Who Pays for the Foundations Everyone Builds On?

Concentrated costs, widely shared benefits and a less visible path to children’s learning make essential foundations for AI in education hard to fund.

This blog reflects on the philanthropy panel at GAILA’s Collective Action Agenda launch in New York on 22 September, featuring Benjamin Piper, Director of Global Education at the Gates Foundation; Annabel Gerry, Director for Human Development at FCDO; Denis Mizne, CEO of Lemann Foundation; and Frank van Cappelle, Global Lead for Digital Education at UNICEF, moderated by Amy Klement of Imaginable Futures.

UNICEF’s 2026 report highlights evidence from 48 low- and middle-income countries showing that over 90% of children were below expected levels in basic reading skills after three years of schooling (Crawford, Michael, et. al., 2025). AI could help address constraints in assessment, accessible content and teaching support, but its value depends on whether tools work for learners’ languages and contexts, whether they can be evaluated reliably and are safe to use.

This requires shared foundations: reusable datasets, benchmarks, evaluation methods, standards and other resources that multiple organisations can build on.

These resources create value across the education ecosystem. Yet their costs sit with whoever develops and maintains them, while their benefits spread across many organisations and users.

Their impact on children is also indirect. A better dataset may improve a tool; a stronger benchmark may guide a ministry’s choice; that tool must then be used effectively in a classroom before it improves learning. The distance between the original investment and the child’s experience makes the outcome harder to observe, measure and attribute to any one grant.

That combination can leave essential foundations underfunded, even when their collective value is substantial.

The panel’s five takeaways suggest a pathway for investment: start with education needs, fund missing shared inputs, establish common ways to judge quality, translate shared standards into product decisions, and support the coordination and stewardship needed for sustained uptake.

The upshot is that alongside funding promising products, philanthropy can also help build the shared foundations that make many products safer, more effective and easier for governments to evaluate.

1. Start with Education Problems, Not with AI Products

The investment case begins with common constraints: a teacher who cannot assess every child’s reading regularly, a learner who lacks accessible materials, a ministry that cannot judge which tools are worth adopting.

Frank van Cappelle of UNICEF described how AI can accelerate content conversion and adaptation, while human review and integration into education systems remain essential. A 2026 peer-reviewed study of Shiksha Copilot in Karnataka found that government-school teachers reported saving an average of two hours per week preparing curriculum-aligned lessons, particularly for Kannada-language materials (Dennison et al., 2026).

For funders, the question is practical: which learning constraint could AI help address, and what shared foundations would make that improvement possible?

Government and educator participation matters from the outset. Their priorities should shape what gets built, tested and funded, providing the connection between upstream investment and what changes for learners.

2. Some of the Most Important Inputs will be Underprovided

In general, organizations have weak incentives to individually build the resources that matter most for AI in learning.

Children’s speech data illustrates this challenge. AI-based literacy tools need to recognise how children speak as they develop reading fluency, yet most speech-recognition systems are trained predominantly on adult speech. Research has also focused heavily on English, with few publicly available datasets capturing children’s speech in other languages (Graave et al., 2024; Jacobs et al., 2025).

Developing representative datasets across languages and contexts is costly, while their benefits extend across multiple AI tools and research programmes. This makes it difficult for any single developer to justify the investment, even when the collective value is substantial.

Philanthropy can help close this gap by funding shared resources that individual developers are unlikely to create independently. The same applies to curriculum datasets, accessible content and technical standards that reduce duplication and enable more locally relevant AI solutions.

3. Governments Need Common Ways to Judge Quality

Even when promising AI tools exist, education systems need reliable ways to determine whether they are safe, effective and suitable for their learners.

This is particularly important for governments making decisions about which solutions to adopt, procure or scale. Without credible evidence and comparable evaluation methods, it can be difficult to distinguish between products that demonstrate meaningful educational value and those that make compelling claims without sufficient validation.

Shared benchmarks and evaluation frameworks can help establish common expectations for learning effectiveness, child safety, accessibility and performance across different languages and contexts.

But these resources need to be designed around the realities of education systems. A benchmark developed for one language, curriculum or learning environment may not translate directly to another. Governments, educators and local researchers therefore have an important role in defining what quality means and ensuring that evaluation methods reflect the needs of their learners.

Common approaches can also inform procurement criteria, quality-assurance processes and the development of education-specific AI policies. They provide a basis for more informed decisions while creating clearer expectations for developers.

The objective is to create credible, adaptable methods for assessing quality, not to impose a single definition of a good AI product across countries.

4. Public Goods Need to be Coupled with Responsible Product Decisions

Shared infrastructure can strengthen product development, but companies still make consequential choices about how their tools behave.

Denis Mizne cautioned that earlier technologies arrived with promises to democratise access to knowledge. Access did not automatically translate into better learning. AI raises that challenge again, with additional questions about children’s safety, privacy and the learning behaviours products encourage.

His comparison with “airplane mode” made the point tangible: widely understood safety expectations can become features built directly into technology through collective action. Education needs that kind of influence over product design, rather than leaving every school to assemble protections after deployment.

Governments, educators and child-safety experts can help define expectations for age-appropriate safeguards and designs that support learning. Governments can reinforce these through education policies, regulations and, where needed, legislative safeguards. Technology companies must then incorporate those expectations into their products.

Shared standards, evidence and evaluation methods can make those expectations clearer and more actionable. Their impact depends on adoption: a standard creates value when it changes how products are designed, selected or used.

Collective investment should therefore support both the development of common expectations and the work needed to put them into practice.

5. Philanthropy and Development Partners can Fund Shared Foundations and Help their Value Spread

For philanthropy, supporting AI in education may require looking beyond individual products to invest in resources that improve the wider ecosystem. “One of the things I am realising is really important for funders in this moment is doubling down on shared public goods. Benchmarks, evaluations, datasets and shared tools are what the whole field builds on, and they are where we need to lean in.” Benjamin Piper of the Gates Foundation pointed out.

This shift is already reflected in efforts to develop shared resources for AI-enabled learning. Curriculum infrastructure developed through IDinsight can become useful to other builders. Components of Pratham’s assessment work can be opened for others to test and adapt. Benchmarks and evaluation methods developed through Fab AI and Google DeepMind can inform decisions beyond the original research partnership. Imaginable Futures’ support for GAILA’s Secretariat similarly reflects the importance of investing in coordination capacity.

Philanthropic foundations and bilateral development partners can play complementary roles in building these shared foundations. Philanthropy can provide flexible, early-stage funding to develop and validate public goods alongside coordination, maintenance and support for uptake, while bilateral partners can help align investments with national education priorities and support their adoption across countries. Governments and educators can define needs, test usefulness and shape the policies, safeguards and public-system decisions governing adoption. AI firms can contribute technical capabilities and incorporate shared standards and evidence into product development.

Coordinating these contributions is itself an important part of the work. As Annabel Gerry of FCDO reflected:

“We want to be part of ensuring that AI contributes to safer, more equitable solutions for better learning outcomes. […] This is the kind of thing that we can share through GAILA. The opportunity to collaborate and to coordinate, bringing together the different settings and contexts.”

Through GAILA, funders, governments, researchers, education organisations and AI firms can identify priority gaps, coordinate complementary investments and connect public goods with the institutions best placed to develop and use them.

These investments may not always produce the most visible product demonstrations. But they can improve the conditions under which many subsequent products are built, evaluated and adopted.