A sixth grader sits down with a chatbot to talk through a science quiz. In most classrooms, that conversation unfolds the way it does with any general-purpose AI tool: friendly, fluent, and mostly improvised, with no real reasoning behind why it says what it says. Clayton Cohn wants to change that and develop learning theory-driven frameworks that support interactions between AI and students.
Cohn, a former Ph.D. student in computer science at Vanderbilt University’s Institute for Software Integrated Systems, worked with Gautam Biswas, a Cornelius Vanderbilt Professor of Computer Science and Computer Engineering and senior research scientist at the institute, to build classroom AI that’s grounded in long-established ideas about how people actually learn, instead of the generic (and often, ad hoc) chatbot behavior found in most educational AI tools today.
“I really liked talking to it because it could easily simplify my scores and it helped me stay on topic, which is really important,” one of the students who worked with Cohn’s system said.
That system, called Inquizzitor, came out of Cohn’s first paper this year, presented at the premier AAAI 2026 conference in Singapore. The idea is simple to state, even if the engineering isn’t: instead of letting an AI tutor improvise feedback, it builds in rules based on decades of research on how students learn and how good teachers respond. Inquizzitor doesn’t just guess at a useful-sounding answer — it scores a student’s quiz against real evidence and the rubrics provided by the classroom teacher, then bases its feedback on that evidence, the way a human tutor reasons from a graded paper rather than a hunch.
Cohn and his collaborators tested Inquizzitor with 104 sixth-grade students at a Nashville public middle school over a three-week Earth science unit. Its scoring matched or beat human graders on two of three assessments and came close on the third—a sign, Cohn’s team says, that an AI tool built this way can grade as reliably as a person while still tailoring its feedback to each student.
A second paper, presented at the International Conference on AI in Education (AIED) 2026 in Seoul, S. Korea, advances the idea further. It outlines a three-step process for the AI to follow: gather evidence on how a student is progressing in their learning task, use a theoretical framework to determine what that student needs next, and then provide feedback based on that decision. An example theory that the agent uses to formulate its feedback includes Lev
Vygotsky’s “Zone of Proximal Development,” which suggests providing feedback that helps the learner overcome the gap between what they have already accomplished and what they can achieve with a little guidance and collaboration.
The team built this approach into Copa (Collaborative Peer Agent), an AI “study partner” used alongside a hands-on, computer-based environment where students learn physics. They tested it with students, working in pairs, at a Nashville high school over six weeks. As students got better at the material, Copa noticeably changed how it helped them — shifting from asking questions that checked their understanding to offering more direct suggestions. Not only that, students leaned on Copa less over time as they grew more confident on their own, which is exactly the pattern researchers hope to see: good instruction should eventually make itself unnecessary.
Across both studies, involving more than 250 students total, Cohn’s research points to the same conclusion: AI feedback built on learning research can adapt to each student while still being easy to explain and check. That’s the key difference from an ordinary chatbot — every response can be traced back to a specific, understandable reason, not just a pattern the AI picked up on its own.
For Cohn, who is now a Postdoctoral Research Associate at the National Tutoring Observatory at Cornell University, the bigger goal is closing the gap between how natural today’s AI tutors sound and how well-tested older, more rigorously studied tutoring systems were. The point isn’t AI that merely sounds like a good teacher — it’s AI whose decisions can be explained, tested, and trusted the way good teaching can be.