Skip to main content

Children and Adolescents in the Age of Generative AI: 3 AI in the school context

Children and Adolescents in the Age of Generative AI
3 AI in the school context
  • Show the following:

    Annotations
    Resources
  • Adjust appearance:

    Font
    Font style
    Color Scheme
    Light
    Dark
    Annotation contrast
    Low
    High
    Margins
  • Search within:
    • My Notes + Comments
    • Notifications
    • Privacy
  • Publication HomeChildren and Adolescents in the Age of Generative AI
  • Publications
  • Learn more about Manifold

Notes

table of contents
  1. Cover image
  2. Title Page
  3. Imprint
  4. Logos
  5. Contents
  6. Figures
  7. Tables
  8. Executive Summary
  9. 1 Introduction
  10. 2 AI as disruptive technology
  11. 3 AI in the school context
  12. 4 The situation in Luxembourg
  13. 5 Reactions and prevention
  14. References
  15. Glossary of Key Terms

3 AI in the school context

AI is reshaping schooling at a systemic level. It analyses data, automates processes, and generates content, thereby affecting diagnosis, lesson design, assessment, and learning organisation. The OECD’s Digital Education Outlook 2023 frames AI as one component of a wider digital education ecosystem that requires trustworthy data and infrastructure, interoperability across platforms (e.g., learning management systems, student information systems, data repositories), and clear governance covering privacy, transparency, and accountability (OECD, 2023).

Beyond tools, the central challenge is cultivating a reflexive learning culture in which responsibility, transparency, and human control guide use, so that new pedagogical opportunities and ethical/didactic risks are both recognised and managed (Fu et al., 2024). This perspective aligns with theories that frame education within an intensifying “digital condition” marked by communality, referentiality, and algorithmicity (Stalder, 2016), and with the account of deep mediatization, i.e., the deep embedding of media infrastructures in social practices and institutions (Hepp, 2020).

Empirical developments are already visible. The National Youth Report in Luxembourg indicates that AI is changing how knowledge is produced, taught, and assessed, and that teacher roles and student–teacher relations are shifting (Biewers Grimm et al., 2025). In classrooms, AI-driven platforms can analyse individual progress and provide adaptive tasks, while analytic tools return feedback on lesson trajectories. A large evaluation by the Education Endowment Foundation found that teachers using ChatGPT with a structured guide reduced lesson-planning time by about 31% without observed reductions in the quality of materials (EEF, 2024). Similar reviews report that LLM-supported tools can generate high-quality tasks and feedback, provided that pedagogical oversight and subject-matter checking remain in place (Alfarwan, 2025; Zhao, 2025). A follow-on randomised controlled trial with Oak National Academy’s Aila planner is now recruiting subjects to test workload and quality at scale (National Foundation for Educational Research, 2025).

Early-warning systems illustrate both promise and caution. Using large longitudinal datasets, machine-learning models can identify learning risks early. For example, one study predicted upper-secondary dropout as early as the end of primary school with moderate accuracy (Psyridou et al., 2024). These datafied forms of steering show how far AI is reaching into pedagogical processes: data are becoming not just the centre of enquiry, but also instruments for shaping teaching and learning — what scholars term algorithmic governance in education. There is an opportunity for personalised support and efficiency; however, the trade-off is that teachers will increasingly interact with decision systems whose criteria or weightings may not be fully transparent (OECD, 2023; Williamson, 2017; Williamson & Eynon, 2020).

3.1 Pedagogical potentials and risks

A central promise of AI in schools is the personalisation and individualisation of learning. The evidence that has accumulated for intelligent tutoring systems indicates sizeable average learning gains in controlled evaluations. Effectiveness depends on subject, alignment to assessments, and implementation fidelity (Kulik & Fletcher, 2016; Ma et al., 2014; Steenbergen-Hu & Cooper, 2013). More recently, a randomised controlled trial in introductory college physics found that a custom GPT-4 tutor designed with research-based practices produced significantly larger learning gains in less time than in-class active learning, while increasing engagement and motivation (Kestin et al., 2025). These results are promising but context-specific; they underscore that design and teacher orchestration matter (Pane et al., 2015; Pane et al., 2017).

Complementing this body of evidence, a recent meta-analysis of 51 experimental and quasi-experimental studies provides robust evidence for the pedagogical potential of generative AI in learning contexts. Across the included studies, ChatGPT demonstrated a large positive effect on students’ learning performance and moderate positive effects on learning perception and higher-order thinking (Wang & Fan, 2025). The analysis found that impact was strongest in problem-based learning models and skills- and competence-oriented courses, particularly when interventions lasted between four and eight weeks. These findings indicate that AI-supported tutoring systems can meaningfully enhance conceptual understanding and engagement when pedagogically aligned and scaffolded, though their effectiveness varies depending on context, duration, and instructional design.

Beyond direct learning effects, AI can also reduce administrative and repetitive workload (e.g., auto-grading, item analysis, differentiation suggestions), freeing up time for individual guidance. Reviews and community position papers stress that impact is greatest where AI is transparent and supports — rather than substitutes—professional judgment (Holmes et al., 2022). There are inclusive benefits: text simplification, translation, text-to-speech, and captioning can lower access barriers for learners with linguistic or visual needs. International guidance recommends that AI tools in education be rights-based, privacy-protecting, and accessible by design, with human oversight and sustained teacher capacity-building (UNESCO, 2023). In all cases, effective use depends on reflective implementation that aligns technological affordances with didactic intentions.

