After learning lagged for some K-12 students in virtual classrooms during the pandemic, students and their parents turned to online tutoring platforms to help fill the gaps.
U.S. school districts did, too, investing more than $700 million in tutoring programs. This shift to virtual, one-on-one learning on various online platforms produced untapped, granular data on effective teaching practices that would have otherwise been impractical to collect at scale.
A Cornell-led collaborative research team has received a nearly $5 million grant from the Gates Foundation and the Chan Zuckerberg Initiative to leverage artificial intelligence and transform that data into insights that can accelerate the science of teaching and learning.
The team - led by Rene Kizilcec, associate professor of information science in the Cornell Ann S. Bowers College of Computing and Information Science - is establishing the National Tutoring Observatory(NTO), a collaboration with tutoring platforms, school districts, educators and researchers to responsibly compile virtual tutoring data at scale. In turn, they hope to create the world's largest open-access dataset on effective teaching practices.
"We have large datasets that help us understand many important behaviors. Teaching is not one of them," said Kizilcec, who directs Cornell's Future of Learning Lab and studies behavioral, psychological and computational aspects of technology in education. "We do not currently have large-scale, machine-readable data about what teachers are doing right."
Along with Kizilcec at Cornell, researchers from the Massachusetts Institute of Technology, Carnegie Mellon University and RAND (a nonprofit research organization) will be involved in launching NTO and building its "Million Tutoring Moves" dataset. The goal of NTO is to offer researchers and developers reliable data to study teaching practices that support student motivation, engagement and learning in math, science and other subjects. The data generated through NTO will help build better AI-powered tutoring tools for student use, Kizilcec said.
"A lot of companies are experimenting with AI tools that are attempting to be teachers or tutors, but these tools generally aren't trained on any teaching or tutoring data; they're trained on data from all over the internet," Kizilcec said. "We need high-quality data to train algorithms to align with the practices of expert teachers."
What makes good teachers, the team contends, is their ability to go off script when they see a student struggling or losing interest. These "improvisational moves" - such as how best to respond to a student's incorrect answer, or when to offer a word of encouragement - are critical to student learning. Yet, until now, these nuanced "teaching moves" were practically impossible to examine at a large scale, Kizilcec said.
But virtual learning companies are already logging millions of minutes' worth of tutoring sessions on their platforms for quality assurance and internal research purposes. By partnering with companies such as Carnegie Learning, Eedi, PLUS, Third Space Learning and UPchieve, NTO leadership will responsibly curate platform data, including: virtual tutoring video; audio and chats; classroom recordings; and teaching simulations.
They'll then use AI to turn these recordings into text that a computer can understand. The goal is to use this data to understand how teachers can best support student learning.
"Some tutoring companies have developed advanced tools that anonymize and process their tutoring data with remarkable efficiency. The National Tutoring Observatory will build on these early advances and scale them up for researchers and edtech developers," Kizilcec said. "The consortium will be able to answer questions such as what works in teaching and for whom - questions that require large quantities of data to answer reliably."
K-12 educators and AI experts who serve on advisory boards for data privacy, access policy and scientific advancement will oversee the NTO, Kizilcec said. It's this collaborative approach involving educators, school districts and researchers that excites Kizilcec, particularly amid a movement in education research that recognizes the importance of using large-scale data for insights.
"Teachers play an enormously important role in classrooms, and we don't know enough about exactly what they do, especially in one-on-one interactions with learners," he said. "We can learn so much from the best of their efforts."
Along with Kizilcec, the leadership team includes: Ken Koedinger, professor of human-computer interaction and psychology at Carnegie Mellon University, director of LearnLab and Simon Initiative, and co-founder of Carnegie Learning; Justin Reich, associate professor of comparative media studies and writing at MIT and director of MIT's Teaching Systems Lab; Rachel Slama, director of the labor and workforce development program and a senior policy researcher at RAND; and Doug Pietrzak, CEO of Freshcognate.
Louis DiPietro is a writer for the Cornell Ann S. Bowers College of Computing and Information Science.
