This chapter identifies barriers and potential risks associated with the use of artificial intelligence (AI) in vocational education and training (VET) curriculum and qualification development. Drawing on OECD survey evidence and country evidence, the chapter analyses the main barriers limiting effective, secure and inclusive AI use, including limited skills and capacity, unequal access to secure and fit‑for‑purpose tools, data constraints, lack of guidelines and support, and concerns about job displacement and devalued expertise. It also examines potential risks, such as weakening collaborative governance, undermining accountability and quality, and exposing VET systems to data protection and security challenges.
Developing Vocational Education and Training with Artificial Intelligence
Abstract
Barriers, potential risks and strategic directions in using AI to support VET curriculum and qualification development
What barriers and potential risks are associated with AI use in VET development?
While artificial intelligence (AI) holds potential to enhance the efficiency, responsiveness and evidence base of vocational education and training (VET) development, its use remains nascent, uneven and largely exploratory across OECD countries. Uptake is constrained by limited AI literacy, resource gaps, fragmented experimentation and the absence of sector-specific guidance.
Concerns among AI users persist regarding data protection, intellectual property, security and the quality and reliability of AI-generated outputs. Effective use depends on access not only to tools but to high-quality, interoperable and machine‑readable data; yet fragmented governance, legal restrictions and weak data infrastructures limit scalability and comparability. Stakeholders highlight risks of over-reliance, diminished professional judgement, weakened multi-stakeholder governance and blurred accountability if AI is not carefully embedded within established validation and decision making processes.
How should countries address these barriers and risks?
These challenges are not purely technical but organisational, pedagogical and cultural: successful integration requires clear purpose, proportionate regulation, robust data governance, transparent allocation of responsibilities and sustained human-centred approaches and oversight. Moving beyond isolated pilots towards system-level impact therefore depends on coherent national and sectoral strategies, targeted capacity building and support for access, clear guidelines, and interoperable data ecosystems and human-centred governance frameworks that balance innovation with quality, equity, trust and regulatory stability.
Artificial intelligence (AI) is emerging as a potentially transformative tool for vocational education and training (VET), with growing use in labour market analysis, curriculum design and evidence‑based decision‑making (Delia, 2025[1]; ILO, 2020[2]; UNESCO-IIEP, 2025[3]; UNESCO, 2025[4]). Across countries, AI is beginning to support faster updates of curricula and qualifications and improved alignment with evolving skill needs (OECD, 2025[5]). Yet its integration into VET curriculum and qualification development remains at an early and uneven stage. Current use is largely informal or experimental, often driven by individual initiatives rather than coherent system‑level strategies. As a result, VET stakeholders face regulatory ambiguity, limited human and data capacity and unclear governance arrangements when determining whether and how AI could be used.
While the policy rationale for using AI in VET development centres on efficiency, relevance and timeliness, its adoption also introduces potential risks and trade‑offs. These include concerns about VET accountability, quality assurance, stakeholder ownership of AI‑generated outputs, and the possible weakening of collaborative, multi-stakeholder decision making that underpins VET systems. Addressing these challenges requires more than technical solutions: they are also organisational, ethical, pedagogical and cultural. Ensuring that AI strengthens rather than undermines quality, equity and public trust requires careful navigation and clear strategic direction. Drawing on OECD surveys and interviews conducted as part of this study (OECD, 2025[5]), this chapter examines main barriers and potential risks associated with AI use in VET development, and outlines strategic directions that VET systems and stakeholders need to consider addressing these barriers and risks.
Across countries, VET stakeholders report significant constraints limiting both uptake and effective use of AI in VET development. Non-users most frequently cite limited capabilities: more than half of surveyed industry partners and providers indicate lacking the necessary capabilities and expertise. While a third of them report that existing procedures are considered sufficient, equal share reported resource constraints or high upfront costs as reason for not suing AI. Concerns around data security, privacy and ownership prevent VET stakeholders, particularly policymakers and industry partners, from using AI to support VET development (Figure 4.1, Panel A).
Even among stakeholders already experimenting with AI, implementation challenges remain. Reported obstacles include lacking AI‑related skills (58% of respondents using AI to support VET development), uncertainty over the quality and reliability of AI‑generated materials (50%), and limited openness or agreement (30%) among stakeholders involved in VET development. Technical and infrastructure limitations, development costs and the difficulty of interpreting or integrating AI‑generated insights are present, yet relatively less prominent (Figure 4.1, Panel B). The subsections discuss these barriers.
