Module 6: AI in EdTech
Education is where AI has the most underappreciated potential and the most overhyped deployments simultaneously. The gap between "AI that shows students different content based on quiz scores" and "AI that builds a complete model of what each student knows and does not know, then sequences learning to maximize long-term retention" is enormous - and most EdTech products are much closer to the first than the second.
This module focuses on the second. The real technical problems in educational AI - knowledge tracing, learning curve estimation, forgetting models, item response theory - are fascinating and largely unsolved. The engineering to put them in production at scale is equally interesting.
Why EdTech AI Is Different
The ground truth is learning, not engagement. Most EdTech products optimize for time-on-platform or completion rates because those are measurable. What you actually want to optimize is long-term learning retention, which requires testing students days or weeks after instruction. This creates a massive delay between model output and label signal.
Individual differences are extreme. Two students can have identical prior knowledge and learn at 5x different rates. A model that performs well on average may perform poorly for specific learning styles, reading levels, or prior knowledge states. Personalization is not a feature - it is the core value proposition.
The data is small by ML standards. A large EdTech company might have 1 million students, each completing 100-500 exercises. That sounds like a lot until you realize that the knowledge tracing model needs to model each student's knowledge state across 500+ distinct knowledge components. The data is both sparse and longitudinal.
Algorithmic fairness has high stakes. An adaptive system that systematically underestimates students from certain demographic groups and routes them to remedial content creates lasting harm. Bias testing and fairness metrics are not optional.
Module Architecture
Lessons in This Module
| # | Lesson | Key Concept |
|---|---|---|
| 1 | Adaptive Learning Systems | Mastery learning, zone of proximal development, sequencing |
| 2 | AI-Powered Assessment | Automated grading, IRT, constructed response scoring |
| 3 | Content Generation for Education | LLMs for problem generation, quality filtering |
| 4 | Knowledge Tracing Models | BKT, DKT, SAKT - modeling what students know |
| 5 | Student Performance Prediction | Early warning systems, dropout prediction, at-risk identification |
| 6 | Personalized Tutoring AI | Socratic dialogue, scaffolding, feedback generation |
| 7 | NLP for Educational Content | Readability, concept extraction, curriculum alignment |
| 8 | Ethics and AI in Education | Algorithmic fairness, surveillance concerns, data governance |
Key Concepts You Will Master
- Knowledge tracing - estimating a student's probability of mastery for each skill over time
- Deep Knowledge Tracing (DKT) - using LSTMs to model knowledge state from interaction sequences
- Item Response Theory (IRT) - calibrating the difficulty and discriminability of assessment items
- Spaced repetition algorithms - scheduling review to maximize long-term retention
- Automated essay scoring - using BERT-based models to score free-text responses
- Learning analytics - building dashboards and early warning systems from learning event data
Prerequisites
- Sequences and Time Series
- LLM Prompt Engineering
- Basic Python and statistics