Chris Agnew is the managing director of Stanford University’s AI Hub for Education, a part of the SCALE Initiative that translates research on AI in teaching and learning into practical guidance for K-12 education leaders. Since its founding in January 2025, the Hub has built a repository of academic research on AI in education and published “Understanding The Evidence Base on AI in K–12,” which identified key trends in AI research. Agnew also recently co-authored “The Learning Experiences that Matter and AI’s Role” with Stanford colleagues Susanna Loeb and Cristina Barnard Gonzales. FutureEd Policy Analyst Tara Moon recently spoke with Agnew about the fast-changing AI landscape in education.
What is the AI for Education Hub at Stanford’s SCALE Initiative and how did you come to lead the organization?
The goal of the Hub is to explore what’s working and what’s not with AI in education and to study how to use AI to reimagine longstanding education systems to better benefit kids.
I spent two decades in education, almost entirely in non-traditional learning environments, like outdoor- and community-based classrooms and apprenticeships. I left that space feeling frustrated because I knew that immersive, experiential, relevant learning was impactful for kids, but it’s way too expensive to be accessible to all.
When ChatGPT arrived, I was working in ed tech using early generative AI as a formative-assessment tool. This work led me to realize the potential of AI to unlock those impactful learning experiences and to chip away at the barriers that have historically kept them out of reach for most students.
Your recent report, “The Learning Experiences that Matter and AI’s Role,” seeks to reframe how policymakers and educators think about AI in schools. Rather than starting with what AI tools can do, as most conversations about AI do, it asks what schools are trying to accomplish and whether AI can help achieve those goals. What prompted that approach, and what did you find?
It can be very tempting with any new tool—a chalkboard, a smartboard—to start with the shiny object and figure out all the ways it can plug in. But that tools-first approach risks locking us into a system of schooling that was designed more than a century ago. We wanted to flip that thinking around: first decide what we want for students, then ask whether AI can help us get there.
We started with the question, “What’s school for?” We identified 10 skills that research links to long-term success, like academic knowledge, higher-order thinking, social skills, and motivation. We then reviewed research to learn which types of learning develop those capacities. That work consistently pointed to five key experiences: personalized instruction, real-world learning, student agency, enriching discussions, and strong, supportive relationships with adults.
From there, we asked a different question than most conversations about AI ask: if these are learning experiences we know matter, what’s preventing schools from providing them today? Through conversations with students, parents, and school leaders, we identified several historical barriers in schools to those important features of student success: rigid scheduling and staffing structures, narrow accountability systems, insufficient teacher training and support, inflexible curriculum, and uneven access to information for families.
Finally, we explored how AI might help address these constraints.
What did that analysis reveal about AI’s potential? How could it help schools overcome those barriers and make important learning experiences more achievable at scale?
We thought about AI as something that can knit together data, logistics, and expertise across schools in ways that were never operationally possible before, making them much nimbler and more responsive to the needs of kids.
For example, right now, kids get grouped by birthday, not by learning level or need. AI could synthesize assessment data with staffing and room constraints to enable much more dynamic grouping. Kids could move between groups as they demonstrate mastery, instead of staying locked into a fixed schedule set months in advance.
Assessment is another example. Today, we measure a pretty narrow slice of academic knowledge, and we do it infrequently, so results come back too late for teachers to help the students. AI could support much more continuous, formative assessment and help make sense of student data so that teachers can intervene quicker. It could also provide visibility into a student’s process—how they revise their thinking, how they collaborate—which is where skills like perseverance, creativity, and critical thinking show up, and which traditional testing rarely captures.
AI can also support teachers professionally. Simulation tools can allow them to practice things like facilitating an open discussion or responding to a disengaged student, skills that traditional professional development doesn’t build well because they need repeated practice with feedback, not a one-off workshop.
What’s getting in the way of AI realizing that potential?
We live in a capitalist market and schools are free to buy what they want, which is a good thing. But a lot of today’s purchasing decisions reward small improvements in the short term over bigger changes in the long term. AI products are often built to fit into how schools already work because that’s what makes them adoptable and what brings ed tech companies revenue. This creates a gap between what might be possible in the long run and what gets built and used today.
That’s why I think there’s a role for larger systems, like states, to help shape innovation. Instead of individual teachers or schools acting in isolation, states can identify promising uses and put public dollars behind them. That kind of public signal can tell the private market to build toward those solutions and longer-term systems change.
That’s the future-facing view. But shifting to what’s happening now, the AI Hub recently published “Understanding the Evidence Base on AI in K-12 Education,” which analyzed research on AI in education up to November 2025. What were the key findings?
One of the clearest findings was just how limited the evidence still is. There is very little research on AI in early childhood or pre-K, and a lot in post-secondary. And until recently, there was no high-quality causal research on AI use by K-12 students in the U.S.
Then we zoomed in on the strongest available studies and looked at AI in two contexts: students using it and teachers using it.
Academic outcomes improved when students had access to AI, but when it was removed, results varied—in some cases, outcomes got worse, and in one case, they improved. AI can ease students’ mental load and make them feel better about their learning but doesn’t necessarily encourage deeper thinking. Lastly, the design of the tool really matters—purposeful, curriculum-anchored AI that provides step-by-step instruction shows greater promise than open-ended use.
We published the first rigorous research on U.S. K-12 students and AI a couple of weeks ago. We studied 355 first- through fifth-grade students in five after-school programs and two schools across two school districts and found that nearly half of the elementary students never used their AI literacy tutor, even when there was dedicated time in the schedule to do so. Secondly, we found that pairing students with human tutors boosted use of the AI support, but not enough to improve reading achievement.
The results for educators as AI users have been more promising. AI shows potential for reducing time on tasks, and automated instructional feedback can improve both teaching quality and student outcomes. Interestingly, the research suggests AI may be most beneficial to less experienced or lower-rated educators.
As you suggest, it’s clear the evidence on AI in education is still limited. But the technology is already in classrooms and evolving rapidly. Given that gap, if you were advising a state superintendent or district leader, what would you tell them to act on now, and what would you tell them to hold off on until the evidence is stronger?
Assuming the basics are in place (a clear AI policy, strong student data privacy practices, and a baseline of AI literacy), I’d tell leaders to lean into educator use. That’s where the evidence is more promising, and adults can apply professional judgment and discernment in ways students can’t yet.
That means using AI to support teachers’ work in ways that ultimately strengthen the learning experience: freeing up time on lower-value tasks to spend more time with students, analyzing data to surface better information about student learning and respond to their needs. I’d encourage leaders to run experiments and work closely with teachers, learn what’s actually useful, and refine from there.
On student use, go slower, especially with unsupervised or unmitigated AI use. The most promising student use case right now is providing additional practice on a specific skill a student is struggling with and, critically, feeding that data back to a caring adult who can exercise human judgment and intervene.
People compare AI integration to ed tech more broadly, having promised major transformation but largely falling short. This has led to widespread calls to limit screentime in classrooms. What will it take for AI to be different?
The debates about screens and AI in schools are asking the wrong questions because they start with the tool rather than what we’re trying to build for kids. When one-to-one devices entered classrooms, many called it a panacea just because it was technology. That framing missed asking the more important question: what do we want for kids, and what are schools for?
Right now, we’re engaging with AI mostly through chat, but multimodal and voice-based AI is advancing fast. We could soon be in a world where screen time has gone way down but engagement with technology has exploded because kids are interacting with AI through voice rather than a screen.
Instead of the screen or no screen debate, exploring the capacities we want to develop in kids and the learning experiences that best develop those capabilities would be a more productive conversation. If screens are involved, we should be thinking about it not from a lens of screen time, but of screen value.