pexels/kindelmedia When my eight-year-old daughter asked me who Nikos Kazantzakis was, I did not recite his biography. I opened ChatGPT, and she watched as I typed. The answer that came back was fluent, polite, and deeply misleading. Kazantzakis, it claimed, was a “20th-century novelist known for his romantic tales and political thrillers,” yet the anguish, the spiritual depth, and the existential struggle that shaped his voice were gone. My daughter blinked and said, “That doesn’t sound very interesting.” This moment did not make me fear artificial intelligence; it reminded me what real education demands: mediation, framing, and human presence. Thus, we rewrote the answer together, keeping the parts she understood and rephrasing the rest; we asked the AI again and compared; we drew his village, listened to Cretan music, and imagined what a conversation with Alexis Zorbas might feel like. Then, we wrote our own simplified paragraph. The process took 40 minutes and was the most meaningful lesson of our week.
Generative AI is not an oracle or a tutor; it is a mirror and a provocateur.
In what follows, I bring together four complementary perspectives on generative AI in education. I ground the argument in everyday family and classroom experiences that illustrate how children already encounter AI in their learning lives. I also examine technical and architectural choices, such as educator-governed interaction layers, which shape how model outputs are delivered to learners. In addition, I connect these choices to pedagogical frameworks of scaffolding, mediation, and teacher–student coregulation, and I situate the discussion within emerging regulatory and policy developments, including the EU AI Act and international guidance from UNESCO and the OECD. Together, these perspectives help clarify not only what AI says, but how we choose to frame what children hear.
For the classroom, I have seen the same vision. Generative AI is not an oracle or a tutor; it is a mirror and a provocateur. In a recent project [1], we explored how large language models could be used to simplify canonical literary texts for younger readers, such as passages from The Tempest and Macbeth. The system generated age-aligned versions, adapting syntax, vocabulary, and emotional tone while preserving semantic depth. These outputs were not final answers but starting points for reflection. Students and educators examined parallel versions of the texts, questioned lexical choices, and reflected on tone, affect, and audience. The goal was not to let AI decide but to support interpretation through guided comparison and discussion. With appropriate scaffolding, learners engaged critically, as editors, interpreters, and coconstructors of meaning. Our emphasis is human-centered: teacher judgment, dialog, and community norms frame the work; technology is a resource shaped by those relationships, not an end in itself.
In addition to principle-level guidance, UNESCO has also published AI competency frameworks for students and teachers [3], [4], which we include here to connect public governance directly to classroom-facing competencies. We treat these frameworks as competency targets; the concrete interface and governance mechanisms described below are our proposed operationalization to support those targets in classroom settings.
AI can dazzle without informing, affirm without explaining, and simplify without understanding.
Generative tools can support rich learning when carefully scaffolded; without such framing, the picture changes. A recent meta-analysis by Wang and Fan [7] reports gains in learning, perception, and higher-order thinking but only under appropriate scaffolds. Without scaffolding, children are exposed to systemic risks: hallucinated facts, incoherent reasoning, and the seductive confidence of statistically plausible language. AI can dazzle without informing, affirm without explaining, and simplify without understanding. The problem is not that AI makes mistakes; the problem is that its voice sounds right, especially to a child. When children receive unframed responses, they may mistake fluency for accuracy or correlation for meaning. The Al’s voice becomes dominant, yet although some systems now attach citations or generate step-by-step explanations, recent evidence indicates that these self-explanations are often not faithful to the model’s actual decision process and may mislead nonexperts [8], so it cannot reliably explain its reasoning or justify its claims without educator mediation.
This is why no child should ever face generative AI alone: the solution is not prohibition, but thoughtful design. We need transparent intermediary layers, software governed by educators, ministries, and communities, which filter and contextualize AI responses for children. This aligns with growing calls for public oversight of AI tools in education, such as those embedded in the EU AI Act’s classification of educational systems as high risk. These interfaces should align outputs with ethical guidelines (e.g., UNESCO’s 2021 Recommendation on the Ethics of Artificial Intelligence [2], UNESCO’s AI competency frameworks for students and teachers [3], [4], the OECD AI Principles [5], and the OECD and European Commission’s AI Literacy Framework [6]), support curricular goals, and scaffold interpretation. This emphasis on pedagogical framing is consistent with our related argument that Al’s educational value depends on ethical framing and educator-led embedding, rather than on technical affordances alone [9].
In Table 1, the mappings indicate how an educator-governed interaction layer can help operationalize these public frameworks in practice, rather than implying that the frameworks prescribe specific technical features. Their purpose is not to suppress AI tools but to reframe them in ways that are developmentally, pedagogically, and culturally coherent. This is a call for public institutions to establish design principles before private tools dominate the educational space by default. To make this explicit, we locate key decisions in public institutions: curriculum authorities, school boards, and parent councils set defaults, approve prompt libraries, review audit logs, and publish clear opt-out routes, while professional associations and civil society monitor impacts on equity, privacy, and labor. More broadly, UNESCO’s 2025 capacity-building work on AI supervisory authorities underscores that effective governance depends on practical institutional routines and supervisory capacity that translate high-level norms into operational practice [10]. Concretely, under regulation (EU) 2024/1689, AI that is used to determine access or admission, evaluate learning outcomes, assign individuals to levels, or monitor exams is classified as high risk (Annex III, 3).
