AI literacy for primary ages: helper, not oracle

Evidence grade: THIN by necessity, principled by design. Generative AI reached children faster than research about children-and-AI could be conducted; direct studies of primary-aged AI literacy instruction are just emerging. This page builds from what IS established — children's credulity toward confident answers, the machine-text and misinformation literatures, transferable principles from search-evaluation research — and labels the rest as reasoned position.

Primary-aged children are now growing up alongside systems that converse fluently, answer any question with total confidence, and generate homework-shaped text on demand. The adults around them oscillate between "it's cheating-machine evil" and "it's the future, embrace it" — neither of which is a mental model a nine-year-old can use. AI literacy at this age is mostly about installing one accurate mental model and three habits before the magic-thinking sets.

The mental model: a very well-read guesser

The child-sized frame we use, defensible to researchers and comprehensible at eight: an AI chatbot is a machine that has read almost everything and guesses what words come next — so it speaks smoothly whether it's right or wrong. Every load-bearing property follows from it: why it's genuinely useful (it has read almost everything), why it sounds certain when wrong (guessing word-by-word produces fluent confidence — there is no "I'm not sure" feeling inside it to express), why it can invent facts, books, and people (a plausible-sounding guess IS its product), and why the too-smooth voice has no fingerprints. The model matters because the default alternative — the oracle model, where the computer knows — is precisely the one children arrive with, and the one every confident wrong answer exploits. Children's testimony research says they weight confidence heavily in judging informants; a machine that is only confident needs demystifying early.

The three habits

Check the checkable. AI answers about facts get the second-source treatment — same habit, new target. "The machine said so" joins "the website said so" as the start of checking, not the end.
You're the boss, and the editor. Useful AI work is directed — the child asks, evaluates, redirects, keeps or discards. The posture is employer, not supplicant; treating output as a draft from an eager, unreliable assistant is both accurate and empowering.
Never feed it what's private. Personal information rules apply to chatbots exactly — a text box that talks back is still a text box, and a friendly voice is not a trusted adult. (The parasocial dimension — children befriending chatbots — deserves naming with parents; the research is young, the design incentives aren't.)

The homework question, answered without panic

The integrity conversation goes wrong when framed as detection-and-punishment — not least because machine-text detection is unreliable and accusations from style alone hurt real writers. The framing that works at this age is the one our fiction uses: homework a machine writes teaches the machine, not you — the point of the writing was never the artifact; it was the practice, and outsourced practice is practice that didn't happen. Alongside it, the legitimate-use conversation: asking AI to explain differently, quiz you, or critique your draft is using the assistant as a teacher; asking it to be you is using it as a substitute — and the difference is who did the thinking.

What the evidence doesn't say

It doesn't yet quantify much — child-specific AI-literacy trials are arriving, not arrived; humility is the honest posture and this grade will move.
It doesn't support AI-abstinence education — the tools are ambient; untaught children get the oracle model by default.
It doesn't make primary children prompt engineers — the model and habits are the curriculum; technique can wait.

In the classroom

Teach the guesser model explicitly — then demonstrate a confident wrong answer live; one witnessed hallucination beats a term of warnings.
Run the editor exercise: AI produces a paragraph, children find what's wrong or weak and fix it — the posture lesson disguised as fun.
State the homework principle in the open — "it teaches the machine, not you" — and name the legitimate uses just as clearly.
Keep the privacy line absolute and boring: no names, no school, no addresses in any chatbot, ever.

How Wiz Kids applies this

The Hall of Mirrors teaches the guesser model in-fiction (AI as a helpful golem that "dreams answers" and must be checked), the machine-text lessons train the too-smooth-voice ear, the homework principle is taught as judgment scenarios ("the machine's essay teaches the machine"), and the privacy rules cover chatbots explicitly. No generative AI runs inside the product itself — every simulated "AI" is authored, so children study the phenomenon in a specimen jar before meeting it wild.

References

Koenig, M. A., & Harris, P. L. — children's trust-in-testimony research program (confidence as a credibility cue in young children).
Breakstone, J., et al. (2021). Students' civic online reasoning. Educational Researcher, 50(8) — the evaluation-skills baseline AI literacy extends.
UNESCO (2023–). AI competency frameworks for students — the emerging institutional scaffolding.
OpenAI/academic literature on LLM hallucination and calibration — the technical basis of the "confident guesser" model, translated.

© Glu IO Pty. Ltd. — Wiz Kids (wiz.kids). Link freely; republication requires permission — see terms. Found an error in our reading of the research? We correct fast: tell any teacher piloting Wiz Kids.