With these potentials come non-trivial risks. Systematic reviews show that algorithmic bias in educational AI can misclassify or mispredict performance — especially for socio-economically disadvantaged or linguistically/culturally diverse students — which leads to unequal support (Baker & Hawn, 2022). Within learning analytics, recent studies probe bias and fairness in widely used models such as Bayesian Knowledge Tracing: aggregate performance can appear unbiased while particular skills or subgroups exhibit systematic disparities, underscoring the need for subgroup audits (Stinar et al., 2025; Zambrano et al., 2024). UNESCO and the OECD have called for transparency, auditing, and impact assessments that surface model assumptions and limits and for governance that separates pedagogical functions from commercial data extraction (OECD, 2023; UNESCO, 2023; Williamson & Eynon, 2020). Schools should require vendors to provide model cards and impact statements documenting training-data provenance, subgroup performance, and known limitations, and to expose independent audit hooks for periodic review, so that AI remains pedagogically governed rather than platform-driven (UNESCO, 2023).

3.2 Teacher professionalisation and school capability

Successful integration depends on teacher expertise and school capacity. Evidence from the OECD indicates that acceptance and effectiveness rise when professional development, infrastructure, and collegial support are in place; lack of training yields uncertainty and extra workload. The OECD Outlook highlights teachers’ AI literacy to understand techniques, critically assess model outputs, and use AI creatively in instruction (OECD, 2023). These needs are mirrored in national youth reporting for Luxembourg, which highlights teachers’ technological, ethical, and didactic competence requirements for meaningful AI use.

3.2.1 Student AI literacy and cognitive load

Students likewise need a baseline understanding of what AI is, how it works, and its limits. Work on AI literacy proposes competency frameworks (data, models, training, ethics) that do not require coding and shows that targeted programmes can improve knowledge, critical thinking, and ethical reflection (Long & Magerko, 2020). From a learning-science perspective, Cognitive Load Theory cautions that generative systems can produce voluminous but not instructionally sequenced content; unfiltered outputs may overwhelm younger learners. Design tactics to manage intrinsic and extraneous load (worked examples, segmentation, modality, variability) remain essential (Paas et al., 2020; Sweller et al., 2019). The implication is that outputs from systems like chatbots should be pedagogically filtered, contextualised, and embedded in existing learning sequences to avoid shallow or merely reproductive learning.

3.2.2 Age-appropriate didactic integration

Implementation should be age-graded.

  • Lower secondary (approx. ages 10–15): spaced vocabulary practice, reading fluency, formative checks, and learning-status diagnostics — paired with clear goals, intuitive interfaces, and immediate teacher-mediated feedback.
  • Upper secondary (approx. ages 15–19): writing support, simulations, and data-driven reflection on study habits — under the condition that students understand system limits (bias, hallucinations) and can justify their use. OECD recommends coupling adoption with evaluation, governance, and instructional design support (OECD, 2023).

This is also in line with the current recommendations outlined in Luxembourg’s KI-Kompass (SCRIPT, 2025).

3.2.3 Supporting schools in implementing safe and effective AI use

For schools to implement AI in a safe, equitable and pedagogically meaningful way, targeted support and awareness activities are required. BEE SECURE can play a key intermediary role by translating technical and ethical standards into practical guidance, resources and training. The following areas outline where BEE SECURE can strengthen its contribution to school-based AI literacy and safe implementation practices:

  1. Promote transparency literacy: Provide workshops and materials that help teachers and students understand why AI systems produce certain outputs or scores, and how to interpret uncertainty or bias indicators (UNESCO, 2023).
  2. Raise awareness of bias and data diversity: Develop training modules that explain algorithmic bias, subgroup disparities and fairness, including classroom-ready examples. Require vendors to publish subgroup bias tests and allow school-level validation (Baker & Hawn, 2022).
  3. Empower human oversight: Emphasise the teacher’s role as final decision-maker in the use of AI tools and grading; communicate limits of automation in high-stakes contexts (Holmes et al., 2022; OECD, 2023).
  4. Provide data-protection guidance: Continue to translate complex data-governance principles into accessible teaching resources and parent information: separate learning analytics from commercial profiling; use strict contracts for retention and access (UNESCO, 2023; Williamson & Eynon, 2020).
  5. Teacher and student AI literacy: provide sustained professional development and student curricula aligned to AI-literacy competencies; include hands-on critique of AI outputs (Long & Magerko, 2020; OECD, 2023).
  6. Promote cognitive-load-aware AI use: Support teachers in designing tasks where AI use enhances – not overwhelms – learning, e.g. scaffolded activities with worked examples and fading prompts; embed AI outputs into structured tasks (worked examples → faded prompts → independent practice); avoid raw dumps of long, unsequenced text (Paas et al., 2020; Sweller et al., 2019).

Beyond optimisation of outputs, education aims at subjectification, orientation, and responsibility. In this sense, AI can support — but must not replace — those pedagogical processes through which students come to judgment and agency (Biesta, 2010).

Annotate

Next Chapter
4 The situation in Luxembourg
PreviousNext
CC BY-SA 4.0
Powered by Manifold Scholarship. Learn more at
Opens in new tab or windowmanifoldapp.org