Can data teach a computer to make decisions—and what happens when the data is incomplete or unfair?
60–75 minutesIntermediate challengeInteractive Data LabNo student login
1
Start: human or machine?
8 mins
Silent prediction:
A school wants a computer to identify whether a piece of rubbish is likely to be metal or plastic. It records conductivity, flexibility and weight.
Which measurements might be useful features?
What answer should be stored as the label?
Could the computer ever make a wrong decision? Explain why.
Assessment checkpoint A · Diagnose prior understandingAsk pupils to hold up 1, 2 or 3 fingers for the feature they think is most useful. Select two pupils with different answers to justify their choices. Listen for the misconception that a computer “understands” the object rather than detecting patterns in recorded values.
2
What you will learn
ExplainUse feature, label, training data and model accurately to explain supervised learning.
ApplyConstruct and test a classification rule using numerical data and calculate its accuracy.
EvaluateJudge whether evidence is sufficient, identify bias and propose a justified improvement.
Success criteria
DevelopingI can identify features and labels and run a supplied model.
SecureI can improve a model using evidence from training and test results.
MasteryI can explain why accuracy alone may hide bias and justify better data collection.
3
Learn: how labelled data becomes a model
12 mins
In supervised machine learning, examples are supplied with correct answers. An algorithm searches for patterns connecting the input features to those labels. The pattern it produces is called a model.
Feature = recorded propertyLabel = correct categoryTraining data = examples used to learnModel = learned decision patternTest data = unseen examples
Object
Conductivity
Flexibility
Weight
Label
Drink can
9
3
4
Metal
Food tray
1
8
2
Plastic
Foil
8
9
1
Metal
Important: the model does not know what “metal” means. It finds a mathematical pattern in the examples it receives.
Teacher modelling questions
Why might weight alone be unreliable?
Why is foil a useful example?
What could happen if every metal object in the training set were heavy?
4
Data Lab: build a classifier
25 mins
Your model may use one feature and a threshold. For example: if conductivity is at least 6, predict Metal; otherwise predict Plastic. Train on the visible examples, then test the rule on data it has never seen.
SupportBegin with conductivity and a threshold of 6. Describe every error.
CoreCompare at least three rules and record the best training and test accuracy.
StretchFind a rule that scores well on training data but worse on test data. Explain why this matters.
Recycling Classifier Lab
Rule: predict Metal when the chosen feature is at least the threshold.
Choose a rule, then evaluate it against the training examples.
Assessment checkpoint B · Hinge questionIf a rule scores 100% on its training examples, does that prove it will classify every new object correctly? Pupils must answer no and justify this using the phrase unseen test data before progressing.
5
Think deeper: accuracy is not the whole story
15 mins
A camera system is trained to recognise recyclable objects. Almost every training photograph was taken on a white table in bright light. The system is then installed beside outdoor bins.
AnalyseWhat pattern might the model accidentally learn?
PredictWhich outdoor objects or conditions could cause errors?
ImproveSpecify new examples that should be added to the dataset.
Bias and representation
A dataset is biased when its examples systematically under-represent important situations or groups. A high overall accuracy can conceal poor performance on a smaller category.
Challenge the shortcut: “The model is 90% accurate, therefore it is fair and safe.” Write a counterargument that distinguishes accuracy from reliability across different conditions.
Assessment checkpoint C · Evidence of evaluationA secure response identifies the missing conditions. A strong response explains the resulting pattern of errors. A mastery response proposes a balanced test set and a separate performance measure for each condition.
6
Independent challenge: design the evidence
A school wants an AI system to identify whether plants in its greenhouse need water. You are responsible for the data—not for writing the final program.
Dataset designerChoose at least four useful features and define the label precisely.
Quality controllerDescribe how you would collect varied, accurate and balanced examples.
Critical evaluatorDesign a test and state when a human must override the system.
Required justification: explain one danger of using too little data, one source of bias, and one reason a highly accurate model might still be unsuitable.
7
Review: prove what you understand
10 mins
Final written exit ticket
In no more than 60 words, explain how data can teach a computer to classify something—and why a human must still evaluate the result. Use at least four key terms from this lesson.
0 of 60 words
8
Create my lesson record
Your record gathers the work you completed on this page into one clear document for saving, printing or handing in.
Important: your work is saved temporarily in this browser on this computer only. It will not follow you to another computer. Create and save your lesson record before leaving this page or changing computer.
Want the complete resource library?
Learn IT with Mr. C is developing sequenced Computing resources for Years 7–11 and preparing for the confirmed direction of the 2028 curriculum.
