Fall is in the air, school is back in session, and we’re feeling a lot of momentum behind our work to help developers build amazing tools for educators.
In that spirit, I’m excited to share a few announcements. Today, we’re making an early release of our Knowledge Graph, which improves the accuracy of content generated by AI-powered tools, available to edtech developers everywhere. We’re also integrating Knowledge Graph with Claude, Anthropic’s large language model. And we’re expanding early access to tools called Evaluators, which help ensure AI-generated text is accurate, rigorous, consistent, and worthy of teachers’ trust.
We think of these releases as building blocks for the entire edtech community — open public infrastructure to support AI tools that truly reflect how students learn, and that embody our highest aspirations for education.
The problem and the opportunity
We’re deeply optimistic about what’s possible. AI systems can give educators deep insights into every student’s learning journey — their individual challenges, strengths, and motivations — and make it possible to guide young people forward in ways that were never possible before, much less at scale.
We also know there are big challenges left to solve. Fundamentally, AI models are designed to help adults find answers quickly. The goal of education is to teach students how to come up with answers themselves. We think AI systems can play an important role in the process of learning, but they need to be carefully trained and tuned for that purpose.
As it stands, edtech developers are needing to do all the work themselves — which means wrestling with academic standards from all 50 states, summarizing decades of learning science research, and trying to distill that knowledge into their products.
For an individual developer, this work is expensive and time-consuming. For the wider edtech ecosystem, the result is fragmentation. Every tool is built on a different foundation and delivers a different level of quality. Teachers are struggling just to get a handle on the technology that’s out there, let alone using it to meaningfully enhance their practice.
But over the last couple of years, we’ve come to understand these challenges as a transformational opportunity, especially with AI as a tool for tool development itself.
CZI is a philanthropy with a world-class engineering team, deep knowledge of learning science, and strong connections to developers, researchers, and educators. We have a clear understanding of the infrastructure that needs to be built, and a longstanding commitment to making technology accessible and open source.
With that in mind, we’re working to create common, high-quality resources that AI systems can easily read and train on, and a shared and open foundation of quality and rigor for the next generation of education technology.
A navigation system for learning
Knowledge Graph is one of those resources. It’s been in private beta for nearly a year, and today, we’re releasing four machine-readable datasets on GitHub that developers can directly integrate into their tools.
Our early release includes academic standards from all 50 states in four core subjects — English, math, science, and social studies. Another dataset breaks the math standards into smaller skills and concepts, which we call learning components. Other datasets in Knowledge Graph connect the components and standards to one another, so AI systems can understand education as a progression of ideas with certain pathways and prerequisites.
In this way, Knowledge Graph is a bit like the data layer that sits underneath products like Google Maps and Apple Maps. We’re assigning a latitude and a longitude to every skill students need to master, then drawing routes between each skill and the rest. Developers can use that data to create tools to help teachers and students get where they need to go — what we might one day think of as a GPS for different paths to learning.
Bringing Claude to the classroom
That brings me to the next announcement, which is that we’ve built a custom model context protocol (MCP) server that connects Knowledge Graph to Anthropic’s LLM, Claude.
A lot of educators already use Claude to help them develop lesson plans and problem sets. With the integration, Claude’s responses will get a lot more specific — reflecting state academic standards, learning progressions, and learning science research that is embedded into Knowledge Graph.
We’re excited for teachers to try it out. And if you’re a developer, we hope the integration is an inspiring example of what you can do with Knowledge Graph. We designed our MCP server to work with any AI system that supports the protocol, and we look forward to expanding access to more edtech developers in the near future.
Evaluating AI tools so teachers can trust them
Finally, I want to turn toward one of the most common applications for AI in education, which is generating practice exercises for students.
Clearly, there’s a lot of potential to tailor material for students’ strengths and interests — and a lot of potential for AI output to miss the mark. Getting systems to generate material that’s always accurate and rigorous is one of the hardest problems in edtech, which is what led us to build tools called Evaluators. Basically, they’re AI models that assess the output of other AI models.
Our first Evaluators are focused on literacy for students in 3rd and 4th grades. They’ve trained on a dataset we built in partnership with Student Achievement Partners — authors of the gold-standard SCASS rubric — and literacy experts at Achievement Network. The models can read AI-generated text and measure the complexity of the vocabulary and sentence structure. We’ve also built an evaluator that helps developers assess the appropriateness of a text for a particular K-12 grade band, based on its text complexity.
Today, we’re doing an early release of the prompts, logic, and scoring code for these Evaluators under open licenses. In the months ahead, we’ll enhance and expand them to cover other measures of text complexity and more grade levels. We’ll also release Evaluators to assess AI-generated output against other rubrics, from alignment with state academic standards to how motivating the exercise is for students.
Our next chapter
Over the past decade, we’ve learned a lot about the challenges and opportunities in our education system. We’ve said since the beginning that technology isn’t a silver bullet for any of them. But the right tools really do have the potential to transform teachers’ and students’ lives for the better — and that’s never been truer than it is today.
We’re committed to building the core infrastructure to support impactful AI tools in the classroom. As we deepen our partnerships with education technology developers and educators and prepare to move our tools from private beta to general availability of our tools in 2026, the Chan Zuckerberg Initiative’s work in education will now be called Learning Commons — a name that reflects our sharpened focus and role within the larger education ecosystem.
While our name changes, our values remain the same. We will keep working for a future where education and technology unlock student potential and accelerate meaningful progress for all. And to do that, we will continue collaborating across the education ecosystem — co-building the future and the technologies we believe in with teachers, school district leaders, researchers, and developers.
We begin this next chapter with boundless optimism for this movement and for the future of education.
We also hope you’ll join us in this work. You can inquire about our new products and partnership opportunities, and follow along with Learning Common’s work at learningcommons.org.
Sandra Liu Huang
Head of Education and Vice President of Product
Chan Zuckerberg Initiative