A major new grant program is targeting a specific problem with AI tutors: they talk too much, give away answers, and fail to recognize when students are struggling. The Gates Foundation is backing an $8 million initiative to develop open-source AI infrastructure designed to make math tutoring systems behave more like skilled human educators.

What's Wrong With Today's AI Tutors?

Digital Promise's K-12 AI Infrastructure Program released a Request for Proposals on June 1, 2026, seeking one team to build what's called the Open Source AI Model for Tutoring, or EDU AI. The initiative addresses a fundamental mismatch between how current AI tutors operate and how effective human tutors actually teach.

The problem is well-documented. Today's AI tutors often exhibit what researchers call a "helpful assistant" bias, meaning they rush to solve problems for students rather than guiding them through the thinking process. They also tend to produce overly long responses, miss signs that a student is losing motivation, and fail to distinguish between a genuine misconception and a simple arithmetic mistake. For math learning especially, this is a serious limitation because productive struggle is essential to building understanding.

"The grant is focused on a specific problem: current AI tutors are not yet behaving enough like effective human tutors," noted Bryan Richardson, Senior Program Officer for R&D Infrastructure and AI at the Gates Foundation.

Bryan Richardson, Senior Program Officer, Gates Foundation

The RFP identifies several specific weaknesses in existing frontier AI models when applied to tutoring. These include weak awareness of what students already know, safety risks when conversations shift context unexpectedly, and a tendency to give away answers instead of supporting productive struggle.

How Will This New AI Tutoring Project Work?

• Timeline: Applications are due by July 31, 2026, with work expected to begin in November 2026 and continue for 30 to 36 months.
• Scope: The funded team will develop open-source model weights, training code, datasets, evaluation tools, and a reference implementation that other developers and school districts can use.
• Licensing: All outputs must be released under open licenses, with code and models using Apache 2.0 or similar permissive licenses, ensuring the infrastructure remains publicly accessible.
• Team Requirements: Applicants must bring together machine learning engineers, K-12 classroom practitioners, learning scientists, and partnerships with at least one major tutoring EdTech provider.

The RFP puts significant weight on safety because the model will interact directly with children. Applicants must provide detailed plans for student data protection, de-identification, anonymization, and compliance with the Family Educational Rights and Privacy Act (FERPA), the Children's Online Privacy Protection Act (COPPA), and relevant state laws. A safety and bias mitigation plan for student-facing deployment is a non-negotiable requirement.

Why Does This Matter for Schools?

The initiative sits within Digital Promise's wider $26 million K-12 AI Infrastructure Program, a multi-year effort launched with partners including Learning Data Insights, DrivenData, Georgetown University's Massive Data Institute, and Catalyst at Penn GSE. The broader program aims to address gaps that make current generative AI less useful for K-12 learning, including limited education-specific datasets, weak support for learner variability, and difficulty applying learning science principles inside AI tools.

The funded work is explicitly not intended to create a closed commercial tutoring product. Instead, the grant is designed to create public infrastructure that researchers, developers, school districts, curriculum teams, and AI model developers can use and build upon. This approach mirrors the open-source software model, where code and tools are freely available for the broader community.

The project is expected to coordinate with an AI tutoring benchmark under development by AllenAI and the Stanford Scale initiative, scheduled for release in late summer 2026. Applicants are also encouraged to leverage existing work including the National Tutoring Observatory, the AI Math Tutoring Benchmark and Open Dataset Project, and the Math Misconceptions Data Challenge.

What Qualifications Do Applicants Need?

The Gates Foundation has set specific eligibility criteria to ensure the funded team has both technical depth and real-world education experience. Lead organizations must have at least one peer-reviewed publication before May 8, 2026, and a documented record of contributing digital public goods, such as public datasets, open-source models, evaluation artifacts, or comparable infrastructure.

Critically, applicants must demonstrate meaningful prior deployment or evaluation using real student or user data. Proof-of-concept work or projects using only synthetic data will not meet the minimum scale requirement. This ensures the winning team has already tested their approaches in actual classroom settings with real learners.

The RFP also requires applicants to explain how their model will be tested for validity, reliability, fairness, safety, efficacy, and cost. They must also show how teachers, school administrators, and other education stakeholders will provide feedback during development. This emphasis on stakeholder input reflects a growing recognition that AI tools designed without educator input often miss practical classroom needs.

The deadline for proposals is July 31, 2026, with the grant period estimated at 30 to 36 months beginning in November 2026. This represents a significant investment in building AI infrastructure that prioritizes learning science over raw computational power, signaling a shift in how the education sector approaches AI development.