Source: OECD (2025[5]), Developing vocational qualifications and curricula with AI: Surveys and stakeholder interviews.
Lack of guidelines and support that address limited, uneven capabilities and access
Despite widespread national AI strategies and emerging AI use cases (OECD, 2025[6]), specific guidance for VET curriculum and qualification development or VET sector-wide guidelines remain limited. Only a small minority of surveyed stakeholders (15%) report the existence of relevant guidelines (OECD, 2025[5]). Where guidance exists, it often focusses on classroom use or general public-sector, internal rules restricting sensitive data processing, rather than addressing the unique challenges of curriculum design, stakeholder consultation or validation processes. In some cases, it is poorly implemented or not widely known among staff – reported by an employers’ association in the Netherlands that contributes employer perspectives to qualification profiles, learning outcomes and skills needs in their industries and manages extensive employer and labour‑market data, including job descriptions, competence profiles and salary benchmark data.
In several countries, such as in Germany, Ireland and Mexico, limited AI, digital and data literacy – affecting curriculum developers, VET teachers and even IT and technical staff in VET agencies and institutions – constrains effective integration (Figure 4.1, Panel A). Generational gaps (OECD, 2025[7]) may be particularly pronounced in VET, where expert committees often include experienced professionals with limited exposure to AI tools and many VET teachers are over the age of 50. Croatia, Estonia and Lithuania stand out in terms of generational gaps in digital skills (Figure 4.2). In Mexico, many trainers and curriculum developers rely on self-learning and experimentation, calling for structured capacity-building programmes. Moreover, stakeholders report that low familiarity can increase mistrust in automated outputs and slow adoption. Conversely, overconfidence without adequate competence risks over-reliance and diminished critical reflection and professional judgement.
Capacity constraints and lacking support and guidance also create equity concerns. Well-resourced institutions may experiment with advanced tools, while smaller or under-resourced providers lack access to training and secure systems, widening the digital divide within VET (OECD, 2025[5]).
Share of Individuals with above basic overall digital skills, 2023
Note: All five component indicators are at above basic level: information and data literacy, communication and collaboration, digital content creation, problem-solving, and safety skills.
Source: Eurostat (2026[8]), Individuals’ level of digital skills (from 2021 onwards) (isoc_sk_dskl_i21), https://ec.europa.eu/eurostat/databrowser/view/ISOC_SK_DSKL_I21/default/table.
Lack of access to effective, secure and tailored AI systems and quality data to inform AI-assisted VET development
Access to AI also has multiple layers: from basic use of generic tools to access to specific models trained on high-quality, domain-relevant data. Effective AI use depends not only on access to tools, but on reliable, structured and interoperable datasets. The quality and reliability of AI outputs is closely tied to the quality, interoperability and representativeness of training data. Yet, many VET systems often lack high-quality, harmonised, accessible and machine‑readable datasets. AI models trained solely on open, historical datasets may fail to anticipate emerging skill needs, resulting in outdated or misaligned qualifications and curricula. Poorly structured and inconsistent data can generate misleading outputs, particularly when AI is used to forecast labour market demand or identify future skills (OECD, 2025[5]).
The difference in AI use among VET stakeholders (Figure 3.1, Chapter 3) may reflect differing capacities, incentives and resource availability. While firms and public VET agencies are increasingly integrating AI into VET development processes as well as other tasks including administrative ones, VET institutions may face barriers – e.g. fragmented data ecosystems, limited digital and analytical capacity, and unclear governance or ethical frameworks. Larger organisations in manufacturing and finance sectors tend to adopt AI more readily (Lane, Williams and Broecke, 2023[9]) and AI adoption correlates with greater autonomy, more complex tasks and higher level of education (Arntz et al., 2025[10]), suggesting similar patterns for VET institutions.
In Mexico, fragmented databases limit AI’s capacity to map learning outcomes to occupational standards, while data gaps in sectors such as cultural and creative sectors – such as performing arts, visual arts, cultural heritage and crafts – introduce bias and implementation gaps. Limited availability of reliable, AI-processable quantitative data leads stakeholders to rely more heavily on qualitative inputs, even when using AI. Preparing dialogue templates and prompts for open-access AI tools to analyse qualitative material is time‑consuming compared to using AI systems tailored to targeted datasets, making dataset development more resource‑intensive. Stakeholders also noted that data privacy concerns, uncertain security in publicly available tools, high costs of secure paid solutions and the absence of clear organisational guidance on standard tools further hinder effective and wider adoption (OECD, 2025[5]).
