when underwritten by explainability,

contestability, and documented human oversight; dataset quality is not a technical nicety but a legal duty;

and learning analytics pipelines become subject to post-market monitoring in which serious incidents

(e.g., systemic misgrading, biased placement recommendations) can trigger reporting and corrective ac-

tion (European Commission, 2025), (European Parliament, 2025).

For assessment, the implications are even more direct: automated scoring, proctoring, and predic-

tive risk flagging now sit within a compliance perimeter where model-risk analysis, bias testing, robustness

to adversarial inputs, and audit-ready documentation are not optional; in practice, this encourages as-

sessment designs that pair algorithmic scoring with calibrated human adjudication (“human-in-the-loop”),

www.ijcrsee.com

142

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

set clear escalation protocols for edge cases, and publish intelligible grading rationales to preserve due

process for students (European Commission, 2025), (European Parliament, 2025).

Because risk classification is the linchpin on which obligations hang, education policymakers can

profitably integrate the OECD Framework for the Classification of AI Systems into their institutional risk

triage: mapping an AI system along the framework’s dimensions—People & Planet, Economic Context,

Data & Input, AI Model, Task & Output—helps expose where a given adaptive tutor or proctoring tool

implicates fundamental rights (e.g., equality in access), what data flows create privacy and representa-

tiveness risks, which model characteristics stress explainability, and how task-output couplings might en-

able or foreclose pedagogically sound override mechanisms; the OECD framework’s lifecycle orientation

(planning/design; data collection; model building/validation; deployment and monitoring) dovetails with

the Act’s demands for pre-market conformity assessment and post-market vigilance, offering a shared

vocabulary by which school systems, vendors, and regulators can align on evidence of safety and ac-

countability (OECD, 2022).

In domains where generative AI becomes a co-author of learning artefacts (lesson plans, forma-

tive feedback, worked examples), UNESCO’s global Guidance for Generative AI in Education and Re-

search supplies a concrete programmatic complement to the Act’s legal minimalism: beyond compliance,

UNESCO calls for age-appropriate guardrails, teacher-led validation workflows, and explicit curricular

integration of AI literacy to sustain a human-centered, rights-preserving adoption; taken together, the

UNESCO guidance and the AI Act suggest that the legitimate pedagogical uses of GPAI are those that

preserve human agency, embed privacy-by-design, and make the limits of model knowledge visible to

learners (Miao and Holmes, 2023).

Notably, the AI Act’s emphasis on robustness, cybersecurity, and accuracy is not decorative: the

ENISA Threat Landscape underscores that ransomware, DDoS, phishing, and related attacks remain

persistent and adaptive, while AI-enabled threats (e.g., automated phishing, synthetic-voice fraud, data

exfiltration targeting model pipelines) complicate institutional risk profiles; in education settings—where

networks carry sensitive learner data and assessment systems may present tempting extortion targets—

the Act’s requirements for logging, incident reporting, and resilience should translate into concrete pro-

curement clauses (e.g., secure model-update channels, adversarial-robustness testing, rate-limiting and

anomaly detection) and operational controls (e.g., separation of duties, recovery plans, red-team exer-

cises) commensurate with the sector’s exposure (European Commission, 2025), (ENISA, 2024).

On governance, the Act’s distributed architecture—European AI Office, national market-surveil-

lance authorities, AI Board and advisory bodies—creates escalation and coordination pathways likely to

affect education in three ways: first, by generating codes of practice and interpretive guidance that lower

transaction costs for small ministries and school networks; second, by standardizing registries and docu-

mentation templates that enable comparability of high-risk deployments (e.g., exam-scoring engines); and

third, by anchoring enforcement in administrative routines that encourage continuous improvement (e.g.,

corrective action plans, proportional penalties) rather than one-off punitive gestures (European Commis-

sion, 2025), (European Parliament, 2025). Strategically, aligning adaptive learning with the Act means

institutionalizing a pedagogy-first compliance culture:

• needs analysis and theory-of-change before tool selection;

• explainability criteria tied to specific learner decisions (placement, feedback, progression);

• calibration studies that check algorithmic recommendations against expert judgments across student

subgroups;

• a duty to furnish students with recourse, including the right to a human review of impactful automated

decisions (European Commission, 2025), (OECD, 2022).

For assessment, the combination of transparency duties and high-risk obligations invites renewed

attention to validity and fairness: if a proctoring classifier generates false positives differentially by lighting,

skin tone, or disability status, then both the Act’s dataset-quality requirement and equity commitments in

education policy are implicated, compelling iterative retraining, alternative accommodations, or—where

irreparable—retirement of the system; conversely, where automated scoring aids formative feedback

without deciding high-stakes outcomes, limited-risk transparency may suffice, provided institutions clearly

signal to learners when AI is involved and how its suggestions are moderated by instructors (European

Commission, 2025), (European Parliament, 2025), (Miao and Holmes, 2023).

www.ijcrsee.com

143

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

Finally, the chapter contends that responsible governance post-Act is a project of institutional ca-

pability as much as legal conformity: ministries and universities should inventory AI-mediated decisions,

classify them using OECD criteria, decide—policy-by-policy—where high-risk thresholds are crossed, and

build UNESCO-aligned AI literacy into teacher education and student orientation; procurement should

prefer vendors who evidence adherence to GPAI guidance, publish training-data summaries, and sup-

port auditability; and CISOs should plan for ENISA-profiled threats with controls that map to the Act’s

robustness and post-market monitoring obligations; done well, the result is not compliance theatre but a

defensible, learner-first operationalization of the Act’s animating principle: trustworthy AI in education that

enhances learning while protecting rights (European Commission, 2025), (ENISA, 2024).

