A.I. Can Transform Education Research—If the U.S. Invests Now
The most exciting example of AI’s potential in education isn’t a chatbot or digital tutor–it comes from a large language model named Centaur. Recently published in Nature, Centaur is trained on over 10 million human decisions across 160 classic cognitive psychology tasks. After training, the model could predict what people would do better than existing cognitive models and also generalize to novel tasks it hadn’t seen before.
Centaur wasn’t built for education, but it hints at what could be possible. If a large language model can simulate adult reasoning across tasks it wasn’t trained on, could similar models help researchers understand how students learn in school? Could they forecast how different types of learners respond to different instructional approaches? Of course, modeling children’s development adds layers of complexity–developmental stages, emotional regulation, social learning dynamics–but these are precisely the kinds of high-risk, high-reward questions federal research agencies were designed to pursue.
Today, discovering what works in education is slow, expensive, and fragmented. Determining whether a new curriculum, instructional model, or assessment improves student outcomes can take years. Results are often context-dependent, and when innovations fall short, it’s difficult to know whether the problem lies in the intervention itself, the study design, or the school environment. And long feedback cycles make it costly and onerous for researchers to learn and iterate.
AI offers a potential new pathway. Rather than relying entirely on trials in classrooms, researchers could simulate learning processes, model the effects of interventions, and accelerate discoveries in learning science. Simulated learners, powered by AI, could help derisk the education research and development process, allowing researchers to identify what’s most promising before investing years and millions of dollars in field studies.
The implications are significant. Testing a new learning intervention today involves months of recruitment, teacher training, and data collection—only to find that many innovations yield minimal effects. Despite growing philanthropic investments, most education innovations still struggle to scale. The Science of Reading offers a sobering example. Decades of research supported systematic phonics instruction, yet widespread adoption only followed a national reading crisis, a viral podcast, and sweeping state legislation. If AI-powered models could help surface effective strategies earlier and more reliably, we might avoid these lost decades.
Critically, progress will depend on access and experimentation. In a promising step, OpenAI recently announced open-weight models that are designed to be “best in class.” Moves like this could dramatically lower the barriers for education researchers, enabling them to build and adapt models tailored to specific learning contexts.
To seize this opportunity, the federal government must act proactively. A national AI strategy for education should include robust investment in basic research at the intersection of artificial intelligence, developmental science, and learning. Agencies like the Department of Education and the National Science Foundation are well positioned to lead this work.
Key questions to explore include: Can models like Centaur be adapted to reflect the minds of young learners? What behavioral and learning patterns emerge across simulated learner populations, and how accurately do they mirror real classrooms? Could large-scale simulations uncover game-changing instructional strategies that traditional research methods miss?
These questions aren’t abstract, and their answers could shape how we design curriculum, empower teachers, and support all types of learners. And they could help the United States lead globally in the science of learning at a time when education and AI are receiving significant public attention.
The recently announced Presidential AI Challenge is an encouraging start, inviting young people and educators to build AI-powered solutions. But classroom apps and tutoring tools are only the surface. What’s needed now is a serious national commitment to the scientific infrastructure that will allow AI to advance our understanding of how children learn. Without that investment, the promise of AI in education may remain confined to consumer products and incremental gains—rather than unlocking the breakthroughs students and educators deserve.
The moment is here. The U.S. can lead in transforming how we learn what works in education—but only if it acts with urgency. Federal investment in AI-powered education research isn’t just an option; it’s the foundation for building a smarter future for all learners.
