From Fragmented Initiatives to Integrated Systems

May 2026 · GDLN Blended Learning Program, KDI School of Public Policy and Management Certificate No. GDLN-2026071D · 12 hours · April 8–30, 2026

The session recordings are publicly available on the GDLN website at gdln.or.kr for anyone who wants to explore the content directly.

I want to be careful not to write a summary of a training program. Those are easy to write and rarely worth reading. What I want to do instead is describe what actually shifted, the questions the program raised that I did not arrive with, and how those questions now sit uncomfortably alongside my daily work.

The 2026 GDLN Blended Learning Program, organized by the KDI School of Public Policy and Management in Korea, brought together eight sessions on AI adoption, digital infrastructure, governance, environmental applications, data visualization, public sector learning systems, and anti-corruption. The participants were mostly public sector professionals from Asia and Africa, alongside practitioners from higher education, myself among them.

As a public sector practitioner from higher education, I found myself approaching each session from a specific vantage point, not the ministry official or the policy analyst, but the person responsible for building and sustaining learning systems inside a government university, and the founder of an EdTech initiative committed to making that work better. That specificity turned out to be generative. It forced me to ask, in every session: what does this mean for the institutions I work in? And in asking that question repeatedly, some things became clearer than they might have otherwise.

The problem the program kept returning to

Every session, in its own way, circled back to the same underlying tension: technology gets adopted, but capability does not follow.

Session 1 framed it through labor economics. AI as a general-purpose technology, like electricity before it, takes decades to fully reshape productivity, and most of the gains require not just adoption but task redesign, complementary investment, and deliberate organizational change. Simply adding the tool does not move the curve.

Session 7 made the same argument through learning systems. The speaker, a former World Bank global head of capacity with nearly thirty years of experience, opened with a number that has stayed with me: only 10 to 20 percent of training actually translates into sustained behavior change. Seventy percent of what people learn is forgotten within days unless it is reinforced through context, repetition, and application.

I have run faculty training workshops. I know this number is not wrong.

The question the program left me with is not "how do we train more people?" It is "what would it look like to stop measuring success by completion rates and start measuring it by what people are actually able to do differently?"

What I kept bringing back to my work

I work as a Computer Science Lecturer and E-Learning Coordinator at Samara University in Ethiopia. My daily work involves supporting faculty in designing and teaching online and blended courses, coordinating digital learning initiatives, and trying to build something durable in an environment where resources are constrained, internet connectivity is unreliable, and institutional priorities shift frequently. I also facilitate the 5 Million Ethiopian Coders Initiative, a national program designed to build digital and coding skills at scale.

When Session 3 described the World Bank's distinction between digital enablers (infrastructure, connectivity, platforms) and digital multipliers (governance, trust frameworks, interoperability, human skills), I recognized the gap immediately. During my coordination of e-learning initiatives, I have observed that the enabling layer, platforms, course development, faculty training, has received considerable investment. The multiplier layer, institutional policies, data governance, integration between systems, is much thinner.

When Session 7 described the three things that cause AI-enabled learning systems to fail, fragmentation across disconnected initiatives, tool-driven adoption without strategic vision, and weak institutional ownership, I was essentially reading a description of the risk landscape I navigate every semester.

The 5 Million Ethiopian Coders Initiative made this concrete for me during the program. It currently allows open enrollment: anyone can register directly for courses like Android Development or Fundamentals of Programming regardless of prior knowledge or career direction. Many learners join advanced tracks without the foundational competencies required for meaningful engagement. The result is high enrollment figures that mask shallow skill acquisition. We are measuring activity in place of capability, and the platform has the tools to fix this. Diagnostic assessments and prerequisite gating are native Open edX features. They are simply not being used.

When Session 8 closed the program with the argument that digital government requires not just hardware (ICT systems) and software (legal frameworks) but also people, cultural change, ethics, mindset shifts among individuals, I thought about how little the culture dimension gets addressed in most e-learning implementation plans I have seen, including some I have written.

The session that surprised me most

Session 5 was on AI applications in environmental science, specifically spatial interpolation of air pollution data and environmental epidemiology. It was the session most distant from my professional context.

And yet it was the one that most clearly demonstrated what it actually means to build a system rather than a tool.

The research team did not just build a model. They built a pipeline: satellite data gaps filled by a convolutional neural network, ground-level monitoring combined with meteorological and land-use data, spatial and temporal lag variables incorporated to capture how pollution persists and spreads, and SHAP values applied to make the model's predictions interpretable and accountable. Each stage was justified. Each limitation was named.

The application mattered too. The exposure estimates were linked directly to health outcome data, a cohort of COPD patients, followed over time, with time-varying exposure assigned annually based on residential address. The better the exposure model, the better the epidemiological evidence. The better the evidence, the better the policy.

That is a complete system. Most of what I see in educational technology, including work I have contributed to, is not a complete system. It is a well-designed component sitting in an incomplete architecture.

On bias, fairness, and what the algorithm cannot fix

Session 4 on AI risks offered a framework I have not stopped thinking about.

