Family learning · Suggested Grades 4–7

Can a sorting rule learn the wrong lesson?

Four training cards suggest that large shapes are triangles. Two new cards break the pattern. Try the sorter, count its mistakes and decide which evidence you need next.

Start the paper activity

Why try this in December 2026?

Code.org’s November 17 announcement places this year’s Hour of AI activities alongside Computer Science Education Week, 7–13 December 2026. Families can explore its official activity catalogue through the source link below. This original BananaLearn paper exercise gives you another way to begin a conversation about examples, predictions and mistakes.

You need paper, two coloured pencils and someone to compare answers with. The exercise uses a tiny made-up dataset so you can inspect every example. A parent or teacher can read the instructions aloud if the table feels unfamiliar.

Prepare six shape cards

A · TrainingLarge triangle
B · TrainingLarge triangle
C · TrainingSmall square
D · TrainingSmall square
E · TestLarge square
F · TestSmall triangle
CardDrawSize labelCorrect shape labelUse
ATriangleLargeTriangleTraining
BTriangleLargeTriangleTraining
CSquareSmallSquareTraining
DSquareSmallSquareTraining
ESquareLargeSquareNew test
FTriangleSmallTriangleNew test

Draw triangles with three straight sides and squares with four equal sides and four right angles. Mark the word large or small on each card so everyone uses the same category. The exact dimensions are less important than keeping the assigned labels consistent.

Put E and F face down. Show the learner only the four training cards first. Ask what patterns they notice without telling them which rule to use.

Build and test a rule

Try this rule first: predict triangle for a large card and square for a small card. It matches all four training labels. Before turning over the test cards, write the predicted label for E and F from their size labels alone.

Now reveal the drawings. The rule predicts triangle for the large square E and square for the small triangle F. Both predictions are wrong. The same rule scored 4 out of 4 on the training set and 0 out of 2 on the new test set.

4 / 4Training cards matched

Size and shape were linked in every training example.

0 / 2New test cards matched

The new cards broke that link.

Ask what the rule actually used. It followed size, although the task was to identify shape. Seeing every training example once gave it an easy shortcut that failed when the examples changed.

Try a better feature

Draw a new decision rule using the number of sides. In this six-card activity, predict triangle when there are three straight sides and square when there are four. Apply it to A through F and count the matches.

The new rule gets 6 out of 6 here. Now draw a rectangle that is wider than it is tall and has unequal neighbouring sides. A rule that calls every four-sided shape a square will mislabel it. Improve the check by considering equal side lengths and right angles, or give the sorter a separate category for other shapes.

A test set checks a specific collection of examples. Add examples that make you question the rule. A perfect score on a small, familiar set leaves many cases unexplored.

What this tells us about AI

You wrote the rule yourself in this activity. In machine learning, a system can fit a prediction rule using training examples. Our paper sorter makes the roles of input features, correct labels and new test examples visible before you meet more complex systems.

The input feature in the first rule was size. The label was triangle or square. The prediction was the label the rule produced. Looking at which examples were missing from training helps you understand the errors on the new set.

Real systems can use many features at once, and the effect of a feature can be difficult to inspect. The habit to carry forward is to ask what was measured, where the examples came from and how a prediction was checked. Use a new test set after changing a rule if you want another fresh check.

Six questions to discuss

Write a prediction or explanation before opening each answer. Use the actual cards to show the evidence behind your choice.

1. What feature does the first rule use?

Size. It predicts triangle for large cards and square for small cards, without checking the sides.

2. How many training labels does the size rule match?

Four out of four. A and B are large triangles, while C and D are small squares.

3. What does the size rule predict for the large square E?

Triangle. That prediction is wrong because E is a square.

4. Why does the first rule fail on both test cards?

The training examples linked size and shape in the same way every time. The test cards break that link.

5. Does a four-sided shape always count as a square?

No. A square also needs four equal sides and four right angles. A non-square rectangle challenges the simple side-count rule.

6. After changing the rule, why keep some new cards hidden?

They give you a fresh check. If you repeatedly adjust a rule to fit the same visible cards, those cards give less information about how it will handle unfamiliar examples.

Run a second round together

Ask one learner to create eight new cards. Include large and small versions of triangles, squares and rectangles. Another learner designs a rule using six of them. Keep two hidden for the final test, and agree how to label shapes outside the first two categories.

Before testing, write down what the rule can observe. A learner allowed to see only size cannot use the number of sides. Changing the permitted information changes the problem, so record it alongside the rule.

After the reveal, write one example the rule handled well and one that needs work. Finish by choosing the next card you would collect and saying why it could help. This turns an error into a question about evidence.

For a family record, copy the headings rule, training matches, new test matches and next example to collect. Keep the cards with the record so another person can reproduce the test.

Keep exploring

Sources and reference models

The activities and practice questions on this page are original BananaLearn examples. Use these references for the underlying explanation and confirmed event dates.

Dates checked on 21 November 2026. BananaLearn is an independent learning platform.