How Learner-Generated Examples Work | Create a Case That Proves You Understand the Rule

eduKateSG Learning Node Series · 0051

Seeing an example can make a definition feel clear. Creating a valid example from scratch asks whether the definition has become usable.

Learner-generated examples move the responsibility for exemplification from the textbook or teacher to the learner. Instead of receiving a case and being told what it demonstrates, the learner must construct a case that satisfies the concept, explain why it qualifies and often modify it to explore where the concept stops.

This can expose the shape of understanding with remarkable precision—but only when example generation is designed as a disciplined task rather than an invitation to guess.

Quick Read: Examples as a Test of the Learner’s Model

An instructor-provided example shows a learner one valid case.
A learner-generated example asks the learner to build a valid case from the defining properties.
The task can reveal which features the learner treats as essential, accidental or interchangeable.
In mathematics education, learner-generated examples have a substantial theoretical and classroom research tradition, especially around concept formation and “example spaces.”
Empirical results are mixed. Some studies show productive conceptual work; others find little advantage over classification tasks, and one communication study found instructor-provided examples produced better recall and application.
The method is therefore not “generation beats explanation.” The useful question is when generation adds diagnostic or constructive work that supplied examples cannot.
The strongest tasks constrain the example, require justification, include non-examples or boundary cases, and provide feedback.

A generated example is valuable when the learner has to use the rule to build the case—not merely invent something that sounds related.

The Difference Between Recognising and Constructing

Suppose a student learns that an even integer is divisible by two.

Recognition is easy to test: Is 18 even? Is 21 even?

Generation asks for more: “Give an even integer greater than 100 that is not a multiple of 4.”

Now the learner must coordinate several constraints at once. The answer cannot be retrieved as a memorised classroom example. It has to be constructed from the properties.

That constructive act is where learner-generated examples become informative.

What Is an Example Space?

Anne Watson and John Mason used the idea of an example space to describe the collection of examples a learner can access or construct for a concept.

A learner’s first example space can be narrow. If every triangle seen in childhood is upright and roughly equilateral, “triangle” may quietly inherit properties that are not part of the definition.

As learners generate, transform and compare examples, the space can widen. They begin to separate necessary properties from familiar appearances.

Learning is not just adding more examples to memory. It can involve reorganising the space of possible cases.

Why One Example Is Dangerous

One example can carry accidental features.

If the teacher introduces a function using one smooth upward-sloping graph, a learner may attach “smooth” or “increasing” to the concept even when those features are irrelevant.

Generating several examples under changing constraints forces the learner to ask:

What must stay the same?
What can change?
How far can it change?
What would make the example stop belonging?

The Research Tradition in Mathematics

Watson and Mason developed learner-generated examples as a way to make mathematical learning more constructive. Watson and Shipman later described learners using example generation to encounter new concepts, arguing that constructing and reflecting on cases can support shifts in understanding through induction, abduction and deduction.

This tradition treats examples not as decoration after the definition but as objects learners can manipulate to explore the definition’s structure.

The Evidence Also Tells Us to Be Careful

More recent experimental work adds useful restraint. George Kinnear and colleagues compared example-generation tasks with classification tasks in introductory university mathematics. Across two studies, they found little difference in later concept-question performance between the two task types.

A later study of mathematical example generation found that the interface itself could change learners’ strategies. Students working through e-assessment and paper-based versions of similar tasks did not necessarily generate examples in the same way, and some digital responses reflected trial-and-error rather than the analytic strategies designers hoped to elicit.

Outside mathematics, an experiment with 348 students learning communication concepts found that instructor-provided examples led to better recall and application than student-generated examples, partly through greater lesson clarity.

Taken together, the lesson is strong: asking learners to generate examples is not inherently superior to showing them excellent examples. Its value depends on prior knowledge, task design, feedback and what capability the lesson is trying to build.

Four Jobs a Generated Example Can Perform

Construction: build a case from defining properties.
Diagnosis: reveal what the learner thinks the properties mean.
Boundary testing: change features until the case almost stops qualifying.
Generalisation: compare several generated cases to infer what remains invariant.

Mathematics: Build a Fraction With Constraints

Instead of asking “What is a proper fraction?”, ask:

Create three proper fractions greater than one-half. Make one reducible and two already in simplest form.

The learner must coordinate numerator, denominator, magnitude and simplification. A correct response demonstrates more than recognition.

Then ask for a fraction that changes category after one small change. Now the learner explores the boundary.

Additional Mathematics: Generate a Function With a Property

Ask for a function that has two stationary points, or a quadratic whose discriminant is zero, or a trigonometric equation with exactly two solutions in a stated interval.

The learner must reason backward from the property to a valid mathematical object.

This reverse construction is powerful diagnostically. A student who can solve a supplied discriminant question may still struggle to construct coefficients that produce a chosen discriminant state.