By contrast, countries such as Germany use AI more frequently for quantitative labour-market analysis, supported by relatively well-developed databases. Still, German stakeholders face fragmented data access arising from decentralised governance, strict data protection laws and sector-specific (not cross-sector) data systems, limiting AI’s potential beyond discrete applications such as labour-market forecasting and large‑scale text analysis, including BIBB’s monitoring of qualification trends. The absence of a comprehensive, interoperable and longitudinal database prevents cross-sector or cross-region analysis and constrains deeper use in curriculum and qualification development. No AI model has yet been specifically trained on VET data for these purposes, and frequent AI model updates may weaken reliability, consistency and comparability over time (OECD, 2025[5]).
Croatia’s barriers include data accessibility and intellectual property constraints. Micro-credential and adult VET content developed by colleges and universities are treated as institution-owned assets and commercial outputs, particularly where they support fee‑based or continuing education provision. This restricts open access to, and reuse of VET content and data sharing. While these training programmes are accredited and verified by national agencies, the full content is not publicly shared; only general information such as learning objectives and duration is disclosed. Institutions retain ownership of the detailed programme design, which creates competitive advantages and limits open access. In addition, this exclusivity means that even if AI tools are used to develop content, questions remain about whether AI-generated materials can be claimed as institutional intellectual property. The absence of clear guidelines on AI-created content further complicates this issue. This limits interoperability and hinders collaborative efforts to build AI-driven tools that rely on large, diverse and relevant datasets for curriculum alignment (OECD, 2025[5]).
Costs further shape access, as mentioned by Mexican stakeholders. Advanced or secure AI solutions can be prohibitively expensive for VET institutions with limited resources. In Estonia, while a government agreement with OpenAI under the AI Leap initiative facilitates access, broader use of professional-grade AI tools may require additional funding or national co‑ordination. Without equitable access arrangements, better-resourced institutions are more likely to benefit, potentially widening the digital divide (OECD, 2025[5]).
Concerns about job displacement and devalued expertise
Stakeholders in some countries express anxiety that AI could displace professional roles (e.g. curriculum developers, education consultants and educators) or diminish the value of expert judgement in VET development. While no country reported replacing human decision making in qualification development, concerns persist that automation could gradually weaken the stakeholder-driven and collaborative processes central to VET governance (OECD, 2025[5]).
Evidence from other public-sector domains suggests substantial automation potential in routine tasks. The Alan Turing Institute estimates that the AI could automate 84% of repetitive public service transactions in the United Kingdom, saving the equivalent of 1 200 person-years of work annually (OECD, 2025[6]). Public service workforce displacement could also occur if governments seek to replace rather than augment civil servants’ capabilities (OECD, 2025[6]).
As shown in Figure 4.1 (Panel B), even stakeholders already experimenting with AI report implementation challenges linked to perceived risks associated with AI use. While the capability issue and limited openness or agreement among stakeholders (30%) and others (technical and infrastructure limitations, development costs and use of AI-generated insights) are rather manageable over time, concerns about the quality and reliability of AI‑generated outputs persist. Interviews confirm survey findings: stakeholders emphasise risks of over-reliance or premature use lowering quality, alongside data protection, security and privacy concerns (Figure 4.1, Panel A). Interviews also revealed more VET-specific issues: weakening collaborative, multi-stakeholder decision making in VET development, if not carefully embedded; undermining transparent and accountable VET development processes as well as overall VET accountability and reputation, if roles and responsibilities are not clearly defined and managed. The following subsections discuss these risks (OECD, 2025[5]).
AI use potentially weakening collaborative, multi-stakeholder decision making in VET development, if not carefully embedded
Shared responsibility among public authorities, VET providers, teachers and trainers, employers and social partners is a longstanding or sought feature of many VET systems. Curriculum and qualification development typically relies on consultative processes to ensure relevance, legitimacy and acceptance by those responsible for implementation. As AI begins to enter these processes, questions arise as to how it interacts with established governance arrangements and stakeholder roles (OECD, 2025[5]).
