Integrating AI in TVET: a practical guide for institutions

📜🔗LINK to the source

1. ANALYTICAL SUMMARY

A gap between actual uses and institutional frameworks. AI reconfigures tasks, job profiles and skill requirements in almost all sectors, faster than training systems can adapt (p. 13). The guide notes a gap: 62% of young people are already using AI in practical or work contexts, but only 30% have received formal training (p. 14). Exposure is uneven — gender, rurality, low income, connectivity — and risks exacerbating existing fractures. The integration of AI remains early, transitional and fragmented, concentrated in pilots and innovation hubs rather than deployed at the institutional level (p. 14).

A human and institution-centred action framework. The document proposes a comprehensive institutional approach: five ethical principles, a four-dimensional ecosystem, phased integration based on actual maturity, and seven areas of application (governance, curriculum, pedagogy, assessment, skills, at-risk jobs, innovation). It asserts that AI should augment and not replace professional judgement (p. 20). It provides directly actionable tools — maturity grid, risk register, indicator framework — and emphasises data protection, equity and security as non-negotiable conditions.

2. KEY POINTS OF THE DOCUMENT

1. Five operationalised ethical principles. The guide bases any AI strategy on five interconnected principles: human agency and teacher authority, equity/non-discrimination, educational purpose, transparency/explainability, accountability/data protection (Figure 1, p. 19). Table 2 (p. 20) translates each principle into concrete policy action and pedagogical justification — for example: imposing a human review step in any AI-assisted assessment, conducting an ‘equity’ impact study before any deployment.
2. A 4-dimensional ecosystem and a non-sequential phased integration. The ‘AI-ready’ institution articulates four dimensions: roles/partnerships, industry engagement, enabling infrastructure, leveraging existing capabilities (Figure 2, p. 27). Integration follows four phases aligned with maturity — diagnosis, low-risk/high-value applications, pedagogical deepening, ecosystem (Figure 3 and Table 3, p. 32-33). Crucial point: these phases do not form a mandatory sequence; one enters at the level corresponding to their actual maturity.
3. A rethinking of assessment in the face of GenAI, in 4 categories. Table 11 (p. 52) distinguishes assessment where AI is excluded, permitted, integrated, or mobilised by the assessor. Each assessment allowing AI must include a ‘non-delegable anchor’: a step that only the learner, in their context, can perform (p. 53). Proposed methods: process documentation, oral defence, layered assessment, project tasks, written reflections in class. Practical demonstration remains the strongest evidence in manual and high-risk professions (p. 54).
4. A tool-based governance: minimal policy, risk register, 8-step process. The Table 4 (p. 36) sets the minimum elements of an institutional AI policy; the Table 5 (p. 37‑39) lists the risks by category (educational, assessment, governance, legal/ethical, equity, skills, security, environment) and their mitigation measures. The Figure 4 (p. 35) outlines a governance process in eight steps, from defining the vision to the continuous evaluation-revision of the policy.
5. Skills by role, and AI treated as a security issue. The guide breaks down the UNESCO framework of teaching competencies (Table 12, p. 56) into indicative expectations by role cluster — management, teachers, learner support, technical staff (Table 13, p. 57) — and into pathways for learners (Table 15, p. 59). In high-risk professions (health, construction, electricity…), it requires independent manual mastery before AI and human supervision of critical decisions (Table 16, p. 62; p. 61‑63).

3. ACTION TRACKS FOR LOCAL ACTORS

1. Launch the diagnosis before any purchase. Gather a cross-functional group and complete the self-reflection tool on maturity (Annex A, p. 72‑76), which covers seven dimensions. In a constrained context, this phase 1 can last 12‑18 months without any technological investment: its value is the clarity it produces (p. 32).
2. Formalise a minimal AI policy and a data reference. Before moving on to pedagogical deepening, ensure that a usage policy exists and that at least one person or committee is responsible for data governance (p. 32). Rely on Table 4 (p. 36) to cover the minimum elements.
3. Maintain a living risk register. Adapt the model from Annex B (p. 77) to its scale: risk category, cause, existing controls, mitigation measures, responsible party, monitoring indicator. Prioritise risks with irreversible physical consequences for high-risk professions.
4. Redefine assessments exposed to GenAI. Reclassify each task according to the four categories of Table 11 (p. 52), add a “non-delegable anchor” (p. 53) and rebalance towards practical demonstration, oral defence and process proof. Reserve the unsupervised written work for verifying fundamental knowledge (p. 54).
5. Sequence learning: manual skill first. In high-risk professions, define a minimum volume of unsupervised practice hours per learner, distinct from simulated practice, and assess professional judgement without AI (p. 63). AI simulation remains a preparatory step, never a substitute for supervised practice.
6. Establish a light monitoring system. Select a limited set of structural and perceptual indicators within the framework of Annex C (p. 78‑85), establish a baseline, then review regularly — rather than measuring everything at once (p. 66). Unmet need to anticipate: the guide assumes organised institutional data, which is difficult to gather when systems are siloed (p. 51) — plan for prior interoperability work.