A promising example solution to these issues is an educator-governed validation layer (e.g., ethical-pedagogical validation layer (EPVL) [1]). This middleware sits between the model and the learner and introduces a structured space for ethical and pedagogical judgment. Rather than offering raw outputs, the layer produces three views: a simplified student-facing text, pedagogical commentary for educators, and reflective prompts to stimulate critical engagement. As illustrated in Figure 1, student queries pass through a school-governed interaction layer that applies public guidelines and teacher review before any model output reaches the classroom, while all interactions are logged for accountability.
Figure 1.Educator-governed interaction layer in a pedagogically controlled environment. Student queries are mediated by teacher review and policy guardrails that implement public guidelines (UNESCO, OECD, Ministry, and so on), with provenance and logging, before any response from the commercial model is shown.
From a regulatory and policy perspective, this interaction-layer pattern supports alignment with key EU AI Act requirements by embedding human oversight before disclosure (Article 14), providing institu-tion-facing instructions and disclosures (Article 13), enabling lifecycle logging for auditability (Article 12), and scaffolding risk management and data governance processes (Articles 9 and 10) [11], [12].
From a technical and architectural perspective, the layer acts as adaptable middleware between the model and the learner. It should be deployed as institution-controlled middleware with on-premises or regionally hosted inference endpoints behind an auditable API gateway that enforces predisclosure validation, policy-as-code guardrails, and data residency constraints. Core functions should run as containerized microservices with role-based access control, versioned artifact storage, immutable and queryable logs, and encryption in transit and at rest to ensure provenance and full classroom oversight. Integration should follow model-agnostic APIs, so schools can switch between commercial services and locally hosted open models without modifying the pedagogical layer while defaulting to strict data minimization, zero data retention unless opted in, and end-to-end logging.
From a pedagogical and learning sciences perspective, its true value is that it delays the presentation of AI-generated responses until a human educator can first review, contextualize, or reshape them. This intentional pause enables teachers to scaffold student reasoning before the model’s answer appears, drawing on principles from Vygotsky’s Zone of Proximal Development [13] and Scho¨ n’s theory of reflective practice [14]. Accordingly, educator agency and classroom context determine when and how the system is used; the layer simply structures the interaction so that mediation remains in human hands. In this article, we use scaffolding to mean contingent, adaptive support that fades as learner competence grows (aligned with the ZPD), mediation to mean the cultural and semiotic tools through which meaning is constructed (e.g., teacher prompts, rubrics, and AI-generated drafts), and an additional layer to denote the infrastructural interaction layer (this edu-cator-governed layer, e.g., EPVL) that operationalizes mediation and enables teacher-led scaffolding at the point of use [13], [15], [16], [17]. We make the target of scaffolding explicit: (a) teacher-facing scaffolding comprises pedagogical commentary, risk flags, and suggested prompts that support planning and in situ mediation; (b) student-facing scaffolding comprises reflective questions, partial exemplars, and age-tiered explanations designed to fade as competence increases; and (c) joint scaffolding combines both through coregulation cycles in which teachers set constraints and timing, while the system surfaces cues and students act on them [17], [18]. We distinguish scaffolding from tool-centric assistance that accelerates production without fading support (e.g., coding copilots that tend to supply solutions by default); in our approach, support aims at independence rather than offloading. This design aligns with the augmentation perspective and with hybrid human–AI learning technologies that articulate teacher–AI coregulation, detect–diag-nose–act cycles, and levels of automation [19], [18]. To address specific weak points in child–AI interactions, the interaction layer implements or supports predisclosure human review to prevent premature anchoring, curriculum tagging and age-tiered simplification to align outputs with learning goals, source disclosure and versioned logging to counter hallucinations, reflective prompts to sustain metacognitive questioning, and accessibility options for multilingual and diverse learners.
From an everyday classroom perspective, in pilots with learners aged 10–12 years [1], students began to treat AI as a draft generator rather than an authoritative source, asking, “Why did it say this?” instead of simply copying the response. This shift from passive use to dialogical engagement reflects the power of interpretive mediation, and the architecture of such a layer enables real-time educator feedback that supports continuous system refinement and alignment with national standards.
This layer is not a product; it is a governance pattern in technical form, a replicable blueprint for jurisdictions seeking to retain pedagogical sovereignty in ecosystems dominated by commercial AI interfaces. It empowers public institutions to define what “responsible AI in education” looks like in practice, by embedding ethical, cultural, and developmental priorities directly into the interaction layer.
