For a parent or older sibling with a child in Classes 6–8. You do not need to know anything about AI. Each session takes about 10 minutes and tells you what to ask, when to wait and which hint to give.
Predict: your child says what will happen before pressing anything.
Try: they use the activity on the screen while you watch.
Explain: they tell you why it happened. You ask; you do not tell.
New example: one fresh question to check the idea, not the screen.
Away from the screen: a short paper or card activity.
A wrong guess your child can explain is good learning. If they are stuck, wait about ten seconds, then give the hint. Stop whenever they are tired; the next session can wait. Every activity stays open for free exploration too.
Session 1: Rule or learning?
The idea: some machines follow a fixed rule; others use patterns learned from examples. The word “smart” alone does not tell you which.
Remember. Ask: “When an advert calls something ‘smart’, what do you think it means?” Accept any answer for now.
Predict. For each case, before your child clicks, ask: “Fixed rule, learns from examples, or need more information? Which words in the description are your clue?”
Try. Let them choose, then read the feedback together. Press Next case and repeat for all four cases.
Explain. Ask: “What evidence helped you decide?”
Hint if stuck: “Does the description mention an instruction someone programmed, or examples it was trained on?”
What you want to hear: your child points to evidence, like “it was trained on labelled photos”, instead of the name of the product. “Need more information” is the right answer for the backpack.
Common mix-ups: “Automatic means AI” (the automatic door follows an if–then rule). “AI is always right” (the photo sorter can still be wrong).
New example
Ask about each: a street light with a timer programmed to switch it on at 6:30 every evening; an email filter trained on thousands of messages that people marked as spam; a “smart” water bottle with no other details.
Answers for the adult
Street light: a rule. Email filter: learning from examples. Water bottle: need more information; we would have to ask how it works.
Away from the screen
Find a timer and one AI feature you already use, such as a phone’s photo search. Your child explains what information each one uses.
Session 2: What a label tells a model
The idea: a training example has features (what is measured) and a label (the answer we give the model). In this activity each dot is an invented fruit with two numbers: redness and length.
Remember. “What is one clue that a machine learns from examples?”
Look. Read the six training examples in the list below the chart, such as “Apple (90, 20)”.
Explain. Ask: “What do the two numbers mean? What does the word ‘Apple’ tell the model?”
Describe. Ask your child to describe the apples and the bananas using only the numbers.
Find the three parts. The drop-down menu is the label you would give a new training example. The model’s prediction appears below the chart once it is trained. The six tests A–F are separate, fixed cards. They are three different things.
Hint if stuck: “The first number is redness and the second is length. Are the apples more red or less red?”
What you want to hear: “The label is the answer we tell the model while training.” And: apples here are red and short; bananas are less red and long.
Common mix-up: “The model knows what an apple is.” It does not see pictures. It only compares two numbers.
New example
“A fruit measures redness 85 and length 25. Using the starter examples, which label would this model predict, and why?”
Answer for the adult
Apple: it is closest to the apple examples, which are very red and short. The reason matters more than the word. A prediction can still differ from a fruit’s real label.
Away from the screen
Make six cards. On each, write an invented fruit: two numbers (redness and length, 0 to 100) and a label. Keep the cards for sessions 3 and 4.
Session 3: Predict before you press
The idea: this tiny model finds the training example closest to a new fruit and copies its label. The dotted line on the chart shows which example it used.
Open:Teach a tiny model and press Starter dataset. Set the sliders to redness 20 and length 25: a green, short fruit.
Steps
Remember. “What does a label tell the model?”
Predict. Before training, ask: “Which training example is closest to this fruit? What will the model say?” Wait for an answer and a reason.
Train and look. Press Train model. The model says banana. Follow the dotted line to the example it used.
Try. Move the sliders. Ask your child to find the place where the answer changes from banana to apple.
Explain. “Why did the answer change there?”
Hint if stuck: “Compare the numbers: which training example has redness and length nearest to 20 and 25?” Go by the numbers, not by how close dots look on the chart; the two axes are drawn at different scales.
What you want to hear: “It copies the label of the closest example.” A wrong prediction explained with “the closest one was a banana” is a good answer.
Common mix-up: “The model thinks about what fruits look like.” It only measures how far apart the numbers are.
New example
“A model has one apple at (90, 20) and one banana at (15, 85). What will it say for (80, 30)? And for (20, 80)?”
Answer for the adult
(80, 30) is closest to the apple, so apple. (20, 80) is closest to the banana, so banana.
