A Public Good Approach to AI Tutoring
Digital Promise has opened a request for proposals offering up to 8 million dollars to develop an open source AI model capable of supporting one to one math tutoring in United States K-12 classrooms. The initiative, run under the organisation's K-12 AI Infrastructure Program and titled the Open Source AI Model for Tutoring, is managed by the Gates Foundation, which handles proposal review and award monitoring. Proposals are due by July 31, and the project is expected to run for 30 to 36 months, with work beginning around November.
The framing is what makes this notable. Rather than fund another proprietary tutoring app, Digital Promise wants an open model that any district, vendor or researcher can inspect, adapt and deploy. The stated goals are to improve student motivation, engagement, metacognition and math learning, the outcomes that decades of research tie to effective one to one tutoring. By insisting on open source, the program treats the underlying tutoring intelligence as shared infrastructure, a public good to be built once and reused widely, rather than a moat for a single company.
The Money Behind the Model
The 8 million dollar tutoring grant sits within a larger 26 million dollar Digital Promise program that will issue awards over four years to support openly shared datasets, models, benchmarks and other digital public goods for AI in education. That structure signals a deliberate strategy. Instead of scattering small grants across many pilots, the funders are concentrating capital on the foundational pieces that individual districts and startups cannot afford to build alone, betting that shared infrastructure will unlock a wave of downstream products and research.
Only a single award is expected from the tutoring RFP, which raises the stakes for applicants and concentrates the bet. Bryan Richardson, a senior program officer at the Gates Foundation, described the ambition as developing the best AI tutoring model using cutting edge methods and applying learning science principles. The phrasing matters. This is not a call for the flashiest chatbot. It is a call for a model grounded in what research actually shows about how students learn mathematics, which is a harder and more valuable target.
Who Can Win, and Why the Bar Is High
The eligibility requirements reveal how seriously the funders take the difficulty of the problem. Teams must bring machine learning expertise, direct K-12 classroom experience, a learning science background and EdTech product partnerships, with at least one major tutoring provider committed at the point of submission. That combination is demanding by design. It rules out pure research groups with no path to the classroom and pure product teams with no grounding in pedagogy, forcing applicants to assemble genuinely interdisciplinary coalitions.
The insistence on a committed tutoring partner is the sharpest signal. Too many educational AI projects produce impressive demonstrations that never reach a real classroom, dying in the gap between prototype and deployment. By requiring a delivery partner up front, Digital Promise is trying to close that gap before the work begins, ensuring that whatever gets built has a realistic route to students. It is a lesson learned from years of edtech pilots that dazzled in the lab and vanished in practice.
The Math Tutoring Problem Is Worth Solving
The focus on math is not arbitrary. One to one tutoring is among the most reliably effective interventions in all of education research, yet it is expensive and impossible to scale with human tutors alone. Math, with its clear structure and verifiable answers, is both a domain where students struggle at scale and one where an AI tutor can plausibly reason about correctness and guide a student through steps. If any subject is a fair test of whether AI tutoring can deliver on its promise, mathematics is it.
The caution from the field is that early evidence is mixed. Research has found that students under exam pressure sometimes resist pure Socratic dialogue and reach for answer first shortcuts, and that what actually repairs their reasoning is a controlled reveal combined with worked examples and step linked grounding. That nuance is exactly why the RFP demands learning science expertise. Building a tutor that helps rather than merely answers requires understanding how students actually behave under pressure, not an idealised model of how they should.
The Adoption Gap
Even a superb open model faces the chasm between availability and use. Districts are cautious, budgets are strained and teachers are wary of tools that add work or undermine their judgement. An open source tutor that no district deploys helps no student, which is why the requirement for a committed delivery partner is so pointed. Building the model is necessary but not sufficient. Getting it into classrooms, integrated with existing systems and genuinely trusted by teachers, is the harder and far less glamorous half of the job, and it is where most educational technology efforts quietly fail.
History here is sobering. Education technology is littered with well funded tools that dazzled in pilots and never scaled, defeated by procurement friction, poor integration or teacher scepticism. The programs that succeed tend to treat adoption as a design constraint from the outset, shaping the product around the realities of the classroom rather than expecting the classroom to adapt to the product. If Digital Promise wants its investment to matter, the winning team will need to obsess over that last mile as much as over the model's benchmark scores, because the benchmark is not what changes a child's understanding of mathematics.
Our Read
We see Digital Promise's approach as a corrective to a market drifting toward proprietary lock in. If the most capable tutoring models are owned by a handful of vendors, districts become dependent on pricing and roadmaps they do not control, and the benefits accrue unevenly to those who can pay. An open source foundation, if it is genuinely good, changes that dynamic by giving every district and every startup a credible starting point they can build on and adapt to local needs.
The risk is that open source alone does not guarantee quality or adoption, and a single 8 million dollar award is a modest sum against the scale of the problem. Success will depend on whether the winning team can produce a model that teachers trust and students actually learn from, and on whether the broader ecosystem picks it up. But the strategy is sound. Funding the shared layer as a public good, with learning science baked in and a delivery partner committed, is a more durable bet than subsidising one more app.
