I have a new chapter out, co-authored with Danielle McNamara, Gregory Goodwin, and Diego Zapata-Rivera, in the latest volume of Design Recommendations for Intelligent Tutoring Systems (Volume 12: Generative Artificial Intelligence), edited by Anne Sinatra, Vasile Rus, Arthur Graesser, and Paige Lawton.

Danielle was originally invited to contribute to this volume and asked if I’d like to take the lead. I jumped at it. I’m not an ITS expert per se, but I’ve been thinking a lot about what LLMs can and cannot do in educational settings, and this felt like an opportunity to bring that perspective to a community with deep expertise in structured, adaptive learning systems.

The core argument is this: LLMs and Intelligent Tutoring Systems are built on fundamentally different logics. ITSs are precise, consistent, structured. They excel at diagnosing student errors, adapting instruction, and ensuring mastery of well-defined content. LLMs, on the other hand, hallucinate (it’s a feature, not a bug, of how they generate language) and produce variable outputs even when given identical prompts. These are serious problems if you’re trying to use an LLM as a tutor.

But here’s the thing. Rather than seeing this as a fatal flaw, we propose a reconceptualization. LLMs aren’t failed tutors. They’re something else entirely: thought partners, creative collaborators, what we call in the chapter an “intelligent creative buddy.” Their variability, their ability to approach problems from multiple angles, their capacity for open-ended creative discourse… these become assets when the goal is fostering curiosity, lateral thinking, and interdisciplinary exploration.

So instead of replacing ITSs with LLMs (or vice versa), we argue for a synergistic framework. Let the ITS handle the structured knowledge acquisition where precision matters. Let the LLM handle the exploratory, creative, perspective-shifting work where variability is a feature. Together, they offer something neither can provide alone.

The chapter also discusses concrete implementations within the GIFT (Generalized Intelligent Framework for Tutoring) platform and lays out research questions for the field. The full book is freely available online, and I’ll link to just our chapter below.

Reference

Mishra, P., McNamara, D. S., Goodwin, G., & Zapata-Rivera, D. (2026). Large language models and intelligent tutoring systems: Conflicting paradigms and possible solutions. In A. M. Sinatra, V. Rus, A. C. Graesser, & P. M. Lawton (Eds.), Design recommendations for intelligent tutoring systems: Volume 12 – Generative artificial intelligence (pp. 37–45). US Army Combat Capabilities Development Command – Soldier Center.