We use EPVL as a reference pattern of an educa-tor-governed interaction layer. See Table 1 for how public frameworks map to requirements that such a layer can operationalize. Concretely, the educa-tor-governed interaction layer turns public guidance into operations: predisclosure teacher review, rolebased permissions, and immutable audit logs make transparency, accountability, and human oversight routine classroom practices, as required by UNESCO’s Recommendation on the Ethics of AI and consistent with the OECD AI Principles. Provenance indicators, versioning, age-tiered defaults, and refusal routes uphold child protection and responsible disclosure, as emphasized by UNESCO. For primary and secondary AI literacy, the layer supports Engage and Manage through configurable gating and logging; it enables Create and Design through staged disclosure (prompts and cues first, exemplars next, and full solutions only when pedagogically justified) and sandboxed exploration, meaning a school-controlled safe environment where students can experiment with prompts and compare outputs under teacher oversight and logging, which keeps authority with teachers, aligning with the OECD and European Commission AI Literacy Framework.
The question is no longer whether children will hear from AI because they already do; the real question is whether what they hear will be framed with care, context, and human intention, or left to the indifferent fluency of an algorithm optimized for prediction, not truth.
Table 2 summarizes how the educator-governed interaction layer operationalizes key EU AI Act provisions. Concretely, it implements versioned, immutable logs that capture prompts, model identifiers, outputs, educator actions, timestamps, and policy decisions to satisfy record-keeping under Article 12; provides inline disclosures of intended purpose, capability limits, provenance, and age-tiered defaults to meet transparency under Article 13, with user-facing indicators when interacting with AI where applicable (Article 50); and enforces predisclosure holds, teacher-approval gates, pause/override controls, and escalation routes to deliver human oversight under Article 14. For Annex III, 3 high-risk educational tasks, role-based access, curriculum tagging, default-on logging, auditable workflows, and clear routes for concerns and redress keep schools and public authorities in effective control. The layer also supports Articles 9 and 10 through policy-as-code guardrails, data minimization, retention schedules, and provenance metadata and assembles evidence packs (usage analytics, logs, policy configurations, and DPIA links) to facilitate the fundamental-rights impact assessment under Article 27. In short, it lets schools apply the act as routine classroom practice rather than ad hoc compliance.
In Estonia, where AI tutors are integrated into national systems, such scaffolding is already visible, while, in Cyprus and other Mediterranean countries, local initiatives are beginning to emerge. However, these efforts are scattered. The private platforms move faster and reach our homes before our ministries do. Thus, the question is no longer whether children will hear from AI because they already do; the real question is whether what they hear will be framed with care, context, and human intention, or left to the indifferent fluency of an algorithm optimized for prediction, not truth. Kazantzakis once wrote, “You have your brush, you have your colors, you paint the paradise, then in you go.” Let us ensure that the brushes we give our children are governed by pedagogical values, framed by reflective practice, and informed by human understanding. Let us not offer them just fluency, but wisdom. Not just power, but trustworthiness. Not just mirrors, but companions for thought.
This requires more than ethical aspirations; it demands concrete, structural safeguards. We need certified educational interfaces that foreground pedagogical purpose over performance, and we need public prompt libraries that are transparently designed, aligned with curricular goals, and accessible to all educators. We also need deliberate response delays that open space for teacher-led scaffolding before any AI answer appears, as well as reflective AI kits that let students explore variations, revise suggestions, and interrogate alternatives. Even where vendors advertise “educational” or “tutor” modes that support iterative, critical questioning, these remain optional interface features rather than enforceable, educator-governed controls; our proposal specifies institutionally set delays, age-tiered defaults, and auditable mediation at the interaction layer. Operationally, staged disclosure is used, prompts and cues first, exemplars next, and full solutions only when pedagogically justified, with control shifting to learners over time and dashboards supporting staged transfer of control; the process is logged so that teachers can withdraw supports as proficiency grows [18].
In parallel, educators need protected time and funded professional development to design prompts, curate examples, and conduct classroom-level reviews, together with school policies that recognize refusal rights and protect teacher autonomy.
In short, specifically designed technology is necessary but not sufficient: educators lead classroom use, school leaders and parent councils provide oversight, public authorities set defaults and standards, and vendors adapt to these constraints.
Finally, we need audit trails for educational AI outputs, including versioning, source attribution, and disclosure of model provenance. These logs directly address the Act’s record-keeping obligation for high-risk systems and facilitate the institutions’ and public authorities’ fundamental-rights impact assessment (Articles 12 and 27).
Technically, these safeguards can be modular and scalable, whereas, pedagogically, they are indispensable. Systems that adopt an educator-governed interaction layer, with delayed disclosure and auditable mediation, offer a promising blueprint: less a final answer than a direction that keeps interpretive sovereignty with learners and teachers.■
Author Information
Savvas A. Chatzichristofis is a professor of artificial intelligence and the vice-rector of research and innovation at Neapolis University Pafos, 8015 Paphos, Cyprus. His research lies at the intersection of AI, computer vision, and robotics, with a particular interest in building systems and practices that support responsible, human-centred use of AI in education and society. He is a member of Cyprus’s National AI Taskforce and contributes to national strategy work, while his outreach and open learning initiatives have been recognised with national distinctions in Cyprus and EU-level recognition through the All Digital Awards. Email: s.chatzichristofis@nup.ac.cy.
_______
To read the full version of this article, including references, click HERE.
________
JOIN SSIT