The Case for a Global Push on Pro-Development AI
Ensuring AI benefits are broadly shared requires a collaborative effort centering lower-income countries.
Windfall Voices is a new series of publications by Windfall Trust, elevating voices from different communities and regions around the world to better understand the AI narratives shaping our collective future.
Our first piece for Windfall Voices explores the concept of pro-development AI: AI applications and capabilities specifically tailored to the humanitarian and development needs of lower-income countries, and built with and by the communities they are meant to serve. It makes the case that a deliberate, large-scale push for pro-development AI is a necessary complement to the efforts around redistributing AI-generated wealth and ensuring broad access to the existing AI capabilities.
Katya Klinova is the Head of AI for Humanitarian and Development Action at the UN Global Pulse, the Secretary-General’s Innovation Lab and an advisor to the Windfall Trust.
We are living through a time of unprecedented investment in AI, driven by expectations that it will yield immense wealth. This wealth might end up heavily concentrated or, according to more optimistic predictions, enable widespread abundance. The optimistic story often goes something like this: powerful, transformative AI is invented and, assuming it is aligned, economic conditions for all humans improve dramatically.
This story is incomplete, of course. In the words of Dario Amodei, himself a strong believer in AI’s potential to transform lives for the better, “Intelligence may be very powerful, but it isn’t magic fairy dust.” The purpose of this essay is to begin mapping out what needs to happen for the positive economic scenario – one in which AI benefits the world’s poorest and most vulnerable, not only a rich and powerful few – to be realized. For the sake of brevity, safety and alignment are assumed as prerequisites and left outside of the scope.
There are several ways the benefits of AI might be shared with lower-income countries. The most frequently discussed include:
Diffusion of AI-enabled innovative technologies and services from high-income to lower-income countries. Some of the AI-enabled improvements in education, health, energy, logistics, and other fields will spread globally through market mechanisms, while others may not because markets often fail to bring the newest technology to the poorest communities. As a result, there is a growing and well-warranted debate over how to ensure broad access to AI-enabled advances.
Fall in consumer prices thanks to AI-driven productivity gains. If AI boosts productivity across the economy, lower prices may spread widely, provided markets are sufficiently competitive (otherwise, producers might simply pocket the savings.) However, even widespread reductions in consumer prices, while welfare-enhancing, do not necessarily translate into broad-based, comprehensive economic development and improvement of living standards: trickle down of affordable goods from higher income countries may not help much to improve transportation, energy or housing infrastructure in lower income countries. For example, smartphones and mobile internet are now affordable enough to be common even in villages with dirt floor housing without sanitation. Such villages can hardly be called prosperous even if people there can buy things like clothing and connectivity at relatively low prices. Furthermore, if AI undermines job growth in lower- and middle-income countries, cheaper consumer goods will be of limited relief.
Redistribution of AI-generated wealth. Proposals under this umbrella range from creating commitments to share profits or equity to paying global dividends. While redistribution in some form may be necessary, it is unlikely to be sufficient on its own as it relies on the willingness of powerful actors to continuously share broadly, and requires robust mechanisms that protect recipients’ political power in a scenario where their labor may become less relevant.
Most likely, none of the above pathways is sufficient on its own to assure broadly shared prosperity, and a combination of them should be pursued along with at least one other crucial one: investing in what I will refer to as pro-development AI—applications and capabilities designed specifically for the development and humanitarian challenges of lower-income countries, with and by the communities they aim to serve. While many “AI for Good” efforts exist, they rarely reach transformative scale. Without a much larger, intentional push, lower-income countries risk being left further behind as AI advances.
Why Pro-Development AI Is Essential Under Any Path of AI Progress and How It’s Distinct
If one believes that in the next few years AI will become capable of performing nearly all economically relevant tasks, then the challenge of ensuring that lower-income countries share in the benefits is less about creating new capabilities for them – an omni-capable AI would, theoretically, already have these – and more about guaranteeing access to the technology and redistribution of the wealth its generates. But the question of how to direct and incentivize the deployment of these capabilities toward uses that benefit lower-income countries remains.
For example, consider a hypothetical AI system capable of effective drug discovery. Its existence does not guarantee that it will be used to develop treatments for diseases prevalent mainly in lower-income countries, for the same reasons the pharmaceutical industry has long underinvested in such treatments. A comprehensive pro-development AI effort would include bringing to life AI uses of this kind when the markets fail to deliver them.
If, however, omni-capable AI does not emerge in the near term, then for the foreseeable future we are likely to face an advancing but “jagged” capability frontier: AI performing impressively on some tasks while underperforming, or failing, on others. In this case, ensuring that capabilities relevant to lower-income countries’ needs are developed at all becomes crucial. Simply providing access to capabilities created in and for high-income countries may not suffice.
Consider one example. Development economists have long argued that one of the most pressing barriers to economic progress is low productivity in services and the informal economy, which employs much of the labor force in many lower-income countries. LLM-based AI systems, however promising for white-collar productivity, are unlikely to play a decisive role in overcoming this challenge. If AI were to help move the needle on it, new and different capabilities would likely be required.