Science: Generate an Example and a Near Non-Example

Suppose the target idea is a physical change.

Ask the learner to generate:

one clear physical change;
one clear chemical change;
one case that students commonly misclassify;
one explanation of the feature that separates the categories.

The fourth requirement matters. Without justification, the task can degrade into examples pulled from memory rather than examples built from a model.

English: Generate a Sentence That Forces the Grammar

For subject–verb agreement, do not only ask learners to correct mistakes. Ask them to create a sentence in which the subject and verb are separated by a distracting phrase but the agreement remains correct.

Then ask for an incorrect version that a hurried reader might accept.

Construction forces the learner to understand what makes the difficulty difficult.

Vocabulary: Generate Context, Not Just a Sentence

A sentence containing a word can be grammatically correct yet semantically weak.

For ambivalent, ask the learner to generate a situation in which a person genuinely holds competing feelings, then explain why uncertain or indifferent would be less precise.

The example now tests meaning, contrast and usage together.

History: Generate a Case That Tests the Category

If students are learning about causes, consequences or turning points, ask them to generate an event that would qualify as a turning point under an explicit definition—and an event that was important but should not qualify.

The task forces criteria into the open. Historical labels stop being merely names attached after the fact.

Learner-Generated Examples Versus Concrete Examples

Concrete Examples owns the use of specific cases to make abstract ideas intelligible.

Learner-generated examples own a different action: the learner constructs the case rather than only studying it.

A teacher can provide a brilliant concrete example before asking learners to generate their own. The two are complements, not competitors.

Learner-Generated Examples Versus Concept Boundaries

Concept Boundaries owns how examples and near non-examples reveal where a concept stops.

This page owns who constructs the examples and what cognitive work construction requires.

A learner-generated boundary task combines both: “Create the smallest change that turns your valid example into a non-example.”

Learner-Generated Examples Versus Analogical Encoding

Analogical Encoding asks learners to compare cases to discover shared relational structure.

Example generation can feed that process. Generate two cases that look different but satisfy the same rule, then compare them to identify what remained invariant.

Learner-Generated Examples Versus the Generation Effect

The Generation Effect is a memory phenomenon involving generated target information.

Learner-generated examples are a richer task. The learner is constructing an instance that satisfies conceptual constraints. Any memory benefit is only part of the educational value.

The Example-Generation Ladder

Level 1 — Copy and vary: change one surface feature while preserving validity.
Level 2 — Fill a constraint: generate any example satisfying one stated condition.
Level 3 — Multiple constraints: satisfy two or three conditions simultaneously.
Level 4 — Contrast: generate an example and a non-example.
Level 5 — Minimal pair: create two cases differing in only one important feature.
Level 6 — Boundary case: create a case as close as possible to the category edge.
Level 7 — Unusual valid case: break superficial stereotypes while preserving the definition.
Level 8 — General family: describe how to generate infinitely many valid examples.

Why the Constraint Matters

“Give an example” is often too weak.

A learner may retrieve the same memorised classroom example everyone knows. No construction has occurred.

Add a constraint and the task changes:

Give an example that satisfies the definition but does not look like the textbook examples.

Now the learner must separate essential features from familiar ones.

Why Justification Matters

A generated example can be correct by accident.

Require the learner to state why the case qualifies, which property is essential and what small change would invalidate it.

This turns a lucky answer into inspectable reasoning.

Failure Mode 1: Retrieval Masquerading as Generation

The learner repeats the exact example used by the teacher.

Repair: impose a condition that excludes the original example or require a different representation.

Failure Mode 2: Trial-and-Error Gaming

In digital tasks, learners may submit guesses repeatedly until the system accepts one. Recent e-assessment research shows why interface design matters here.

Repair: ask for a prediction and justification before submission, limit blind retries or require the learner to describe the construction strategy.

Failure Mode 3: Generation Before Clarity

If the concept is still poorly understood, asking for examples can multiply misconceptions.

The communication-study evidence is a useful warning: well-chosen instructor examples can create clarity that unguided learner generation does not.

Repair: teach the concept and provide discriminating examples first, then transfer some of the example-building work to the learner.

Failure Mode 4: Easy Examples Only

Learners often generate central, familiar cases. This can leave misconceptions at the boundary untouched.

Repair: require unusual, extreme, minimal or near-boundary examples.

Failure Mode 5: Personal Relevance Replaces Conceptual Accuracy

Teachers sometimes ask for personal examples because personal relevance feels engaging. But a vivid personal story can still be a poor instance of the target concept.

Repair: accuracy first, relevance second. The example must pass the definition before it earns points for familiarity.

Cross-Domain Lens: Unit Tests

Software developers write test cases to check whether code behaves correctly under specific conditions. Good tests include ordinary cases, boundary cases and failure cases.

