Sandra Liu Huang explains how Learning Commons is building open A.I. infrastructure to make classroom tools more reliable and useful.”>
In 2015, Mark Zuckerberg and his wife, Dr. Priscilla Chan, pledged the vast majority of their wealth to launch the Chan Zuckerberg Initiative (CZI), driven by a broad and ambitious goal to “build a more inclusive, just, and healthy future for everyone.” In the decade since, Zuckerberg’s tech empire underwent sweeping overhauls, including a high-profile rebranding to Meta, in response to shifting tech landscapes and a maturing founder’s evolving vision. New technology, specifically the rise of A.I., eventually forced a reckoning for his family philanthropy. Last year, CZI restructured its mission into two independent pillars: Biohub, focused on A.I.-driven scientific and biomedical research; and Learning Commons, an initiative dedicated to building open A.I. infrastructure for K–12 education.
At the helm of Learning Commons is Sandra Liu Huang. A seasoned product manager with stints at Google, Facebook and Quora, Huang joined CZI in 2017 as head of technology before narrowing her focus on education. She faces an intriguing question for product managers: How do you apply Silicon Valley discipline to public-good tech where the ultimate metric isn’t user retention or revenue, but meaningful societal impact?
Today’s schools stand at a chaotic crossroads. Teachers and students are overwhelmed by a fragmented sea of A.I. tools, which threaten to derail learning with shortcuts and hallucinations. In an interview with Observer, Huang broke down what Learning Commons is actually building and how its tools are deployed in real classrooms. More importantly, s
The following transcript has been edited for length and clarity.
Observer: You spent years in product development at Google, Facebook and Quora before leading product at CZI and now presiding over Learning Commons. How is building a “product” at a philanthropic organization fundamentally different than at a tech startup? And how do you define success?
Sandra Liu Huang: The fundamentals of building a great product are actually pretty similar: you have to understand the problem deeply, build alongside the people you’re trying to serve, and keep listening and iterating based on what works. But the key differences for a philanthropic organization typically lie in the mission, success metrics and time horizon.
At Learning Commons, we’re building tools that serve as infrastructure for A.I. in education, for the public good. Our goal is to create open resources that many different developers can use to make their edtech products more impactful, reliable and trustworthy for teaching and learning. Success for us means the broader ecosystem is producing better tools for teachers and students because of our infrastructure.
We’re starting to see that happen. Knowledge Graph is now being used by edtech companies and A.I. systems to connect their products to state academic standards, high-quality curricula and learning science. It powers integrations with tools such as Claude for Teachers and ChatGPT for Teachers. We are investing in making Knowledge Graph richer and bringing more learning science datasets forward.
Our Evaluators are helping developers assess the quality of A.I.-generated educational content. The ultimate measure of success is how tools that leverage us better support educators in creating more consistent, high-quality learning experiences for students.
Give me an example of how Learning Commons’ tools are being used in a real classroom right now.
One example I love is RightOn!, a math platform that helps teachers understand how students think by looking not just at whether they got an answer right or wrong, but at the reasoning and misconceptions behind it.
RightOn! is piloting an A.I.-powered tool called MicroCoach that uses our Knowledge Graph to connect patterns in student mistakes to learning progressions and prerequisite skills. So instead of simply telling a teacher, “Your students are struggling with adding and subtracting decimals,” it can help identify whether the issue is a foundational concept that needs to be revisited or something that can be addressed with targeted feedback before moving on.
That’s the kind of role we want A.I. to play: not replacing a teacher’s judgment, but giving teachers more sophisticated resources to complement instruction. RightOn! describes MicroCoach explicitly as a teacher-in-the-loop tool, and they’re now developing an efficacy framework that looks at both math proficiency and students’ willingness to persist through challenging problems.
Another increasingly accessible example is lesson planning. The Knowledge Graph connections to Claude for Teachers and ChatGPT for Teachers allow educators to create and differentiate instructional materials using A.I. that has access to specific state academic standards and detailed maps of learning progressions, rather than relying only on the general information a model was trained on.
How do you measure whether a tool is actually accelerating student learning versus just making a teacher’s day slightly easier?