An effort to advance pro-development AI would consist of both ensuring that the currently existing capabilities get deployed in response to the relevant humanitarian and development challenges, creating pro-development AI uses, as well as investing in the development of novel pro-development AI capabilities. As such, pro-development AI would be characterized by the following features:
Being problem-driven, not solution-driven: the goal should be to address humanitarian and development challenges (some of which can be profoundly helped by AI and some of which cannot), not try to apply AI to as many challenges as possible.
Most likely narrow: since pro-development AI capabilities would be created in response to specific development challenges, they are more likely to be narrow vs general-purpose.
Benchmarked in novel ways: in order to design and measure the progress of pro-development AI, new and different benchmarks datasets would need to be created, tailored to and grounded in the specific humanitarian and development challenges the AI is intended to help addressing (i.e. in the example above, a benchmark could be productivity of human+AI teams performing service sector jobs).
Built for, and by the communities it is meant to serve: while AI created in high-income economies might have plenty of useful applications in lower-income economies, in order to maximize relevance to the challenges that are unique to the latter pro-development AI should be designed and owned by the people with a first-hand understanding of these challenges.
To accelerate pro-development AI, it is useful to unpack the barriers that currently impede its advancement.
A taxonomy of barriers to the advancement of pro-development AI
Below I sketch a map of the barriers that any effort to accelerate pro-development AI – whether market-driven, philanthropic, or policy-based – must overcome. Reducing these barriers would help unblock market mechanisms for advancing the relevant capabilities and applications and lower the need for philanthropic or policy interventions. But only to a point: some AI uses and capabilities critically relevant to lower-income countries may remain unprofitable for private actors, or far less profitable than developing products for higher-income markets. Bringing such applications and capabilities to life will require coordinated action from policymakers, philanthropies, the private sector, and development organizations.
Barriers Affecting the Advancement of Pro-Development AI Capabilities
Path dependence in AI R&D: as I described elsewhere, AI researchers often prefer building on existing advances to pursuing entirely new capabilities, which are or may appear riskier.
Lack of shared understanding of priority capabilities: there is no broadly agreed list of the capabilities that would most benefit lower-income countries, nor of the benchmarks needed to measure progress on those.
Barriers Affecting the Advancement of Pro-Development AI Uses
Inefficient discovery of high-impact AI use cases relevant to lower-income countries due to
Weak mechanisms for decentralized use case discovery and experimentation, in part fueled by the lack of a market, low awareness of opportunities, and scarcity of examples;
Fragmentation and duplication across the AI for Good ecosystem.
Challenges around access and use of existing models and capabilities due to
High API access costs;
Limited language support;
Weaker performance in specific domains or local contexts;
Country availability restrictions.
Barriers Affecting Both Capabilities and Uses
Capital constraints and limited investment;
Infrastructure constraints including access to computing resources and broadband;
Data availability constraints;
Relative scarcity of necessary expertise;
Regulatory barriers;
Heightened risks and sensitivities of deployment and use contexts, which, when present, impose strict requirements for model performance and accuracy validation, as well as accountability and explainability.
Coordinating a global effort to turn pro-development AI from concept to reality
If we want AI to enable broad-based economic development, we must commit to building the frameworks, institutions, and incentives that bring into being capabilities and applications designed for and built with the lower-income countries and communities they aim to serve. An effort to accelerate pro-development AI and address the barriers discussed above could take the form of a large-scale collaboration platform and coordination mechanism – one that identifies shared needs that AI could address but that markets currently overlook, engages deeply with participating countries, and channels funding, talent, compute, and data toward those needs.
At the UN Secretary-General’s Innovation Lab, we have been prototyping such a platform through DISHA, an initiative that brings together partners across technology, philanthropy, academia, IGOs, and NGOs. DISHA bridges the gap between relevant AI research emerging from the private sector and academic labs and putting it in service of humanitarian and development needs at scale. Since DISHA's inception, we have accumulated substantial learnings on navigating structural, logistical, legal, and organizational challenges to enable colleagues from very different organizations to work together as one team and take advantage of pooled resources and capacities.
Initiatives like DISHA show that cross-sector collaboration can surface neglected needs, align global actors, and translate breakthrough research into real-world impact. As AI continues to advance, scaling efforts like this becomes an urgent imperative. Advancing pro-development AI needs to become a first-class priority, backed by serious funding, commitment of actors across the ecosystem, and political will. Without that, we are more likely to face the future of deepened and entrenched global inequalities than universal abundance.
AI may shape the next chapter of human development. Whether that chapter uplifts the world’s poorest depends on what we choose to do now.
Disclaimer: The opinions expressed are solely those of the author and do not necessarily reflect the official views of the United Nations or its member countries.
Acknowledgements: The author is grateful to Miles Brundage for the in-depth discussion of the first draft of this argument, to Deric Cheng, Anna Yelizarova, Brandon Jackson, and Yolanda Laanquist for the insightful suggestions on the subsequent versions.
 | A guest post by
|