Evidence from countries suggests that early AI uses often occurs through isolated experimentations or pilots by individual teachers, institutions or departments, operating alongside rather than within existing collaborative structures – without broader co‑ordination. In England and Estonia, for example, stakeholders noted that AI-supported tools initiated by public authorities and AI-generated outputs, though efficient, could marginalise VET teachers and industry partners, resulting in a “shadow curriculum”: formally approved and registered curricula may diverge from what is actually delivered, if AI outputs are insufficiently reviewed or validated. If relevant VET stakeholders are not actively involved in AI-supported curriculum design and validation, their sense of ownership and engagement may diminish, potentially undermining the quality and acceptance of new and revised qualifications and curricula. Similar sensitivities arise in Switzerland and Germany, where consensus‑based traditions emphasise collaborative decision‑making and stakeholder engagement, and in Nova Scotia (Canada) where stakeholders stressed maintaining the primacy of industry expertise (OECD, 2025[5]).
These experiences highlight that the issue is not AI use per se, but how it is embedded within the existing processes. When AI is perceived as supporting dialogue and evidence‑informed discussion, stakeholder engagement may be reinforced. When it is perceived as substituting for consultation or professional judgement, ownership and involvement may decline. Maintaining clear roles for VET stakeholders throughout AI‑supported processes therefore remain important for preserving trust and legitimacy in VET development.
AI use potentially undermining transparent and accountable VET development processes and overall VET accountability and reputation, if roles and responsibilities are not clearly defined and managed
Across countries, VET stakeholders emphasise that the introduction of AI into VET development raises fundamental questions about accountability, transparency and responsibility. While legal restrictions on AI use remain minimal in many countries, inaccurate, biased or poorly validated AI-generated outputs can have significant consequences in systems where VET qualifications carry strong signalling value for employers and learners.
In Nova Scotia (Canada), VET stakeholders highlighted reputational risks if flawed and irresponsibly used AI‑generated outputs undermine confidence in VET curricula and qualifications even if technical errors were corrected later. In Croatia, VET experts expressed uncertainty about the suitability of commercially developed and driven AI models for public‑sector VET functions, particularly VET qualification and curriculum development where transparency and data protection requirements are high. In Germany, stakeholders emphasised that training regulations and qualifications carry strong legal and reputational weight, and warned that AI‑generated drafts, even when technically sound, could weaken accountability if it becomes unclear who is responsible for content, validation and final decisions (OECD, 2025[5]).
These examples point to a shared emphasis on maintaining clear lines of responsibility. Stakeholders across countries underlined that accountability for decisions and outcomes should remain with human actors and institutions, even where AI tools are used extensively. Interviewed stakeholders noted that transparency about AI’s role, alongside explicit validation processes and clear allocation of responsibility (e.g. for correction) are central to sustaining public trust in VET as AI adoption evolves. Over-reliance on AI without oversight could introduce unpredictable changes in qualification frameworks, undermining stability and predictability for employers and learners (OECD, 2025[5]).
AI use potentially lowering VET quality through over-reliance or premature use
VET development is lengthy, resource‑intensive and complex for good reason: it delicately balance labour market relevance, pedagogical coherence and contextual nuance. Although widespread over-reliance on or premature use of AI-generated outputs has not been observed in VET development, stakeholders warn of diminished critical reflection if drafting and analysis become overly automated.
AI cannot replace expert judgement. AI tools trained on biased or incomplete data may embed errors, outdated content or inappropriate terminology into instructional recommendations if not critically reviewed (OECD, 2025[11]). In Germany, VET stakeholders caution against the premature circulation of AI-generated drafts of training regulations, noting their sensitivity could result in misinterpretation or reputational damage if disclosed prior to formal approval – which could lead to legal and economic consequences. In Mexico, VET stakeholders report that roughly half of AI-generated content currently requires substantial revision to ensure appropriate terminology and contextual relevance for VET, particularly where AI tools are trained primarily on higher education contexts and more cognitive learning than job-specific, practical and operational material. Estonia similarly emphasises that AI may assist technical validation and comparison of standards for speed and efficiency, but cautions that AI must not replace human judgment in evaluating the relevance and quality of learning outcomes, as it often lacks pedagogical depth or fails to reflect the nuanced needs of learners and educators, without proper oversight (OECD, 2025[5]).
Lowering quality linked to AI use also raises accountability concerns. When AI-generated outputs are flawed or data breaches occur, responsibility may be unclear, complicating corrective action and exposing institutions and the wider VET sector to reputational risk. AI-generated content may also inadvertently breach legal norms or use terminology inconsistent with regulatory standards. Without rigorous human oversight, over-reliance on AI may result in the risk of errors, outdated material and legal or terminological inconsistencies.
