Blog

Could, would, should: What philanthropy owes AI in global development

Tony Senanayake , Mallika Sobti , Rebecca Gong Sharp 21 August 2026

Opinion: Here is a concrete agenda for philanthropy’s investment into AI for good.

Children, who do not have access to internet facilities and gadgets, use laptops in an open-air class outside a house at Joba Attpara village in Paschim Bardhaman district in the eastern state of West Bengal, India. Photo by: Rupak De Chowdhuri / Reuters

This blog was originally published on Devex.

“AI for good” has been the refrain of the past year, and the money is following.

In late 2025, 10 of the largest U.S. foundations committed $500 million over five years to guide artificial intelligence toward human-centered goals. Both traditional and fast-growing AI-aligned funders are putting real money and convening power behind AI deployments designed to target some of the world’s hardest developmental problems. We may be at the precipice of another step change in philanthropic capital, and much of it is likely to come from, and point back toward, artificial intelligence.

Philanthropy and private capital are funding what AI could do — the pilots, research and development, and innovation. The bottleneck is what it would do at scale, and the question almost no one is funding is what it should do.

The next wave of philanthropy must invest at the “should” layer, where the market failures are deepest and the stakes highest, because that is what decides whether we build an AI world worth living in.

Philanthropic capital exists precisely to fund the things markets underfund. But knowing philanthropy has a role to play does not tell us what to fund. For that, it helps to separate three questions that run together in most AI-for-good conversations: what AI could do, what it would do, and what it should do.

Level 1: What could work

Most of the momentum and most of the money sits at the frontier of what the technology makes possible. Private capital has largely funded model capability. Philanthropic capital has followed with what one recent analysis called a technocentric bent: emphasizing the transformative potential of AI for complex problems and urging faster diffusion of the technology to ensure low- and middle-income countries are not left behind.

Model capability is improving at light speed and demonstrating what AI could do to transform how social programs are delivered and how nonprofits themselves operate is important. And where raw thinking power has been the bottleneck to scientific progress, we can expect real gains — new vaccines, faster drug discovery, better seed genetics.

Yet there is still a wide gap between what AI models can do in controlled settings and what a health worker in Kaduna or a teacher in Kolkata can actually use them for.

The question the field needs to answer next is not what AI could do in a research setting, but what it would do when introduced into real health systems, real classrooms, and real welfare programs at scale.

Level 2: What would work

Behavior and institutional change is the barrier for new technology. The risk is that the sector fails to apply the lessons of previous waves of innovation; overinvesting in technical deployments and underinvesting in the public goods, user input, trust building, regulation, institutional change management, evaluation infrastructure, and redress systems that make deployments succeed at scale.

Consider three points on the spectrum:

On the private-sector end:ConveGenius in India has reached millions of students with AI-assisted learning, but the traction has come from years of aligning with state-run assessment systems, not from the model itself.
On the public-private partnership side: IDinsight’s work on HEP Assist with community health workers has shown that a promising conversational AI tool is useful only to the extent it fits the workflow, cadence, and referral protocols the workers already navigate.
On the government scale: Rwanda’s rollout of AI in public administration has moved faster than most peer countries in part because the state is unusually integrated — a starting condition few other governments share.

Filtering the “coulds” through “woulds” goes beyond what is technically possible and prompts a focus on the hard work of listening to users, redesigning workflows, and letting the technology bend around the people rather than the other way around.

Level 3: What should work

What could work is a technical question. What would work is an institutional and behavioral one. What should work is a normative question, and often the one that gets asked last, usually when the damage has been done. In public health, we regulated tobacco advertising after two decades of cancer, and trans fats after a cardiovascular disease epidemic. We started asking whether social media was safe for adolescents years after the harm was measurable in the data. We are extraordinarily good at building things, and extraordinarily bad at pacing their deployment against the institutions that should govern them.

Consider the following:

AI systems are already competitive with clinicians on some diagnostic tasks and with teachers on some tutoring tasks. But on whose terms is the AI tool deployed, with what fallback when it fails, with what say for the community it serves, and with what plan for the human workforce it augments or displaces?
Take algorithmic targeting of social protection: AI could plausibly decide who receives a cash transfer or a malnutrition referral. But should it? Under whose oversight? And with what recourse for those wrongly excluded?

Questions like these cannot be answered by builders alone, and they will not answer themselves. Answering the “should” question well is inherently collaborative work: It requires the communities most affected at the design table, and it requires governance structures — public, private, and community — capable of driving accountability and redress where needed. If “AI for good” in global development is going to mean anything, these questions have to be asked before deployments, not after.

The stakes are that we build a world in which AI does remarkable things — genuinely massive, and capable of changing behaviour at scale — and simultaneously not a world any of us would actually want to live in.

Where the money should move

Closing the gap between what AI could, would, and should do is unglamorous, iterative work.

A concrete agenda for philanthropy at the “should” layer is needed. Here is a nonexhaustive list to get the conversation started:

Meaningful participatory design, so the people most affected by a deployment have a say in how it is built and recourse when it fails.
Safety and quality standards and grievance mechanisms for AI tools serving vulnerable users, developed with the communities that use them, with clear escalation paths when a model is out of its depth in a clinical, educational, or welfare decision.
Data governance infrastructure that lets governments and communities decide what happens to the data their citizens generate, rather than defaulting to the vendor’s terms of service.
Institutional integration grants that fund the non-AI work around an AI deployment — the workflow redesign, training, and change management that currently gets smuggled into project budgets as “implementation support” and cut when budgets tighten.
A pooled evaluation fund for AI deployments in real service delivery systems, runningfour-level evaluations (is the model good, is the product good, does anyone use it, does it change lives) continuously as tools are built and adapted.
Regulatory sandboxes and controlled rollouts that let ministries and developers test AI tools together under real conditions — with monitoring, guardrails, and the ability to pause or adapt — so regulation and the tool evolve in step rather than the regulator perpetually catching up

We have spent our careers watching what happens when smart capital, patient implementers, and community voices line up behind a good idea. AI is a bigger idea than most, arriving at a moment of unusual philanthropic ambition, and the questions being asked are the sharpest the sector has heard in a long time. A technology this powerful, aimed with this much care, and matched by this much capital could shift what is possible in health, learning, and livelihoods for hundreds of millions of people.

After a stretch of unprecedented damage to global development funding and to the data and health systems on which so much depends, we have an unusual opportunity to rebuild.

If the alignment holds — “could” becoming “would,” filtered through “should” — the next decade could be the one global development has been waiting for. It is worth every dollar it takes to get right.