Tutoring is not just conversation. It is a system that requires:
– goal formulation,
– learner modeling,
– reliable assessment,
– diagnosis of misconceptions,
– planning of outcomes and next steps,
– and targeted remediation.

These are not primarily linguistic processes — they are logical and organizational processes.
Modern AI systems like Khanmigo are strong in associative tasks: pattern recognition, language generation, and conversational fluency. But they lack what is essential for tutoring: explicit structure, transparency, and goal-driven reasoning. This limitation is not accidental. It reflects a deeper property of intelligence itself.

Human intelligence, as shaped by evolution, combines two tightly integrated components:
– an associative (neural) system — fast, flexible, generative;
– and a symbolic/logical system — structured, interpretable, and goal-oriented.

Current AI systems largely replicate the first component, while the second remains underdeveloped in their architecture. Without this second component, we inevitably observe the same failure symptoms described in the original post:
– passive interactions,
– weak engagement,
– lack of measurable learning impact,
– and difficulty moving from conversation to understanding, knowledge, skills, and capability.

In this sense, Khanmigo did not “fail” — it exposed a missing layer required for effective AI tutoring.
The field of Intelligent Tutoring Systems has spent decades developing symbolic and logical models of tutoring — including learner modeling, diagnostic reasoning, and adaptive planning. However, these models have historically been difficult to author, scale, and maintain.

What is new today is that neural AI can finally address this bottleneck. In our work at iTutorSoft, we have been exploring a unified approach:
– using symbolic, mathematically grounded models to define the structure of tutoring,
– and neural AI to support the authoring, conversation, and instantiation of these models at scale.

This combination enables tutoring systems that are:
– transparent and interpretable,
– goal-driven and adaptive,
– responsible and controllable,
– and engaging and intellectually challenging for learners.

From this perspective, the key question is not whether AI can tutor, but: What kind of AI architecture is required for tutoring to work? I believe this remains an open Learning Engineering problem — and a highly promising one.

We would be very interested in connecting with others exploring hybrid approaches that combine neural and symbolic intelligence for real tutoring systems.

— Vlad Goodkovsky
@itutorsoft.com

The post What is wrong with Khanmigo? first appeared on Intelligent Tutoring Agents.