Cognitive Load Theory & Educational AI

The Science Behind Personalized Worked Examples

Andrew Ellis

18 November, 2025

Learning Objectives

By the end of this workshop, you will understand:

  1. Cognitive Load Theory: How working memory limits affect learning
  2. The Worked Example Effect: Why studying solutions helps novices
  3. The Personalisation Effect: How familiar contexts reduce cognitive load
  4. How AI can scale these principles in education

Two Memory Systems

Working Memory

Limited capacity (approximately 4 items)

Volatile (information decays quickly)

Bottleneck for learning

Long-Term Memory

Unlimited capacity

Permanent storage via schemas

Automatic access without effort

The Challenge: Get information from working memory into long-term memory without overload

Three Types of Cognitive Load

Type Description Effect on Learning Example
Intrinsic Complexity inherent to the material Necessary for learning Learning calculus is inherently complex
Extraneous Load from poor design or irrelevant info Hinders learning Confusing layouts, unfamiliar contexts
Germane Effort directed at learning Helps learning Studying patterns, making connections

Goal: Minimize extraneous load, manage intrinsic load, maximize germane load

The Worked Example Effect

What the Research Shows

“Novice learners who are given worked examples to study perform better on subsequent tests than learners who are required to solve the equivalent problems themselves.”

— NSW Centre for Education Statistics and Evaluation (2017)

Evidence:

  • Effect size: 0.52 (medium to large)
  • Strongest for novices
  • Diminishes as expertise grows
  • Works across domains (math, programming, science)

Why It Works

Problem Solving

High cognitive load:

  • Search for solution
  • Test approaches
  • Monitor progress
  • Learn concept

Little capacity left for learning

Worked Examples

Lower cognitive load:

  • Solution provided
  • Focus on understanding
  • See patterns clearly
  • Build schemas efficiently

More capacity for learning

The Personalisation Effect

Generic vs. Personalised Examples

Generic

Calculate the correlation
between Variable X and Y.

X: [10, 15, 20, 25, 30]
Y: [2, 4, 6, 8, 10]

Higher extraneous load:

  • Abstract variables
  • No connection to prior knowledge
  • No personal relevance

Personalised

You love cycling. Calculate
correlation between training
hours and speed:

Hours: [10, 15, 20, 25, 30]
Speed: [25, 27, 29, 31, 33]

Lower extraneous load:

  • Familiar context
  • Connects to existing knowledge
  • Personally relevant

Why Personalisation Works

“Familiar contexts require less cognitive effort to process, reducing extraneous cognitive load.”

— Cordova & Lepper (1996)

Benefits:

  • Reduces extraneous cognitive load
  • Increases motivation and engagement
  • Better learning outcomes
  • Particularly effective with worked examples

How AI Enables Scaling

The Challenge of Personalisation

Creating personalised worked examples manually:

  • 30 students × 10 concepts = 300 unique examples
  • 15 minutes per example = 75 hours
  • Impossible to do at scale

The AI Solution

AI can generate personalised worked examples:

  • Create unique examples on demand
  • Incorporate personal contexts naturally
  • Maintain quality with structured outputs
  • Scale to unlimited students

Today’s Goal: Build a tool that generates personalised worked examples using PydanticAI

Key Takeaways

Summary

Three core principles:

  1. Worked examples reduce cognitive load for novices
    • Effect size: 0.52
    • More effective than problem solving for beginners
    • Allows focus on understanding patterns
  2. Personalisation reduces extraneous load
    • Familiar contexts easier to process
    • Increases motivation and engagement
    • Better learning outcomes
  3. AI enables scaling of personalised worked examples
    • Generate unique examples on demand
    • Maintain quality with structured outputs
    • Impossible to do manually at scale

Next Steps

Let’s See It in Action

Next: Demonstration of the complete application

  • 3 domains (Programming, Health Sciences, Agronomy)
  • Live personalised example generation
  • Structured outputs with PydanticAI

Then you’ll build your own version!