Vocational Training in the Generative AI Era — Can Institutional Design Keep Up with Technology?
Structural analysis of generative AI's labor market impact and the effectiveness of reskilling policies. Examining the gap between institutions and technology
What's Happening
More than three years since ChatGPT's public release, generative AI capabilities continue to improve at a pace that exceeds initial predictions. Text generation, code writing, image creation, data analysis—many intellectual tasks once considered "uniquely human" can now be performed by AI.
Generative AI Labor Impact vs Institutional Response Gap
Generative AI Impact Spectrum
Office workers, accounting, translation, support
Programmers, marketers, designers, researchers
Caregiving, construction, agriculture, in-person services
Current Vocational Training System
IT courses = 18% of total; AI-related ≈ 0
Most eligible courses are legacy certifications
Utilization rate < 2% of eligible workers
OJT-heavy; systematic digital training at only 23% of large firms
GPT-4 → GPT-5
~12 months
Training curriculum revision
~24–36 months
Legal reform (labor law etc.)
~36–60 months
Institutional Gap
- 1Speed: AI evolves monthly vs training programs revised annually
- 2Targeting: Highest-risk office workers are underserved by reskilling
- 3Scale: Japan ALMP spend = 0.1% GDP (1/5 of OECD avg 0.5%)
- 4Evaluation: No systematic outcome measurement for training programs
What impact will this transformation have on labor markets? According to OECD's 2024 estimates, approximately 27% of workers in member countries are exposed to "high risk of AI automation"—those engaged in occupations where over 60% of tasks can be replaced by AI. Clerical work, accounting, translation, customer support—routine intellectual labor faces the highest risk.
Meanwhile, about 40% of workers are predicted to benefit from "complementary effects," where AI enhances productivity and increases the value of existing skills. Programmers, marketers, researchers, and other professionals who can effectively utilize AI as a tool fall into this category. The remaining 33% work in fields like care, construction, and agriculture involving physical or interpersonal tasks, where AI's direct impact is considered limited.
The question is how well institutions can keep pace with this impact. The Japanese government has intensified reskilling support since 2023, announcing a 1 trillion yen investment over five years. However, actual utilization rates remain below 2% of eligible recipients. There exists a structural gap between the pace of technological evolution and institutional response.
Background and Context
Beyond the "Substitution" vs. "Complementation" Dichotomy
When discussing generative AI's impact on labor markets, the debate often falls into a binary of "AI will steal jobs" versus "AI will enrich work." However, reality unfolds not at the "occupation" level but at the "task" level.
Research by MIT's Professor David Autor and colleagues shows that most occupations comprise multiple tasks, with AI substituting only some of them. For instance, a lawyer's work includes case law research (easily replaced by AI), legal logic construction (partially complemented), and building trust relationships with clients (difficult to replace). Rather than "lawyers being replaced by AI," a more accurate description is "the composition of lawyers' work is changing."
However, for occupations where most tasks can be replaced by AI, the survival of the profession itself becomes precarious—call center operators, data entry clerks, junior translators. For people working in these fields, acquiring new skills becomes an urgent challenge.
What's crucial here is the asymmetry: those most at risk are also those with the least access to reskilling resources. Knowledge workers with advanced AI skills can learn independently. But expecting middle-aged non-regular workers engaged in routine clerical tasks to self-study data science is unrealistic.
Japan's Vocational Training System — What's Missing?
Japan's public vocational training system centers on Polytechnic Centers operated by the Japan Organization for Employment of the Elderly, Persons with Disabilities and Job Seekers (JEED). In fiscal 2023, approximately 200,000 people participated. Looking at training course breakdown, skills training for manufacturing and construction still dominates, with IT courses accounting for only about 18%. Courses related to AI and data science are virtually nonexistent.
The Education and Training Benefit system exists, but most eligible courses are tied to obtaining existing qualifications like bookkeeping, certified care worker, or real estate broker licenses. New skills aligned with industrial structural changes—cloud infrastructure operations, prompt engineering, AI-powered data analysis—are insufficiently covered as eligible programs.
There's a more fundamental problem: Japan's ALMP (Active Labour Market Policy) spending is 0.1% of GDP, one-fifth of the OECD average of 0.5%. This pales compared to Nordic countries (Denmark 1.3%, Sweden 0.9%)—a difference of orders of magnitude. Simply put, Japan doesn't "invest money" in vocational training.
Effectiveness of Reskilling Policies — Who's Being Reached?
The reskilling support measures launched in 2023 were a flagship policy of the Kishida administration's "investment in people" agenda. The 1 trillion yen scale over five years was unprecedented. However, questions remain about policy effectiveness.
First, low utilization rates. Actual usage among eligible recipients stays below 2%. Main factors include low awareness, complex application procedures, and time constraints for working people to participate.
Second, mismatch between training content and labor market needs. While private reskilling courses have increased, quality varies significantly. Many cases fail to directly lead to employment or career advancement upon completion, creating challenges in maintaining participant motivation.
Third, absence of impact measurement. There's virtually no systematic follow-up research on how training programs affect participants' employment rates or income. This represents investment without evidence—the current reality.
Reading the Structure
The Speed Problem — Can Institutions Catch Up with Technology?
Generative AI's evolution speed differs qualitatively from conventional technological change. From GPT-3.5 to GPT-4 took about 12 months. During that period, AI capabilities surpassed bar exam passing standards and reached medical licensing exam levels. In contrast, revising vocational training curricula takes 24-36 months, while labor law amendments require 36-60 months.
This speed differential is structural and cannot be resolved through effort alone. Rather than aiming for institutions to "catch up," we need to design mechanisms that continuously "adapt" to change.
Denmark's flexicurity (flexibility + security) model offers one reference. By relaxing dismissal regulations while combining generous unemployment benefits with active vocational training, it increases labor market fluidity while having society bear individual risks. In an era of accelerating technological change, systems premised on "spending one's entire career in a single occupation" struggle to function. What's needed is institutional design that assumes occupational transitions and supports those transitions.
From "Training" to "Learning Infrastructure"
Existing vocational training follows a "teach specific skills" model. However, in the generative AI era, the skills taught may become obsolete within years. What matters is not individual skills but building infrastructure that supports the "ability to keep learning" itself.
Specifically, three shifts are necessary:
First, training personalization. Rather than providing identical curricula to everyone, design personalized learning paths based on individuals' existing skills, experience, and aptitude. Ironically, AI itself can facilitate this personalization.
Second, diversified assessment. Beyond certificates and completion credentials, systematically recognize micro-credentials (mechanisms that certify short-term learning outcomes) and portfolio assessments. Systems premised on long-term qualification acquisition are too cumbersome for rapidly changing times.
Third, joint design by companies and educational institutions. Institutionalize mechanisms where companies participate from the training program design stage and directly connect to hiring upon graduation. Germany's dual system and Britain's apprenticeship programs serve as reference models.
Generative AI has already begun showing "what will happen" in labor markets. The question is whether institutions can respond to that speed. There's no need to fall into technological determinism. However, if institutions continue ignoring change, the bill will be paid by those most at risk. Delayed institutional design is synonymous with expanding inequality.
Related Columns
References
OECD Employment Outlook 2024 — The impact of AI on the labour market
OECD. OECD
Read source
IT人材需給に関する調査
経済産業省. 経済産業省
Read source
「人への投資」施策パッケージ
内閣官房 新しい資本主義実現本部事務局. 内閣官房
Read source
New Evidence on the Effect of Technology on Employment and Skill Demand
Autor, D., Salomons, A.. Annual Review of Economics
Read source