AI use potentially exposing sensitive, unpublished or commercially valuable VET data to misuse, without adequate safeguards
AI use in VET development potentially exposes sensitive, unpublished or commercially valuable data and content to misuse where safeguards are insufficient or unclear. In addition, reliance on skewed, incomplete or biased data in AI systems can cause harmful or misleading outputs, undermining decision making (OECD, 2025[6]). Without robust data governance, security measures and clear accountability, AI use may result in data breaches, loss of control over confidential content or the inappropriate reuse of sensitive information.
Across countries, data insecurity is cited not only as a risk associated with AI use, but also as a barrier to effective use in supporting VET development. In Germany, even with its strict data protection frameworks, VET stakeholders express concerns about AI use in handling of sensitive and confidential data, draft training regulations and other non‑public materials, noting that inappropriate AI use could compromise confidentiality and disrupt established governance processes. Similar concerns are reported in Mexico, where COSFAC points to risks related to data privacy and the misuse of unpublished curriculum and qualification materials (OECD, 2025[5]).
These risks are heightened where AI tools rely on poorly governed data inputs or publicly available platforms that do not meet strict security and privacy standards. Stakeholders across countries emphasise the vulnerability of personal data collected through focus groups, consultation processes and stakeholder interviews, as well as commercially sensitive employer information used in skills analysis. In Croatia and Mexico, concerns centre on the processing of such data through generic AI tools, which may lead to loss of control over confidential content or unintended data disclosure. In the Netherlands, employer organisations stress the need for manual verification and strict controls when using AI‑based web‑scraping of job vacancies to extract skills information for qualification development, particularly where firm‑level or high‑value information is concerned; while effective for identifying emerging skills needs, these tools can generate data quality issues, including misattribution, if safeguards are weak (OECD, 2025[5]).
These experiences underline that without robust data governance, secure infrastructures and clear responsibility for data handling, AI use risks undermining data security and protection, as well as quality and trust of VET development processes.
References
[10] Arntz, M. et al. (2025), Digital Transformation and the Changing World of Work (DiWaBe 2.0): A Data Source for Research on Artificial Intelligence and Other Technologies in the Workplace, https://www.baua.de/DE/Angebote/Publikationen/Berichte/F2573-2.
[1] Delia, C. (2025), European insights: adoption of AI in TVET institutions – challenges, opportunities and recommendations, UNEVOC Research Briefs, https://atlas.unevoc.unesco.org/research-briefs/european-insights-adoption-of-ai-in-tvet-institutions-challenges-opportunities-and-recommendations.
[8] Eurostat (2026), Individuals’ level of digital skills (from 2021 onwards) (isoc_sk_dskl_i21), https://ec.europa.eu/eurostat/databrowser/view/ISOC_SK_DSKL_I21/default/table (accessed 8 April 2026).
[2] ILO (2020), The Digitization of TVET and Skills Systems, https://www.ilo.org/publications/digitization-tvet-and-skills-systems.
[9] Lane, M., M. Williams and S. Broecke (2023), “The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers”, OECD Social, Employment and Migration Working Papers, No. 288, OECD Publishing, Paris, https://doi.org/10.1787/ea0a0fe1-en.
[5] OECD (2025), Developing vocational qualifications and curricula with AI: Surveys and stakeholder interviews, Unpublished.
[7] OECD (2025), “Digital skills for seniors: Key steps for effective training programmes”, OECD Policy Briefs, No. 35, OECD Publishing, Paris, https://doi.org/10.1787/9edfa7ef-en.
[6] OECD (2025), Governing with Artificial Intelligence: The State of Play and Way Forward in Core Government Functions, OECD Publishing, Paris, https://doi.org/10.1787/795de142-en.
[11] OECD (2025), “OECD Teaching Compass: Reimagining teachers as agents of curriculum changes”, OECD Education Policy Perspectives, No. 123, OECD Publishing, Paris, https://doi.org/10.1787/8297a24a-en.
[4] UNESCO (2025), Why AI Belongs in Every Skills Classroom: Training TVET Educators for a Smarter Future, https://www.unesco.org/en/articles/why-ai-belongs-every-skills-classroom-training-tvet-educators-smarter-future-0.
[3] UNESCO-IIEP (2025), Using AI and digital tools to modernize TVET in a shifting global job market, https://www.iiep.unesco.org/en/articles/using-ai-and-digital-tools-modernize-tvet-shifting-global-job-market.