What is Infrastructure for AI in K-12 Education?

The rapid advancements in artificial intelligence (AI) are creating opportunities for innovators to explore how they can transform and improve K-12 education. Yet educational challenges are more difficult than they initially look. Today’s powerful AI models are missing many kinds of specific educational mechanisms that will be necessary to address learners and teachers’ strengths and needs. Many innovators struggle with similar challenges in applying these technologies, and there are promising approaches being developed that are not yet widely available.

Education needs AI infrastructure that closes the gap between general purpose productivity tools and education-specific learning tools—learning-sciences-based building blocks that can accelerate and deepen innovation. To have an impact on future educational products and services, this infrastructure must be available as a digital public good.

Overall, the rapid progress in AI builds on three kinds of infrastructure: data, models, and benchmarks. Data is the core input to machine learning and leads to developing AI models, evaluating them, and debating what is biased or missing. Models serve as the underlying components that solve specific problems well and are used to assemble real AI services and products. Benchmarks are the beacons that signal the direction forward in the work of improving AI. Rapid progress in AI occurs in an ecosystem where data, benchmarks and models are broadly shared, discussed, and re-used.

Consequently, in the K-12 AI Infrastructure Program, we are focused on these three kinds of infrastructure in the context of education:

1. AI Datasets: Collections of data—such as transcripts, videos, instructional materials, or student work—curated and structured specifically to train, fine-tune, support, or evaluate AI systems and AI models toward K-12 education use cases. They are collected responsibly and ensure that student privacy is respected, while also enabling good actors to evaluate if they are meeting target benchmarks and serving students.
2. AI Models: Machine learning algorithms that process K-12 educational data and perform tasks specific to evidence-based teaching and learning, such as identifying effective instructional practices or generating real-time feedback for students. By sharing models, innovators, researchers, developers, and others can build on the successful research of others.
3. AI Benchmarks and Evaluators: Standardized tools, tasks, metrics, and methods used to evaluate how well AI models (or AI-powered education technologies) perform in education-specific contexts, based on metrics like accuracy, fairness, or relevance to learning goals. These metrics enable us to track both how well solutions work and improvements over time to ensure that AI is safe, reliable, and accurate. For example, in the area of reading, innovators rely on benchmarks that measure how good automated speech recognition tools are at accurately hearing what children say as they read aloud. Automated Speech Recognition (ASR) benchmarks—like all educational benchmarks—can be updated as progress occurs in the field, or in this case, as more information is collected on the characteristics of diverse readers’ speech.

Further, we are focused on building blocks that can be adopted broadly, helping many companies, non-profits, researchers, and educators advance their development of AI products and services. To facilitate broad adoption, the building blocks will be licensed for re-use as digital public goods, available for appropriate use at no cost. We will ensure that they adhere to high standards for safety, privacy, and fairness. Aligned with these issues, setting a high bar for data governance will also be important.

To stimulate brainstorming of what could be possible, we highlight several outside projects that are also building AI infrastructure:

• The National Tutoring Observatory (NTO), led by Cornell University, is creating a large, well-annotated dataset that will offer researchers and developers reliable and secure data to study teaching practices that support student motivation, engagement, and learning in math, science, and other subjects.
• The Learning Commons at the Chan Zuckerberg Initiative has released “literacy evaluators.” These evaluators assess different dimensions of literacy to ensure key characteristics of text are grade-level appropriate.
• The Mind Wandering database provides 20 datasets that identify student mind wandering, based on eye gaze coordinates collected while students participate in a variety of tasks. This dataset includes students from diverse family backgrounds and is fully de-identified, opening up research for people without the resources and laboratory conditions needed to conduct these experiments.

To aid in brainstorming the future of AI infrastructure, we suggest a few guiding questions:

• In what ways are existing AI tools missing the mark in education?
• How can we translate the best of the learning sciences and educators’ wisdom of practice into “machine readable” formats and algorithmic solutions, and thereby strengthen applications of AI in education?
• What existing high quality digital resources could be transformed into public goods, while protecting safety, privacy, and fairness?

In late 2025, we solicited public comments through an open request for information to shape our consideration of funding tracks. In early 2026, we will announce specific tracks or themes for AI infrastructure funding. To stay informed on future developments of the K-12 AI Infrastructure Program, sign up for our newsletter.