Building an AI learning tool is becoming easier. Knowing whether it deserves a place in a classroom is not.

Does it support learning? Reflect the curriculum and language learners use? Work within the realities of teachers’ time and school infrastructure? These questions cut across individual models and products—and answering them requires more than any one firm’s technical capabilities.

At the Global AI for Learning Collective Action Agenda launch on 22 September, Mike Fleming, Learning Product & Ecosystem Strategy at Google; Jayna Devani, Education for Countries & Research Partnerships Lead at OpenAI; and Adam Barton, Global Education, Beneficial Deployments at Anthropic, moderated by Gaurav Gupta of Dalberg, explored a practical agenda for collaboration: which foundations would help the whole education ecosystem if they were built, tested and maintained together?

Four themes stood out from the discussion:

1. As AI makes it easier to build learning tools, quality becomes the harder problem
2. Local context has to shape the technology from the beginning
3. Evidence needs to keep pace with a fast-moving field
4. Some parts of the building blocks that many AI education tools need are more useful when they are shared

1. Easier to build. Harder to know what works.

An AI model can draft a lesson plan or assessment in seconds. Teachers and smaller education organizations can increasingly turn those outputs into tools without large engineering teams. But the ability to create a tool tells us little about its educational value.

Having more tools does not mean having good tools, ones that work in practice and help produce positive learning outcomes.” – Adam Barton

As building becomes easier, the challenge shifts towards judging quality: Is the content accurate and pedagogically sound? Does it reflect the curriculum and the learner’s needs? And does using the tool actually improve learning?

Answering these questions requires different kinds of testing. Model benchmarks can assess specific capabilities, such as recognising a misconception or generating an appropriate assessment item. Product evaluations can examine how those capabilities translate into a learning experience. Classroom research can then test whether that experience supports learning in practice.

This is also where individual innovations can become useful to the wider ecosystem. Pratham’s work on verifying AI-generated assessment items and Fab AI’s foundational-literacy benchmarks illustrate how quality-assurance methods developed for specific applications can become reusable resources for others to build on.

These are foundations that many developers and education systems need. Building and sharing them can reduce duplicated effort and give governments, educators and developers a stronger basis for deciding which tools to use and how to improve them.

2. Local context has to shape the technology, not be added afterwards

An AI tool that works well in one education system may be ineffective in another. It may follow the wrong curriculum, fail to support the language learners actually speak, or assume teaching practices and infrastructure that do not exist in the schools where it is deployed. These are not problems that can be solved simply by translating or adapting a finished product. Local education needs must shape how the technology is developed from the outset.

Teachers need to be in the driver’s seat of any use of AI in a classroom.” – Jayna Devani

The same principle applies at system level. Governments need to help define the problems, standards and evidence requirements that shape AI for education, rather than simply choosing between products once they are finished.

This changes how the sector should think about scale.

Scale does not have to mean taking one finished product and distributing it across more systems. It can also mean giving local actors stronger building blocks, including evidence, technical components, curriculum structures, standards and implementation knowledge, so they can adapt or develop solutions around their own needs.

Shared public goods can make that possible. IDinsight’s work on reusable curriculum infrastructure illustrates this approach. By making foundational resources available to different developers and education systems, shared infrastructure can support local adaptation without prescribing a single standardized application.

3. AI creates an evidence-speed problem

An education intervention can take years to evaluate rigorously. In that time, the AI model underpinning it may have changed several times.

To face this challenge, governments need evidence that keeps pace with technological change without sacrificing rigour. This requires both independent research and faster ways of generating and sharing findings.

We need researchers to be independently researching what’s going on, and it’s important that we engage with them on the right infrastructure to do that. We supply datasets, we supply evaluation approaches and similar tools. And that work needs to live outside in the public domain. And that work also needs to be packaged up and shared so that anyone, any researcher can benefit, not just two or three who are working on one individual study.”
– Jayna Devani

Countries, companies and education organizations are already learning how teacher training affects adoption, where products fail to fit local contexts, which safeguards matter and what needs to change during implementation. When these findings remain within individual partnerships, each new deployment starts from a smaller evidence base than it should.

Rapid-cycle evaluation can complement longer-term research by generating evidence while products and implementation approaches are still evolving. The classroom research presented by Fab AI and Google DeepMind, together with their work on reusable evaluation methods, illustrates how these approaches can be shared for others to learn from and adapt.

The public-good opportunity is to make evidence cumulative. Making datasets, evaluation methods and implementation insights more widely available allows evidence to accumulate across products and contexts, rather than remaining fragmented across individual projects.

4. Distinct products can draw on shared foundations

Google, OpenAI and Anthropic will continue to build their own models and education offerings, and pursue unique strategies. Collective action does not require changing that.

The more useful question is which capabilities each organisation needs to develop itself, and which could serve many more learners if developed collectively.

No one actor, even one as large as Google, can do this alone.” – Mike Fleming

The case for collaboration becomes strongest where different actors repeatedly need the same foundational capabilities.

Collective action helps us avoid reinventing the wheel, a problem we constantly run into across education technology. One way to avoid this is to create common goods, like knowledge graphs, that support local building and ownership. Another is to foster partnerships between organizations and teachers so those shared goods get put to use in the field.”
– Adam Barton

Curriculum structures, datasets, benchmarks, safety approaches, evaluation methods and implementation evidence can all create more value when many actors can use and improve them.

That principle extends beyond AI firms. Governments can help determine which shared resources are actually useful to public systems. Researchers and implementers can test whether they work. Funders can support resources where the benefits extend well beyond the organization that creates them.

The public-goods question is therefore not whether everyone should build together. It is which foundations should not need to be rebuilt separately at all.

GAILA’s task is to turn these shared needs into a small number of practical collaborations, with clear contributions, reusable outputs and governments and educators helping shape them from the outset.