How to Teach Generative AI to Families

A generative model is trained by being shown vast quantities of existing work and repeatedly asked to predict a missing piece — the next word, the next patch of an image. Over billions of these attempts it builds an extremely detailed statistical picture of how such material usually fits together. When it produces something, it is not retrieving a stored answer or consulting a fact; it is generating one piece at a time, each choice shaped by everything it has produced so far. That is why the output reads fluently and why it can be confidently, elaborately wrong: fluency and accuracy are separate properties, and the model is optimised for the first.

The part worth getting right early is the misconception. Most people assume that when it states a fact, it has looked that fact up somewhere. In fact it generated the sentence one piece at a time because that is how such sentences usually go. A fabricated citation is produced by exactly the same mechanism as a correct one, which is why the two look identical. Correcting that once, early, saves a great deal of confusion later — and it is the single idea most likely to stick with the children you teach.

Reading level: Mixed — the adult reads, the children do. Maths assumed: Whatever the youngest participant can follow. Realistic focus in one sitting: about 30 minutes. Pushing past that produces activity, not learning — the child keeps clicking but stops forming a model of what is happening.

Supervision: The adult is a participant rather than a supervisor. Half an hour together beats three hours apart. The value is in the conversation the activity provokes.

The mixed-age problem is real and has a clean solution: do not scale the activity down to the youngest, split the roles instead. In a classifier activity the youngest child collects and sorts the examples, the middle child runs the training, and the oldest designs the test that tries to break it. Everyone is working on the same artefact at their own ceiling, and nobody is watching.

The payoff of doing this as a family rather than individually is the disagreement. When a nine-year-old and a fifteen-year-old predict different results and then watch what actually happens, the conversation afterwards does more than the activity did. Build in the prediction step explicitly — ask everyone to commit to a guess out loud before you run it.

Human next-word game (about 15 minutes, ages 5–10). You need: Paper and two or more people. 1. One person writes the first three words of a story. 2. Pass it on; the next person adds only one word, then folds it so only the last few words show. 3. Keep going for twenty words. 4. Read the result out loud and ask whether anyone planned it. You will know it worked when the child can explain that a sentence can be built one word at a time with nobody knowing the ending in advance.

Same prompt, five ways (about 30 minutes, ages 9–16). You need: A laptop and one approved generative tool. 1. Write one prompt and record the output. 2. Rewrite the same request five different ways — shorter, more specific, with an audience, with a constraint, with an example. 3. Line the five outputs up next to each other. 4. Mark which changes actually improved it and which just changed it. You will know it worked when the learner can point to a specific wording change and say what it did to the output.

Fact-check the confident answer (about 40 minutes, ages 11–18). You need: A laptop, an approved tool, and access to a library or reliable sources. 1. Ask it something factual in an area the learner knows well, then something in an area they do not. 2. Ask it for sources for both. 3. Try to find each source independently. 4. Record which existed, which did not, and which existed but did not say what was claimed. You will know it worked when the learner has personally caught at least one fabricated or misattributed source.

The shortlist is deliberately short. A child who uses one tool properly and finds its limits learns more than one who samples six. Start at the top of this list and only move on when the current tool stops being able to answer the next question.

The most common failure is starting with the mechanism instead of the behaviour. Adults reach for how the system works internally, because that is the interesting part to an adult. A child needs to see the thing behave — make a right guess, then a wrong one — before any explanation of the internals means anything.

The second failure is treating a correct output as the end of the lesson. The learning is concentrated in the failures: the lighting that broke the classifier, the accent it could not parse, the example nobody thought to include. Budget deliberate time for breaking the thing on purpose, and treat every break as the result rather than as a problem to hide.

The third is over-supervising or under-supervising relative to age. The adult is a participant rather than a supervisor. Getting this wrong in either direction costs you — too little and the session drifts, too much and the learner stops making the guesses that teach them anything.

First, every age given is the age the tool becomes genuinely usable, not the vendor's marketing age. Those differ often. Nothing in this collection was excluded on age grounds for this band.

Second, only tools with a genuinely free tier are listed — free meaning a real project can be finished without paying, not a trial that expires mid-activity. 2 of the 5 can be used with no account at all: Quick, Draw!, Scratch. That matters more than it sounds at this age, because an account is a data-collection decision a parent has to make on a child's behalf.

Third, "no screen tool is appropriate yet" is treated as a valid answer rather than a gap to fill. Where this page recommends physical objects over software, that is the recommendation, not an omission.

You can verify all of this yourself in about ten minutes: open each tool listed, check whether it demands an account or payment before producing anything, and see whether the child you have in mind can reach a first result without an adult reading the interface aloud. If any recommendation here fails that test, it is wrong and worth telling us about.