From Surface Learning to Deep Understanding: A Grounded AI Tutoring System for Moodle

Synthesis: Ostrowska, Kukla & Majstrak (2026) present an AI tutoring system integrated into the Moodle LMS designed to scaffold students from surface-level fact recall to deep conceptual understanding through adaptive questioning and feedback.

How It Works

The system operates within Moodle's existing infrastructure, using Large Language Models (LLMs)-based tutoring to:

• Detect the depth of student understanding from their responses
• Adaptively scaffold toward deeper comprehension via Scaffolding techniques
• Move students through Formative Assessment cycles that promote conceptual mastery

This grounded approach — embedding AI tutoring in an existing LMS rather than building standalone tools — addresses deployment barriers identified in the Comprehensive Review of Intelligent Tutoring Systems. Many intelligent tutoring systems fail to achieve real-world impact because they require new infrastructure; Moodle integration lowers the adoption threshold.

Connections to the Knowledge Base

The system's focus on deep vs. surface learning connects to Metacognition research — students must recognize when they have only surface understanding. The adaptive approach aligns with Adaptive Learning but emphasizes qualitative shifts in understanding rather than quantitative difficulty adjustment. The Moodle deployment strategy echoes lessons from AICoFe: Implementation and Deployment of an AI-Based Collaborative Feedback System for Higher Education (AICoFe) about integrating AI tools into existing educational workflows.

Unlike The Evidence Base on AI in K-12: A 2026 Review debates about specialized tutors, this system shows how general LLMs can be scaffolded into tutoring roles within familiar platforms. The focus on deep understanding complements The Pedagogy of AI Mistakes: Fostering Higher-Order Thinking work on using AI errors for higher-order thinking.

Open Questions

• How does in-LMS tutoring compare to standalone systems in learning outcomes?
• Can the deep learning scaffolding generalize across disciplines beyond the tested contexts?
• What is the instructor's role when AI handles adaptive questioning within Moodle?

What this means for practice

• Instructors. Keep every AI-generated quiz or study guide in the "Unapproved" state until a teacher reviews it; participants rated the system's resistance to hallucination at 4.44/5 precisely because nothing reached students without human sign-off.
• Instructional designers. Tune retrieval to the discipline instead of using one default: 512-character chunks with a temperature of 0.3 reached a RAGAS faithfulness of 0.97 on STEM material, while humanities material needed 1000-character chunks at 0.1 to hold a context recall of 1.00 and an average faithfulness of 0.98.
• Software developers. Hold retrieval at Top-K = 10 — raising it to 15 gained no recall and significantly worsened context precision through irrelevant-material distraction — and evaluate a Socratic "deep understanding" mode on context recall (0.98) rather than answer relevancy, which it scores low by design.
• Instructors. Ground the assistant strictly in teacher-verified course documents and use the dashboard logs to see which topics students actually query, letting the tutor carry the Formative Assessment workload that 14 out of 18 users named, with the chat interface, as the system's most useful features.

Limitations

• The human evaluation rests on 18 participants from the authors' own university, of whom 83.3% were students and 11.1% instructors — a single-site usability sample, not a representative or multi-institutional one.
• Outcomes are self-reported: a 1–5 Likert survey plus open-ended questions, with an overall recommendation of 4.00/5 and relevance-and-accuracy rating of 4.06/5. There is no control group, no comparison with Moodle without the tutor, and no measure of learning gains.
• The shift from "surface learning" to deep conceptual understanding is the system's design goal, not a tested result: the automated scores (Ragas LLM-as-a-judge faithfulness of 0.97 and 0.98) judge grounding rather than correctness, and the human side measured usability and perceived utility rather than understanding.
• All configuration findings come from one deployment on a single Moodle server using NLP and Machine Learning course notes and three specific models (Gemini Flash 2.0, GPT-4o-mini, Llama 3.1), so the optimal chunk size and temperature may not transfer to other courses or models.

Citation

Ostrowska, A., Kukla, M., Majstrak, G., Opala, J., Pergała, S., Skwarek, J., & Wróblewska, A. (2026). From Surface Learning to Deep Understanding: A Grounded AI Tutoring System for Moodle.