Making a teacher’s day easier is valuable; teachers have incredibly demanding jobs, but efficiency alone isn’t the end goal. Ultimately, the goal is to ensure all students have access to rigorous and motivating instruction.
One reason we built Evaluators is that you need rigorous ways to define and measure quality before you can improve it. Our Evaluators assess A.I.-generated educational content against expert-informed criteria and research-backed rubrics. For example, they can measure whether a reading passage is appropriate for a particular grade level, whether its vocabulary and sentence structure are appropriately complex, and what scaffolding a student might need.
We’re still early in this process, and it’s important to be transparent about that. Over time, we also want to understand what happens once these tools reach classrooms.
When a student asks an A.I. for help, a hallucination or incorrect step can completely derail their understanding. How does Learning Commons prevent A.I. from generating mathematically or factually wrong content?
Our approach is to improve both what goes into A.I. systems and ensure we rigorously evaluate what comes out.
Knowledge Graph gives A.I. systems access to trusted, structured educational information — detailed, interconnected maps of state academic standards, learning progressions, high-quality instructional materials, and research about how students learn. That gives models a stronger educational foundation than relying on general training data or the web.
Then Evaluators address the other side of the equation. They allow developers to systematically test A.I.-generated content against trusted educational rubrics and expert-informed criteria, rather than simply assuming that because something sounds plausible, it’s appropriate for a student.
This matters because “accuracy” in education is broader than getting a fact right. A passage can be factually correct but far too difficult for a third grader. We want developers to be able to evaluate across important instructional dimensions of quality consistently and improve their systems over time.
Given how rapidly generative A.I. is expanding in schools, is there any part of teaching or learning that you believe technology should never attempt to automate?
Technology should empower teachers rather than try to replace the human relationships at the heart of learning.
Teachers do things that are much bigger than delivering content. They engage and motivate students. They notice when a student who usually participates has gone quiet. They understand the dynamics of a classroom and know when a child needs to be challenged, encouraged or simply seen. They bring their own passions and energy to inspire and create sparks in students. These relationships matter enormously to learning.
The exciting opportunity for A.I. is to give teachers more capacity to do those things well. If technology can help a teacher differentiate a lesson, identify a student misconception earlier, or spend less time stitching together disconnected materials, that can create more room for the parts of teaching that are deeply human.
Ultimately, students still need to do the hard work of learning. Productive struggle is part of how we build knowledge, judgment and expertise. This struggle is the process of building up the brain, according to neuroscientists. Technology should support that process, not remove every moment of difficulty from it.
Mark Zuckerberg and Priscilla Chan pledged 99 percent of their wealth to philanthropy with a multi-decade horizon. Five to ten years from now, what quantitative or qualitative milestone will prove to you that Learning Commons successfully transformed K–12 education?
Learning Commons is not trying to single-handedly “transform” K–12 education. Education is an enormous ecosystem, and meaningful change happens through educators, researchers, school systems, developers, communities and many others who are working together to bring their expertise to bear.
What I would love to see five or ten years from now is that learning science has become part of the basic infrastructure of education technology—something developers don’t have to reinvent, and teachers don’t have to go searching for.
Today, so much of what we know about how students learn is difficult to access or apply at scale. Academic standards vary across states, curricula sit in different systems, and valuable research can be disconnected from the products educators actually use. We’re trying to create the connective tissue between those things.
Success would mean that whether a teacher is using a tool from a large A.I. company or a small edtech startup, they can expect it to be grounded in high-quality instructional materials, aligned to what students need to learn, and rigorously evaluated for educational quality. Ultimately, we want to see evidence that those tools are contributing to better learning experiences and outcomes for students.
I know Learning Commons focuses on primary and secondary education, but higher education is currently struggling with widespread frustration among both professors and students regarding A.I. What do you see as the biggest challenges of A.I. in higher ed, and could Learning Commons help solve them?
Though we partner with researchers in higher education, our focus is on elementary through secondary schools, so I want to be careful not to prescribe solutions for higher education. But some of the underlying challenges are similar, and the principles for addressing them translate. If A.I. is going to become an ongoing part of learning, we need rigorous, transparent ways to ground these systems in trusted knowledge and evaluate whether their outputs actually support learning. That’s a challenge the entire education sector will have to address.