Learner-generated examples can function similarly for a concept. The learner builds cases that test whether the mental rule accepts what it should accept and rejects what it should reject.

The analogy is especially useful because it changes the question from “Can I repeat the definition?” to “Can my definition survive a test suite?”

Cross-Domain Lens: Materials Testing

Engineers do not learn a material only from one normal load. They test it across conditions and near limits.

Concept knowledge also becomes more trustworthy when it survives varied cases. A learner who can create a normal example, an unusual example and a near non-example has begun to stress-test the category.

A 25-Minute Home-Study Routine

Minutes 0–4: write the definition or rule from memory.
Minutes 4–8: generate one ordinary valid example.
Minutes 8–12: generate a second valid example that looks different.
Minutes 12–16: generate one near non-example and explain the violated property.
Minutes 16–20: create a boundary case or unusual valid example.
Minutes 20–23: compare all cases and state the invariant features.
Minutes 23–25: check against a trusted source and repair any false cases.

A Teacher Protocol

Provide enough initial clarity that learners know the target definition or relation.
Choose a constraint that forces construction rather than recall.
Require justification.
Ask for more than one example.
Include non-examples and boundary cases.
Have peers inspect each other’s examples against explicit criteria.
Use incorrect generated examples diagnostically rather than simply rejecting them.
End by generalising what the examples collectively reveal.

A Student Protocol

State the rule before generating.
List the properties your example must satisfy.
Construct the case deliberately.
Test every property.
Explain why it qualifies.
Change one feature and test whether it still qualifies.
Create a case that nearly fails.
Compare your examples and identify what cannot be removed.

A Tutor Protocol for Three Learners

Give all three learners the same concept but different generation constraints.

Student A creates a standard example.
Student B creates an unusual valid example.
Student C creates a near non-example.
Each student must justify the case without naming its category first.
The other two classify and explain their decision.
The group then writes one rule that separates all three cases.

The activity creates a compact laboratory for concept boundaries without requiring a large class.

Using AI Without Outsourcing the Example Space

AI is excellent at producing examples. If the educational goal is for the learner to generate them, that strength can become a problem.

Use AI after the learner’s attempt:

ask whether the example satisfies the definition;
ask for the smallest change that would invalidate it;
ask AI to generate a counterexample to the learner’s over-broad rule;
compare the learner’s example space with the AI’s suggestions;
require the learner to explain which examples are actually useful and why.

The learner should remain the constructor and judge, not merely the recipient of an infinite example stream.

What Generated Examples Diagnose

Definition ownership: can the learner use the rule rather than recite it?
Constraint coordination: can multiple properties be satisfied at once?
Stereotype dependence: does the learner confuse familiar appearance with necessary structure?
Boundary knowledge: can the learner tell when one small change crosses the category line?
Representation flexibility: can the concept be instantiated in more than one form?
Generality: can the learner describe a family of examples rather than isolated cases?

Canonical Owner Boundaries

This page owns learner-generated examples: the learner’s construction of cases that satisfy, violate or probe a concept’s defining properties.

Concrete Examples owns how specific cases make abstraction understandable.
Concept Boundaries owns the use of examples and near non-examples to reveal where a category stops.
Analogical Encoding owns comparison across cases to extract shared relational structure.
The Generation Effect owns memory effects from self-generating target information.
This page owns the constructive act of building the examples themselves as a learning and diagnostic task.

Evidence and Limits

The research literature supports learner-generated examples as a rich way to investigate and sometimes develop conceptual understanding, especially in mathematics. It also warns against assuming that generation automatically beats classification or instructor-provided examples.

The method is likely to be most useful when learners have enough prior structure to generate meaningfully, when the task imposes informative constraints, when justification exposes the reasoning, and when feedback prevents incorrect examples from becoming rehearsed misconceptions.

In some lessons, a carefully chosen teacher example is more efficient and clearer. The educational question is not who should own every example. It is when transferring the work of example construction to the learner reveals something worth learning.

The Return Path

Return to the definition.

Do not ask the learner to repeat it.

Ask for a case that satisfies it.

Then one that looks different.

Then one that almost qualifies.

Then ask what single property decides the boundary.

The learner is no longer merely looking at the teacher’s examples. The learner is operating the concept.

Learner-generated examples work when constructing a case forces the learner to expose, test and refine the rule inside their own head.

Use This Tomorrow

Choose one definition from a subject you are studying. Create four cases: a normal example, an unusual valid example, a near non-example and a clear non-example. For each, write the property that makes the classification correct. Then change exactly one feature of your unusual example and predict whether it crosses the boundary. Check the result against a trusted source or teacher.

Research and Further Reading

eduKateSG Learning Node Series · 0051 of the continuing series. Previous: 0050 — How Problem Posing Works. Continue through the Study & Learning Methods Hub and the wider eduKateSG Learning Hubs.