Away from the screen
The closest-card game: put four of your fruit cards on the table. Read out a new fruit’s numbers. Your child finds the card whose two numbers are nearest overall and copies its label, just like the model. A quick paper method: add the gap in redness to the gap in length; the smallest total is a good guess. The model measures straight-line distance, so very close calls can differ.
Session 4: Training and testing are different
The idea: to see how a model handles new cases, we test it on examples it has not seen. Testing on its own training examples cannot show how it handles new ones.
Predict. Before testing, ask: “Will it get all six right?”
Test and look. Press Test 6 unseen examples. Five of six match. Test C, the green apple, is predicted as banana.
Explain. Ask: “Why are these called unseen examples? If we tested the model on one of its own training examples, would it be right? What would that prove?”
Think further. “Does five out of six mean the model is right five times out of six in real life?”
Hint if stuck: “If you practise ten maths questions and are tested on the same ten, does a full score show you can solve new ones?”
What you want to hear: testing on training examples is like checking answers you already copied. Six invented tests show how this model behaves; they do not prove it works in real life.
Note: the activity will not let you add a test example’s exact numbers to training. It says that example is reserved for testing. That is on purpose.
New example
“A spelling checker learned from a list of example words, and was tested only with words from that same list. It scores 100%. What should we test instead?”
Answer for the adult
New words it has never seen, including unusual ones. The 100% shows it handles those familiar words; it does not show how it handles new ones.
Away from the screen
Shuffle your fruit cards and hide two. Your child “trains” on the others, then predicts the labels of the hidden two using the closest-card game. Check the answers together.
Session 5: Change one thing and look again
The idea: missing kinds of examples can cause mistakes. Change one thing at a time, so you know what caused a new result.
Remember. “Which test did the model get wrong last time? Why?” (Test C, the green apple.)
Write it down. Press Starter dataset, Train model and Test 6 unseen examples. Note: C was predicted banana.
Change one thing. Set the sliders to redness 30 and length 20, choose Apple, press Add labelled example, then Train model again.
Predict, then test. “Will C change? Will any other test change?” Then press Test 6 unseen examples.
Explain. C is now apple and the others stay the same. Ask: “What was the only thing we changed? So what caused the new answer?”
Be fair about it. We chose this change after seeing C fail, so getting C right now is a fix for a known case, not proof of a better model. To judge a real improvement, test on more examples that you did not look at while choosing the change.
Part B: what did we leave out?
Before opening the activity, tell your child: “This sorter has seen only a blue ball and a red block. Which of these four toys might it get wrong: a blue ball, a red ball, a blue block, a red block?”
Open the activity and press Only two combinations. Compare the results with the prediction.
Ask: “Why did it go wrong?” With so few examples, and a comparison rule we programmed to give colour extra weight, colour decides. Then press Add missing combinations.
Ask: “Do four correct tests mean this model is fair to everyone?”
Hint if stuck: “Look at the examples it has seen. Which kind of toy never appears?”
What you want to hear: “Adding one green apple changed the answer for C.” And: a few correct tests show one behaviour; real systems also need people to check who might be left out.
Common mix-up: “More data always fixes it.” The kind of examples matters, not just how many.
New example
“A model sorts bags into ‘school bag’ or ‘shopping bag’. Every school bag it learned from was black and every shopping bag was blue. Which test bags would you add?”
Answer for the adult
A blue school bag and a black shopping bag, plus bags of other colours in each group. Otherwise it may simply sort by colour. The point is to test the kinds it has never seen.
Away from the screen
Show only two examples: a blue round object goes left, a red square object goes right. Ask your child to guess the sorting rule. Then bring a red round object and a blue square one. Where do they go? Did the two examples really show whether colour or shape matters?
Revisit later: a few minutes, a few times
Recalling an idea after a gap, with a new example each time, can help learning. These timings are our practical suggestion:
The next day, after 3 days and after about a week, ask two or three short questions from earlier sessions.
Use the new example from a session, or invent a similar one. Do not repeat the exact screen.
Note whether your child answered on their own or with a hint. Both are fine; it simply shows what to revisit.
If an idea is still hard, revisit just that idea next time. There is no need to repeat a whole session.
Keep notes on paper. Nothing on this page saves or sends anything.
What this guide can and cannot do
The sessions use small, made-up examples so your child can see why a model is right or wrong. They introduce a few big ideas; they are not a full course. The way the sessions are built, with prediction, explanation and later recall, is informed by research on how children learn, but we have not yet tested this guide with families.
Tried it? Tell us what was clear and what was confusing at hello@ai4kiddos.in. Please do not include your child’s name, photo or school.