A3 How Machines Learn
How do we know if an AI is any good?
Lesson plans ready, classifier tool in development
- Stage
- Creators, ages 9 to 11
- Lessons
- 6 × 45 minutes (4.5 hours)
- Ends with
- A mini AI project and a one-page “how it works and where it fails”
- Before you start
- A2, or experience with block coding and simple data in a spreadsheet
Lesson plans ready, classifier tool in development. Lessons 1, 2, 5 and 6 work with cards and discussion today. Lessons 3 and 4 use the classifier tool, which is still being built; until then, build the classifier from sorted drawing cards and a tally sheet.
For children to do on their own. Every lesson has a step-by-step do-it-yourself version with a self-check. Open the kids’ version.
By the end, children can say
- I can explain the difference between a program with rules and a model that learns.
- I can explain why a model must be tested on examples it has not seen.
- I can build and test a classifier with my own data.
- I can calculate accuracy and describe what kinds of mistakes a model makes.
- I can use simple checks to judge whether something online is reliable or computer-made.
- I can present my AI project, say where it can fail and describe how I would use AI responsibly.
About this course
Format. Classifier tool plus simple tally tables. Some lessons work with cards alone.
Tools. The classifier tool and a spreadsheet or table for tallying results. The tool is in development.
Key words. rule, pattern, training data, test data, accuracy, error, bias, source.
Curriculum tags.
- Computer science: data and algorithms
- Digital literacy: evaluating content and sources
The current national curriculum does not mention AI by name, so each course is tagged to the strands it supports.
Lesson plans
Open a lesson to see the steps. Printing this page opens them all.
Lesson 1: Rules Versus LearningI can explain the difference between a program with rules and a model that learns.
- Starter (5 min). Write rules for telling a cat from a dog. Where do they break?
- Teach (10 min). A rule-based program checks fixed rules; a learning model finds patterns in many examples.
- Activity (22 min). Sort mystery cards using made-up rules, then using examples, and compare which works better when something new appears.
- Reflect (8 min). Which approach would you choose for spam email? Why?
Lesson 2: Training and Testing DataI can explain why a model must be tested on examples it has not seen.
- Starter (5 min). A student who memorises answers versus one who understands: how can we tell?
- Teach (10 min). Split the data: a training set to learn from, a test set to check.
- Activity (22 min). Split a card deck into training and test sets, train a partner, then test only on the unseen cards.
- Reflect (8 min). What would go wrong if we tested on the same cards we trained on?
The “Teach the machine” activity on the Try page uses exactly this idea: the machine is tested on the creatures you did not teach it.
Lesson 3: Build a ClassifierI can build and test a classifier with my own data.
- Recap (5 min). Training and test data.
- Demo (8 min). Choose a question (for example, “is this a drawing of a happy or a sad face?”) and gather examples from the class.
- Build (25 min). Train the classifier with at least 20 examples per class, hold back 5 for testing, and record the results.
- Reflect (7 min). Which examples confused it?
Lesson 4: Accuracy and MistakesI can calculate accuracy and describe what kinds of mistakes a model makes.
- Starter (5 min). If 8 out of 10 test items are right, what is the accuracy? Is 80% good enough for a medical test?
- Teach (10 min). Accuracy as a fraction; two kinds of mistake (a false alarm and a miss).
- Activity (22 min). Record a tally of right, false alarm and miss for the classifier from Lesson 3, then improve the training set and retest.
- Reflect (8 min). Which mistake is worse for a smoke alarm, a false alarm or a miss?
Lesson 5: Real, Fake or Made-up?I can use simple checks to judge whether something online is reliable or computer-made.
- Starter (5 min). Two pictures, one real and one computer-made. Can you tell?
- Teach (10 min). Computers can make convincing pictures, text and voices. Check the source, look for the original, compare with other sources, and ask a trusted adult.
- Activity (22 min). Evidence detectives: examine pre-checked examples with a checklist (who made it, why, where else is it, does it feel too good or strange to be true?).
- Reflect (8 min). Three rules for checking before you share.
Teachers choose the examples in advance so children do not meet harmful or distressing content.
Lesson 6: Using AI WiselyI can present my AI project, say where it can fail and describe how I would use AI responsibly.
- Plan (8 min). A one-page template: what it does, how it was trained, how I tested it, where it fails.
- Build (25 min). Complete the project and the one-pager.
- Share (10 min). Peer review: “One thing it does well, one way it could fail.”
- Reflect (7 min). Badge and a class “wise AI user” charter.
For grown-ups
- Ask: “How do we know if an AI is any good?” and listen for “test it on new examples”.
- Look at a picture or video together and play “real, edited or computer-made?” Check the source before deciding.
- Ask which mistake matters more for a smoke alarm: a false alarm or a miss.
In every lesson: the “check it” habit
Who made this? What was it trained on? Where else can I check?