When I searched my mind for the right word to describe my thoughts on AI and development, Albert Hirschman’s notion of ‘possibilism’ sprung into view. ‘Possibilism’ is an approach to development that rejects both pessimism — or a sense that the world is already determined by forces beyond our control — and an over-optimistic adherence to one-size-fits-all solutions. Instead, Hirschman was interested in finding “the possible” within all contextual constraints, and urged us to be creative and experimental in looking for the incremental openings for change.
Living in an extractive and unequalizing AI economy
Most readers of this newsletter will know that a small number of companies have established dominant positions in AI. For starters, NVIDIA’s advanced chips and CUDA platform provide the structural backbone to the AI ecosystem, while the widespread adoption of its libraries among developers reinforces this dominance. Meanwhile, Microsoft and OpenAI, Google, Amazon, Anthropic, and Meta have amassed huge databases, combining publicly available data with private or subscription-based sources like scientific journals.
These corporate behemoths control the chokepoints through which data flows through our world and make it extremely difficult for newcomers to break through. Most governments in low- and middle-income countries cannot simply copy China in creating “national champions” in IT, AI hardware or large language models. They have much smaller markets, much more limited state capacity, and far more constrained bargaining power.
The rise of global value chains, partly enabled by digitization, has increasingly simplified and automated production tasks and trapped the workforces of many developing countries in low-skilled segments of the global economy.
It is also important to situate AI within a set of broader transformations that have restructured global economies and knowledge systems in ways that steepen the development landscape. The rise of global value chains, partly enabled by digitization, has increasingly simplified and automated production tasks and trapped the workforces of many developing countries in low-skilled segments of the global economy. Meanwhile, in education and science, the promotion of academic capitalism and the strengthening of multilateral intellectual property protections have shored up the commercial dominance of their scientific research.
Alternatively, in many global majority countries, universities have instead faced budget cuts, and scientific initiatives have become more beholden to donor funding and foreign research agendas. Even international collaborations aimed at addressing “global challenges” like world hunger and climate change often end up strengthening the commercial advantages of high-income countries. AI is likely to accelerate these dynamics, potentially reinforcing the dominance of high-income countries while also disrupting professional networks and high-wage economies built around their intellectual property.
It is no wonder that we live in a haze of pessimism, bombarded with narratives of data colonialism and extractivism. And yet, much of the world has yet to be “datified.”
Leveraging cross-border frictions to shape counter-realities
In many sectors and regions, data remains fragmented, incompatible, and shaped by sector-specific market structures, proprietary systems, privacy regulations, and institutional bottlenecks. Efforts to standardize and share data entail huge coordination efforts and prohibitive costs. This is very much the case in agricultural research worldwide. Big institutional customers are also cottoning on and beginning to exert greater control over their data. Governments are rethinking IP laws and the regulations surrounding copyrighted material (as my PhD student Javiera Caceres’ research is highlighting). Publishers and scientific actors are likewise rethinking subscription and data-access models.
Much of the world has yet to be captured in data, and there are many points of leverage left. All this fragmentation and friction, far from being a problem to be “fixed” by Data for Development and Digital Inclusion initiatives, represents a form of possibilism, in my view.
While talk of data colonialism helps highlight power asymmetries, its totalizing narrative can distract us from the distributed agency and variation that mark our world. Much of the world has yet to be captured in data, and there are many points of leverage left. All this fragmentation and friction, far from being a problem to be “fixed” by Data for Development and Digital Inclusion initiatives, represents a form of possibilism, in my view.
Most countries are unlikely to produce the next OpenAI or NVIDIA, but their scientists, workers, and firms can still find ways to use AI in creative and incremental, sector-specific ways to address their development challenges. But rather than offer advice to countries seeking to engage with the sector as a whole, my advice would be to situate a discussion of AI within existing sectors and ask some fundamental questions first:
- What are the core sectors that employ the most workers and firms within each region and country?
- What technological capabilities do they have, and what do they need to move into more lucrative activities, sectors or markets?
- How can we mobilize investment and build the technical capacity required to achieve those upgrading and structural change objectives?
And only then should we turn back to digital architectures and artificial intelligence and ask how they might help achieve those goals. We should not put the cart before the horse.
It is here that fragmentation and friction come in, moving our attention away from abstract, technocratic predictions about AI and its global trajectories and instead seeking out situated understandings of how AI might be shaped within very specific development contexts. Across the world, people can make more careful decisions about when and how to share or withhold data, how to leverage contracts and local content requirements, how to manage ethical clearances within research, and how to build local or domestic data pools and databases that serve specific development goals.
Finding possibilities in local requirements and contexts
This approach seems much more sensible than trying to integrate into the AI value chain itself. For example, some countries have opted to try to host infrastructure in hopes that they might attract foreign investment. Yet, data centers in northeast Brazil currently offer limited benefits to the local communities and broader domestic economy. Similarly, another of my PhD students, Dr. Tin El Kadi’s recent research on North Africa has shown that contracting Huawei to build and manage data centers to achieve ‘Digital Sovereignty’ won’t necessarily bring economic benefits without a clear domestic strategy to create linkages for local firms or to boost learning in universities or public institutions.
Meanwhile, some governments have offered up their low-skilled, low-wage workers to help train AI systems. This kind of engagement may very well create short-term jobs, but it is very unclear how it will create meaningful opportunities for skill development, economic security, or long-term structural change.
Meanwhile, some governments have offered up their low-skilled, low-wage workers to help train AI systems. This kind of engagement may very well create short-term jobs, but it is very unclear how it will create meaningful opportunities for skill development, economic security, or long-term structural change. Indeed, these workers may be training for their own obsolescence. Finally, employing AI firms to increase institutional efficiency may be convenient, fast, and cheap, but it does not prioritize domestic learning and risks reproducing the very dependency development aims to surmount rather than fostering domestic capabilities. Developing countries need to be much shrewder about how they want to engage strategically with this technology.
I would argue that AI must be understood within specific places and value chains. The topic of the technology itself should temporarily recede into the background so that policymakers and analysts can focus on learning challenges within key sectors, before considering how digital tools can amplify those capabilities.
This text was originally presented as a lightning talk at the ReGen AI conference organized by ITforChange in Bengaluru on November 1, 2025. It has been revised for publication.
