Last November, Digital Promise launched the
K-12 AI Infrastructure Program to fund datasets, models, and benchmarks with the potential to improve teaching and learning. We invited the public to provide comments through an open Request for Information (RFI) to shape the program’s grantmaking strategy. We received more than 100 responses from researchers, startups, and practitioners, all pointing toward a shared conclusion: To realize the promise of artificial intelligence (AI) in the classroom, we must shift from generic models to
infrastructure grounded in learning sciences and the realities of diverse classrooms.
Challenge Areas to Address
The RFI responses highlighted critical gaps where current public goods (products made available for appropriate technical use at no cost) and datasets fall short.
- Championing learner variability: New datasets must reflect the full spectrum of learners including multilingual learners, students with language impairments, and students who speak with different dialects and accents.
- Student data interoperability: There is an opportunity to leverage untapped diverse data sources such as Student Information Systems (SIS), Learning Management Systems (LMS), and Individualized Education Plans (IEPs) so that data has thicker descriptions of students and settings.
- Teacher support: Teachers often face administrative burdens that reduce instructional time. AI can automate assessment and feedback mechanisms, streamline data analysis across fragmented ecosystems, and translate data into actionable insights—ultimately freeing teachers to focus on instruction and student relationships.
- Prioritizing digital wellbeing: As AI-enabled tools become more integrated into classrooms, protecting student mental health is critical. Future infrastructure should include safeguards against unmonitored synthetic relationships and emotional distress.
- Targeted universalism: Educational interventions fail by focusing solely on closing the disparity gap between learner groups, which often reinforces a deficit framing of specific students. Implementing targeted strategies to address these structural barriers is needed to develop the inclusive infrastructure necessary for all students to thrive.
Infrastructure Needed to Advance K-12 AI
To begin to address these challenges, the community identified specialized public goods the current AI landscape doesn’t provide:
- Multimodal datasets that better reflect diverse students and learning progress: Most AI systems train on generic internet text, without context for how students actually learn. A core priority is building foundational data schemas for interoperability, including machine-readable curriculum context and portable learner profiles that allow a student’s strengths, accessibility needs, and prior knowledge and interventions to travel with them safely. This data must prioritize comprehensive representation to ensure AI supports multilingual learners and students with disabilities.
- Models that integrate learning sciences and effective pedagogy: In addition, there is a call for Automatic Speech Recognition (ASR) models optimized for children’s voices and noisy classroom environments. Respondents identified a need for models that encode learning science principles to better support learning progressions and guide productive struggle. There was interest in domain-specific knowledge tracing models that map content and assessment items to granular standards, learning progressions, and skill taxonomies across U.S. states to support the implementation of High Quality Instructional Materials (HQIM).
- Benchmarks that measure what matters: Respondents pressed for methods that move beyond simple accuracy scores to evaluate how well models support learning. They also emphasized the importance of fairness, safety guardrails such as bias audits to see if a model performs equally well across diverse student groups, and age-appropriateness to test if tools can adapt complexity for a 3rd grader versus a high school senior. Finally, evaluation approaches need to check for reliability, whether models reinforce misconceptions, and predictive validity that accurately predicts student learning gains.
Building Sustainable Public Goods
Infrastructure is more than code and hardware; it’s a socio-technical system. True public goods require prioritizing these three areas:
- Governance and privacy frameworks: Given the sensitive nature of student data, the community identified critical components of secure infrastructure: responsible data sharing, standardized legal frameworks, consent templates, and privacy-enhancing strategies such as synthetic data generation.
- Long-term usability: Public goods require ongoing maintenance, versioning, and evolution as technology and educational needs change. Without long-term stewardship, datasets and models risk becoming obsolete or inaccessible.
- Human expertise: This work requires human judgement and expertise. Some examples shared included: expert teacher annotators to label misconceptions; learning scientists to design benchmarks; and collaborative governance with educators and communities historically left out of the edtech ecosystem.
What’s Next?
The insights from this RFI will directly inform our first Request for Proposals (RFP), which is set to launch in late January 2026. This program will fund the creation of public goods—datasets, models—to ensure that the future of AI in education is equitable, safe, and effective for every learner.
To stay informed on future developments of the K-12 AI Infrastructure Program, sign up for our newsletter.