While this paper outlines foundational principles for responsible AIED, further empirical research is

needed to evaluate real-world impacts across diverse educational systems. Future work should investi-

gate longitudinal learning outcomes, the effectiveness of human-in-the-loop models, and equity effects of

algorithmic interventions. Policymakers should prioritize capacity-building and iterative evaluation as core

components of AI adoption in education.

Conclusion

Education is changing rapidly, and AI is one of its most powerful catalysts—but it is far from a

neutral instrument, and deploying it responsibly demands sustained attention to ethical, pedagogical, and

legal standards. This paper has assembled a framework that weaves together theory, real-world applica-

tions, and regulatory alignment, foregrounding the roles of human agency, sound data governance, and

transparency at every turn. With the EU AI Act now redefining the ground rules for how adaptive systems

and automated assessments can be designed and deployed, educators and policymakers face both an

obligation and a rare opportunity to shape a pedagogy that embraces innovation without abandoning its

responsibilities. Going forward, what the field needs most is empirical work—studies that evaluate these

systems across different institutional contexts, with a persistent focus on whether they actually promote

equity and whether the institutions using them are ready for the challenge.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial,

or not-for-profit sectors.

Conflict of interests

The authors declare no conflict of interest.

Author Contributions

Conceptualization, G.D., Z. S and D.V; methodology, Č.V.; software, R.I., D.S. and M.N.; formal

analysis, R.I. and Č.V.; writing—original draft preparation, G.D, Z.S. and R.I.; writing—review and editing,

D.S., Č.V. and D.V. All authors have read and agreed to the published version of the manuscript.

References

ENISA. (2024, October). Threat landscape. European Union Agency for Cybersecurity. https://www.enisa.europa.eu/topics/

cyber-threats/threat-landscape

European Commission. (2025, August 1). AI Act | Shaping Europe’s digital future. https://digital-strategy.ec.europa.eu/en/poli-

cies/regulatory-framework-ai

European Commission, Directorate-General for Education, Youth, Sport and Culture. (2022). Ethical guidelines on the use of

articial intelligence (AI) and data in teaching and learning for educators. Publications Ofce of the European Union.

https://data.europa.eu/doi/10.2766/153756

European Parliament. (2025, February 19). EU AI Act: First regulation on articial intelligence. https://www.europarl.europa.eu/

topics/en/article/20230601STO93804/eu-ai-act-rst-regulation-on-articial-intelligence

www.ijcrsee.com

144

Dimić, G. et al. (2026). Education in the Age of AI: Adaptive Systems, Assessment, and Responsible Governance, International

Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 14(1), 135-144.

European Union. (2024, June 13). Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down

harmonised rules on articial intelligence (Articial Intelligence Act) and amending Regulations (EC) No 300/2008,

(EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/

EU, (EU) 2016/797 and (EU) 2020/1828. Ofcial Journal of the European Union, L (12 July 2024), 1–144. http://data.

europa.eu/eli/reg/2024/1689/oj

Friedman, L., Black, N. B., Walker, E., & Roschelle, J. (2021, November 8). Safe AI in education needs you. Communications

of the ACM Blog. https://cacm.acm.org/blogs/blog-cacm/256657-safe-ai-in-education-needs-you/fulltext

Gartner. (2025). Augmented intelligence. Gartner Glossary. https://www.gartner.com/en/information-technology/glossary/ aug-

mented-intelligence

Ivković, R. (2025). Transformation and economic aspects of software engineering through the implementation of the EU AI Act.

Pravo – Teorija i Praksa, 42(1), 186–200. https://doi.org/10.5937/ptp2501186I

Maslej, N., et al. (2023). AI Index 2023 annual report. Stanford University, Institute for Human-Centered AI (HAI). https://aiin-

dex.stanford.edu/report/

Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://www.unesco.org/en/

articles/guidance-generative-ai-education-and-research

OECD. (2022). OECD framework for the classication of AI systems. OECD Publishing. https://www.oecd.org/content/dam/

oecd/en/publications/reports/2022/02/oecd-framework-for-the-classication-of-ai-systems_336a8b57/cb6d9eca-en.pdf

The White House, Ofce of Science and Technology Policy. (2022, October). Blueprint for an AI Bill of Rights: Making auto-

mated systems work for the American people. https://www.whitehouse.gov/ostp/ai-bill-of-rights/

U.S. Department of Education, Ofce of Educational Technology. (2022). Advancing digital equity for all: Community-based

recommendations. https://les.eric.ed.gov/fulltext/ED624655.pdf

U.S. Department of Education, Ofce of Educational Technology. (2023). Articial intelligence and the future of teaching and

learning: Insights and recommendations. https://tech.ed.gov

Wang, S., Wang, F., Zhu, Z., Wang, J., Tran, T., & Du, Z. (2024). Articial intelligence in education: A systematic literature

review. Expert Systems with Applications, 252, 124167. https://doi.org/10.1016/j.eswa.2024.124167

IEEE-USA Board of Directors. (2017, February 10). Articial intelligence research, development and regulation [Position state-

ment]. IEEE-USA Global Public Policy. http://globalpolicy.ieee.org/wp-content/uploads/2017/10/IEEE17003.pdf