4. ADDITIONAL REFERENCES 

1. UNESCO – AI Competency Framework for Teachers (2024). Direct basis of the guide (15 competencies, 5 components); French version available. → https://www.unesco.org/fr/articles/referentiel-de-competences-en-ia-pour-les-enseignants
2. CNIL – Development of AI systems: recommendations for compliance with the GDPR (July 2025). Completes the dimension of "data protection" in the guide within the French and European legal framework (minimisation, legal basis, information for individuals). → https://www.cnil.fr/fr/developpement-des-systemes-dia-les-recommandations-pour-respecter-le-rgpd
3. Ministry of National Education (France) – Framework for the use of artificial intelligence in education (June 2025). National operationalisation of the same principles (educational added value, data protection, frugal use, integrity of assessments, integration of generative AI). → https://www.education.gouv.fr/cadre-d-usage-de-l-ia-en-education-450647

5. FREQUENTLY ASKED QUESTIONS (FAQ)

1. Who is this guide aimed at? At the leaders of institutions, teachers and trainers, curriculum and assessment staff, and anyone involved in planning. It also concerns decision-makers and qualification authorities who shape the system conditions (p. 12).
2. Will AI replace trainers? No. The guide makes a central distinction between augmentation and replacement: AI should extend human capacity, never substitute it. The teacher retains a role of supervision and interpretation (p. 20).
3. How do we know where to start if we are starting from scratch? By conducting a systematic diagnosis of existing capabilities (infrastructure, human, processes, partnerships), without mandatory prior technological investment (Phase 1, p. 31‑32; Appendix A, p. 72).
4. How to assess when learners can use GenAI? By classifying each task according to four categories of AI involvement and integrating a "non-delegable anchor". Practical demonstration and oral remain the most reliable evidence (p. 52‑54).
5. What are the main risks and how to address them? Eight families of risks (pedagogical, assessment, governance, legal, equity, skills, security, environment) with, for each, mitigation measures — see Table 5 (pp. 37‑39) and the register in Annex B (p. 77).
6. What to do in high-risk professions (health, construction, electricity)? Require independent manual competence before AI, define task boundaries, maintain human supervision of critical decisions, and embed responsibility in policy (pp. 61‑63; Table 16, p. 62).
7. How to measure if integration is working? By combining structural indicators (existence of a policy, skills objectives) and perceptual indicators (staff's perceived preparedness, learners' perception, employer feedback), on a limited set and a baseline (Annex C, pp. 66, 78‑85).

6. REWRITING IN EASY LANGUAGE

What the document says

Artificial intelligence is changing work and training.

It is changing the jobs and skills required.

Many young people are already using AI.

But few people have learned to use it well.

This guide helps institutions to take action.

Important ideas

AI should assist trainers. It should not replace them.

The trainer retains control. They are the one who decides.

Each institution must write clear rules.

These rules protect learners' data.

AI must be fair to all learners.

To get off to a good start

First, look at what you already have.

Then, write a first simple rule.

Choose a person responsible for the data.

Test the AI on a single easy action.

Show that it works before going further.

For the assessment

With AI, it is easy to cheat.

Ask the learner to show their work.

Ask them to explain orally what they have done.

In manual trades, watch them do it for real.

For dangerous trades

Some trades have risks: health, construction, electricity.

Here, the learner must first know how to do it alone.

AI comes only after.

A person must always check important decisions.

7. CROSS-ANALYSIS — VALUES OF HEALTH PRACTICES

• Literacy: yes — the guide requires information in clear language, without jargon, on each AI tool used, and integrates AI literacy into the programmes rather than as an isolated module (p. 20-21, 26).
• Empowerment: the learner must retain their agency — see their data, understand its use, be able to contest an AI decision (p. 21).
• Participation: explicit co-construction through participatory design workshops and representation of learners in ethics committees and curriculum reviews (p. 21, 28).
• Community health (collective dimension): innovation is linked to the economic, social, and environmental needs of local communities (p. 65).
• Ethics: cultural and social biases are identified and addressed as an ongoing operational responsibility, with an AI ethics committee and bias testing (p. 21‑22, 24).
• Human rights : equity, inclusion and non‑discrimination are the primary lens of evaluation, anchored in the Recommendation on AI ethics (p. 21).
• Intersectorality : recommended partnerships with industry, qualification authorities, government, communities and international networks (p. 28‑29).
• Partnership : formalised models — mixed ethics committees, joint curriculum groups, regional networks, research partnerships (p. 28‑29).
• Combating discrimination : yes — overexposure of women to GenAI and underrepresentation in technical fields, disability, rurality, language explicitly addressed (p. 14, 21, 60).

8. EVALUATION OF RESOURCE RELIABILITY

Scientific relevance : solid on the sources mobilised (WEF, ILO, OECD, World Bank, Cedefop, UNESCO corpus), with recent data (2024‑2026). Two editorial honesty reservations: (1) several cited references are dated 2026 and appear to be preprints or forthcoming publications (e.g. Lai et al., 2026, marked “Preprint”; Gmyrek et al., 2026; Romeo & Conti, 2026) — to be verified before reuse; (2) the case boxes are institutional self‑statements, not independent evaluations, and the guide itself acknowledges that evidence on adaptive AI simulation “remains at an early stage” (p. 50) and that the evidence “often comes from institutional cases rather than systematic surveys” (p. 14). Bias to note: instrument produced by UNESCO, which promotes its own normative corpus — counterarguments (additional costs, risk of showcase effect, dependence on platforms) are mentioned but remain secondary.

Operational relevance: very high. The three annexes are directly transposable (4-level maturity grid, risk register, framework of structural and perceptual indicators), complemented by an 8-step governance process and decision tables. This is the main field asset of the document.

Internal inconsistencies to report (production remnants): the legal mentions page accumulates two contradictory blocks — a mention “Author: Robert Palmer” and a “Cover photo: © xx” (uncleaned template placeholders) coexisting with the actual authors (Yang, Wu, Tegelbeckers), as well as a double mention “© UNESCO 2025” and “© UNESCO 2026”. Without effect on the content, but to be mentioned if the document is formally cited.

10. STRATEGIC HASHTAGS

#GenerativeAI #VET #DigitalLiteracy #AIethics #EmergingSkills #DigitalDivide #InclusiveEmployability #healthpractices