Reflections from ICT4D 2026

Introduction

There is immense pressure to use AI. At ICT4D 2026, it was impossible to avoid the topic and it was part of almost every session. This prevalence is reflective of the realities faced in everyone’s work. AI and its promises are all around us, whether voluntarily adopted, introduced by leadership or forced upon us by tech giants that have us locked into their systems. This piece is the first of a reflection of my notes and observations during the three days of ICT4D. 

The ICT4D Conference brings together practitioners, researchers, policymakers, and technology experts working at the intersection of digital innovation and international development. The conference explores how data, AI, digital tools, and information systems can help address global challenges while critically examining issues such as adoption, ethics, governance, localisation, and real-world impact. The sessions I attended focused  on the gap between technology and implementation, the responsible use of AI, decolonising data and knowledge systems, predictive analytics, and the importance of building data and learning cultures that enable better decisions. 

More data than we can handle and the promise of AI

What is becoming clear now is that the problem is no longer a lack of data or evidence. There is an abundance of it. About every topic and from every context, whether this data equally reflects different voices and perspectives is another question. The problem now has turned to what to do with all this data and what its actual purpose should be. Many times, it is extracted for reporting: As much data as possible just to show as much as possible. Sometimes it is used to build some sort of dashboards (at least when it’s only quantitative data) with the initial intention of informing decisions. In any case, by now, whether in dashboards, in folders or drives, in reports or across people’s minds, there is too much data to process manually. 

Because of this overload, the use of AI has come as, what feels like, a salvation to this issue. Finally, an easy way to deal with all the analysis and synthesis. One place to drop all survey responses, all baseline reports, all minutes of meetings and out comes: one neatly summarised paragraph to include in the report. 

The appeal of AI for this sector is clear. Everyone has a lot to do. There are not enough resources and often not enough capacity. At the same time, everything is urgent and many processes sacrifice efficiency over consensus. Dumping all information, data and knowledge into a free tool that gives back seemingly perfect outcomes feels like the assistant everyone wishes they had but the sector can’t afford. 

Beyond the obvious issue with that, namely that free tools are never free and that most of them belong to tech giants that give us little to no control over our data, another increasingly prevalent issue is that there is an expectation for AI to do it all and solve it all. Examples mentioned at the conference included the expectation for AI agents to just run through huge data bases and output what’s useful. Maybe the agents should also just go ahead and make the decisions, and maybe even put them into action right away. 

The need for human judgement 

This expectation of AI brings with it several pitfalls. One is that AI can put certain things into action but in the end, especially for humanitarian responses and development impact, there is currently still a human on the ground executing decisions. The second is that AI is biased. For most of these tools it is hard to understand how they evaluate data that is inputted, and what happens with, often sensitive, data when it is used to train new models. While humans are also inherently biased, they are often trained on critical analysis. For AI, we don’t know. Lastly, and this is a point that often seems to be forgotten: If the data is messy or unclear, AI cannot magically fix it. The idea of garbage in, garbage out still applies. Worse, without human judgement, we risk creating a loop where AI outputs feed more AI outputs without critical interjection.

It now just happens faster and at a greater scale, and with a veneer of credibility that makes it harder to question. In the development and humanitarian sectors, building decisions on the wrong data does not just produce a bad output today. It shapes programme design, funding priorities, and operational choices that may ripple forward for years. When AI outputs start feeding other AI processes without a human checking the logic at any point, you are no longer working with intelligence of any kind. You have built a loop where automated outputs validate automated assumptions, and nothing in the system is asking whether any of it is true. Someone at the conference hit the nail on the head and called it “lazy thinking creep”. A gradual erosion of critical judgement, not because anyone decided to stop thinking, but because the tools made it easy not to. The useful role of AI is not to replace human thinking, but to help us with large amounts of information, recognise patterns, and other connections that would be difficult to discover manually.

More, not less

Nevertheless, the conclusion here might be: more tech, more AI, not less. Just in fundamentally different ways than we are thinking about now. The argument made at the conference was mainly around going back to how AI is actually useful. Our human brain is good at structuring and prioritising. Let AI come in where it is useful. Let it help deal with huge amounts of qualitative data we cannot process in the limited time we have. Let it help with recognising patterns. But in the end, the human still has to structure, decide and use it. 

AI should be there to help provide the human with a fuller picture of what is already there. The point made very clear across panels is that “AI should be the last question, not the first.”. Questions about programme effectiveness, responsiveness, better decision-making, data, knowledge, and learning have to start with the problem and the people, and end with the tech. From there, we can ask which tech can best support us and whether the LLM behind is trust-worthy, where it is pulling the data from, and how we can stop the creep of “lazy thinking.” The aim should be to allow for critical thinking in the process, rather than just accepting what the bots say. As a panelist perfectly put it: “In the end, all this tech is there to make your life a little easier, you’re still going to have to think and keep the human in the loop.”. 

Since the conference, one idea I have been thinking about comes from a recent article on building AI-first companies. The argument is relatively straightforward. When a technology fundamentally changes what is possible, you do not simply add it to existing workflows. You redesign how the organisation operates around the new reality. The article is set in a very different context but the thought behind it still applies, I would argue. What is happening all around is what the interviewee, Hans Scheffer, describes as: “That’s [using AI] all just applying technology to an existing operating model,” he said. “You get short-term wins and isolated improvements. But the core underneath isn’t getting stronger. The intelligence of the organisation doesn’t compound – it just gets locally faster.”

Looking at the development and humanitarian sector, I wonder whether we are doing the opposite. Much of the current conversation focuses on how AI can help us analyse reports faster, search repositories more efficiently, or generate better summaries. But these are all attempts to improve systems that have not worked particularly well for decades. Knowledge remains fragmented. Lessons remain trapped within projects. Institutional memory is routinely lost. Reporting often takes precedence over learning. The sector is stuck in colonial systems where power is distributed unequally and where funding is not reaching the places it should. 

We currently use tech as a band-aid on a leaky ship. However, the question is not only how we integrate it into our existing systems, but whether those systems should exist in their current form at all. Rather than asking how AI can help us process more information, perhaps we should be asking what a genuinely (AI-enabled) data and knowledge infrastructure would look like. What would become possible if we designed our knowledge, learning, and decision-making systems from scratch today, knowing that AI exists? What would a true collective brain for the sector look like?