The speaker decomposed algorithmic bias into three sources: base rate differences (real underlying inequalities in the data), measurement error differences (imperfect proxies for what we actually want to measure), and estimation error differences (the algorithm fitting some groups' data better than others). Each source requires a different intervention:

• Base rate differences require structural social policy. You cannot fix them by adjusting the model.
• Measurement error differences require better data collection.
• Estimation error differences are the only ones that can actually be addressed by improving the algorithm itself.

This distinction matters enormously for anyone designing AI-assisted systems in education. When an adaptive learning platform performs less well for students from certain regions or language backgrounds, the failure is likely a measurement problem or a base rate problem, not an algorithmic one. Tweaking the model will not solve it. Collecting better data and addressing upstream inequalities is the only path.

A cautionary lesson worth noting

After the program ended, I came across something that added necessary nuance to the broader conversation about AI in education.

South Korea piloted AI-driven digital textbooks in 2025 for mathematics, English, and computer science, adaptive tools that tailor content to individual learners in real time. The ambition was exactly right. The execution fell short. Teachers were not adequately prepared. Infrastructure was uneven. The rollout was mandated rather than phased. Adoption rates reached only 30 percent, and South Korea's National Assembly subsequently stripped the AI textbooks of their official status as core teaching materials.

(Source: The Korea Herald, August 2025)

The lesson is not that AI in education does not work. The lesson is that ambition without institutional groundwork produces backlash, not transformation. A country as digitally advanced as South Korea, with significantly more resources than Ethiopia, still stumbled when it moved faster than its teachers, its infrastructure, and its culture could follow.

That is a lesson worth sitting with, especially for those of us designing learning systems in resource-constrained contexts.

Data visualization as a design responsibility

Session 6 on data visualization raised a question I think about often in my instructional design work: who is the intended reader, and what do I owe them?

Every chart, every slide, every assessment rubric I design makes choices about who has to do the work of interpretation, me or the reader. Good design shifts that work toward me. In instructional design, we sometimes talk about cognitive load theory in abstract terms. This was a concrete demonstration.

The historical examples, John Snow's cholera map in 1854, Florence Nightingale's rose diagram of Crimean War mortality, were reminders that visualization has always been an argument, not just a representation. The question is whether the argument is honest, whether the design choices serve understanding or obscure it, and whether the intended audience can actually access the insight.

What "integrated system" actually means

The phrase I have returned to most often since the program ended is one from Session 7: "Technology enables reach. Leadership enables adoption. Integration is what enables impact."

In my context, reach is achievable. We have platforms. We have courses. We have faculty who have been trained. We have students enrolled. The 5 Million Ethiopian Coders Initiative is enrolling learners at scale. Reach is not the constraint.

Adoption is harder. It requires leadership that understands what it is asking people to change, and that provides genuine support, not just permission, for that change. This is inconsistent. Some departments engage seriously; others treat the LMS as an obligation.

Integration is the longest road. It means learning systems connected to performance systems. It means course design informed by data on what students are actually struggling with. It means prerequisite-based progression built into enrollment before a learner can register for Android Development without knowing how to write a loop. It means faculty development that happens in context, embedded in the workflow, rather than in a workshop held once a semester and largely forgotten by week three.

Estonia was the example the program offered of what fully embedded learning looks like: public servants not periodically trained but continuously supported, guidance integrated directly into the systems they use every day. That is a long way from where most institutions, including mine, currently sit. But it is a useful direction.

What I am taking back

Twelve hours is a short time. But this program did something valuable, it opened my eyes. It sharpened the questions I bring to my work, clarified the gap between where we are and where we need to be, and gave me a clearer sense of what building integrated systems, rather than accumulating tools, actually requires.

I am less interested now in asking "which tool should we adopt?" I am more interested in asking: What is the institutional problem this tool is supposed to solve? Who owns the outcome? How will we know if it is working? What happens to the data? Who is being served and who might be harmed?

I am more attentive to the gap between completion and capability, and more committed to designing for the latter rather than reporting the former.

I am more aware that fragmentation is not just an inconvenience. It is the primary mechanism by which well-intentioned digital initiatives fail to produce lasting change. The 5 Million Ethiopian Coders Initiative has the platforms, the content, and the national mandate. What it needs is the architecture, structured pathways, prerequisite-based progression, career alignment, and competency-based measurement. Not new tools. A better system.

The program was titled Advancing Public Governance through Digital Transformation. In Ethiopia, public universities are government institutions, so the question was never distant from my work. It sits at the center of it: how do we build institutions that can actually use digital tools to serve people better, rather than institutions that have digital tools and remain structurally unchanged?

That is the question I am carrying forward. It is also the question that shapes what Gere EdTech is working toward, not just a platform, but a deliberate attempt to build the kind of learning system this region needs.

The views expressed in this article are the author's own and do not represent the official position of any institution.

Gebremedhn Mehari Haylu is a Computer Science Lecturer and E-Learning Coordinator at Samara University in Ethiopia, and the founder of Gere EdTech, a practice-oriented initiative working toward a longer-term vision: an online academy committed to expanding access to quality, thoughtfully designed education in resource-constrained contexts across the Horn of Africa and beyond.

This reflection was developed with AI assistance in drafting and structuring, based on my session notes.

← Back to Insights