- Updated: June 16, 2026
- 6 min read
Learning after COVID-19 and the ICT career aspirations: Are students entering the AI era with weaker skills?
Direct Answer
The paper Learning after COVID‑19 and the ICT career aspirations: Are students entering the AI era with weaker skills? introduces a cross‑national, longitudinal analysis that links post‑pandemic learning environments to shifts in students’ ICT‑related career aspirations. It matters because the study uncovers which educational levers—digital skills, teacher support, and student autonomy—most reliably predict readiness for an AI‑driven labor market.
Background: Why This Problem Is Hard
COVID‑19 forced schools worldwide into emergency remote teaching, exposing stark gaps in digital infrastructure, pedagogical expertise, and student self‑regulation. Traditional assessments of “learning loss” focus narrowly on test scores, ignoring how those losses translate into future occupational choices. Moreover, existing research treats digital competence, teacher support, and career expectations as separate variables, making it difficult to infer causal pathways or to design coordinated policy interventions.
Compounding the measurement challenge is the heterogeneity of national education systems. PISA provides a common framework, yet each country’s curriculum, socioeconomic context, and post‑pandemic recovery strategy differ dramatically. Without a unified analytical lens that can capture both observable indicators (e.g., test scores) and hidden dimensions (e.g., latent readiness), policymakers lack actionable insight into whether students are truly prepared for the generative‑AI era.
What the Researchers Propose
The authors propose a multi‑layered analytical framework that fuses descriptive statistics, regression modeling, clustering, and deep latent‑representation learning (via a Variational Auto‑Encoder). At a conceptual level, the framework treats each country as a data point described by three pillars:
- Student Autonomy: measures of self‑directed learning, time‑management, and motivation.
- Digital Skills: proficiency in coding, data handling, and online collaboration tools.
- Teacher Support: frequency of feedback, instructional scaffolding, and technology‑enhanced pedagogy.
These pillars feed into two parallel analytical streams. The first stream uses conventional regression to quantify how each pillar predicts changes in ICT career aspirations between 2018 and 2022. The second stream compresses the high‑dimensional indicator space into a low‑dimensional latent vector using a VAE, allowing the researchers to uncover hidden patterns that traditional models miss. Finally, discriminant analysis and probabilistic modeling translate these patterns into interpretable clusters that describe distinct “readiness profiles.”
How It Works in Practice
Imagine a policy dashboard that ingests the latest PISA release, runs the VAE‑based encoder, and instantly assigns each country to one of four readiness clusters: “Digital‑Strong/Support‑Rich,” “Autonomy‑Driven,” “Low‑Skill/Low‑Support,” and “Transitional.” The workflow proceeds as follows:
- Data Ingestion: PISA 2018 and 2022 datasets are harmonized, cleaning missing values and normalizing scales.
- Feature Construction: The three pillars are operationalized through composite indices (e.g., a weighted sum of questionnaire items).
- Latent Encoding: The VAE learns a compact representation that captures non‑linear relationships among the pillars.
- Clustering & Classification: K‑means (or Gaussian Mixture Models) groups countries, while discriminant analysis validates the separability of clusters.
- Policy Insight Generation: Regression coefficients are overlaid on cluster maps, highlighting which levers matter most for each group.
What sets this approach apart is its hybrid nature: it respects the interpretability of classic econometric models while leveraging deep learning to surface hidden structures. The result is a decision‑support tool that can advise ministries on where to invest—whether in teacher professional development, broadband expansion, or autonomy‑enhancing curricula.
Evaluation & Results
The researchers evaluated the framework on three fronts:
- Descriptive Consistency: Across 78 economies, ICT career aspirations rose on average by 7 percentage points from 2018 to 2022, but the increase was uneven—some nations saw gains above 15 pp, while others stagnated.
- Predictive Power: In multivariate regressions, the digital‑skills index exhibited the strongest positive coefficient (β ≈ 0.42, p < 0.001), dwarfing teacher support (β ≈ 0.18) and autonomy (β ≈ 0.07, context‑dependent). The VAE‑derived latent score added an extra 5 % explanatory power (ΔR² = 0.05) beyond the linear model.
- Cluster Validity: Silhouette analysis confirmed four well‑separated clusters (average silhouette = 0.62). Nations in the “Digital‑Strong/Support‑Rich” cluster (e.g., Singapore, Estonia) showed the highest ICT aspiration growth, whereas “Low‑Skill/Low‑Support” economies (e.g., several Sub‑Saharan countries) lagged behind.
These findings demonstrate that digital competence is the primary engine driving ICT career interest, while teacher support acts as a catalyst and autonomy plays a nuanced, context‑specific role. Importantly, the latent representation captured cross‑country synergies—such as the interaction between broadband penetration and teacher training—that were invisible to plain regressions.
Why This Matters for AI Systems and Agents
For AI practitioners building educational agents, talent‑pipeline simulators, or workforce‑forecasting tools, the study offers a data‑grounded blueprint for feature selection. Digital‑skill metrics should be treated as high‑impact signals when training recommendation engines that match students to AI‑related career pathways. Likewise, teacher‑support variables can enrich reinforcement‑learning reward functions that adapt tutoring strategies in real time.
From an orchestration perspective, the cluster‑based readiness profiles enable dynamic routing of resources: an AI‑driven learning platform could automatically allocate more adaptive content to “Low‑Skill/Low‑Support” regions while offering advanced project‑based modules to “Digital‑Strong/Support‑Rich” cohorts. This aligns with the broader trend of “AI‑as‑a‑service” for education, where intelligent agents act as both diagnostic and prescriptive layers.
Integrating these insights into existing ecosystems is straightforward. For example, the OpenAI ChatGPT integration can ingest the latent readiness vectors and generate personalized learning pathways. Similarly, the Workflow automation studio can trigger alerts for policymakers when a country’s cluster shifts toward a lower‑skill profile, prompting timely interventions.
What Comes Next
While the analysis is robust, several limitations warrant further work. First, the study relies on PISA’s self‑reported questionnaires, which may suffer from cultural bias. Second, the VAE architecture was trained on a relatively small sample of countries, limiting its ability to generalize to sub‑national regions. Third, the temporal gap (2018‑2022) captures only the immediate post‑pandemic window; longer‑term trends remain unknown.
Future research could extend the framework with richer data sources—such as learning‑management‑system logs, code‑submission platforms, or real‑time skill‑assessment APIs—to refine the digital‑skill index. Incorporating causal inference techniques (e.g., instrumental variables) would also help disentangle whether teacher support directly boosts ICT aspirations or merely co‑varies with other systemic factors.
Practically, education technology vendors can embed the readiness clusters into product roadmaps. The Enterprise AI platform by UBOS could host a multi‑tenant dashboard where ministries upload their own PISA‑style data and receive cluster assignments in seconds. Start‑ups might leverage the UBOS templates for quick start to prototype AI‑driven career‑guidance chatbots that speak the language of the latent readiness vectors.
Finally, policymakers should consider a balanced portfolio of interventions: scaling broadband access, investing in teacher upskilling programs, and designing curricula that nurture autonomous, project‑based learning. By aligning these levers with the evidence presented in this study, nations can better equip the next generation for the AI‑centric economy.
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.