AI and homework integrity: the primary-school version of the problem

Evidence grade: the capability facts are established; the responses are early practice, honestly graded. That current AI systems can produce competent homework-grade text and code is not in dispute. What works pedagogically in response has practitioner consensus forming but little trial evidence yet — and the primary-age version differs enough from the secondary panic to deserve its own page. Detection's failure is covered in depth on its own page.

The homework-integrity conversation arrived panicked from secondary schools, and primary teachers inherited the panic without the analysis. The primary version of the problem is different: less student-driven cheating (eight-year-olds rarely orchestrate deception pipelines), more well-meaning adult and tool contamination — the parent who "helps" via a chatbot, the household assistant that answers the maths sheet, autocomplete finishing sentences. The question isn't mainly "how do we catch them?" — it's what is homework evidence of, now that fluent output no longer proves a child produced the thinking?

What actually breaks (and what doesn't)

AI assistance breaks homework's assessment function — output no longer certifies the child's capability — while leaving its practice function damaged only when the tool does the thinking rather than supporting it. That split points at the answer: stop using take-home output as evidence, keep (and redesign) take-home practice. Conveniently, primary education is structurally positioned for this: the assessment evidence that matters can live in class, where performance is observable — a luxury secondary essay subjects envy.

The response repertoire

Move certification in-class. Performance under observation — the mastery-evidence model — is the one integrity mechanism no tool undermines. Homework then becomes low-stakes practice whose honesty benefits the child rather than guarding a grade (which also changes the family conversation: "it only helps if it's your brain" is true and checkable).
Ask for process, not just product — drafts, workings, "tell me how you did it." At primary age this is easy to make natural, and a two-minute conversation about a suspiciously polished piece resolves what no detector can.
Design AI-assumed tasks where it fits — "ask a grown-up or a helper for three ideas, then pick the best and say why" makes the tool part of the work and the judgment the assessed thing: the AI-literacy curriculum as homework design.
Never outsource accusation to detectors — they don't work, and a false accusation of an eight-year-old (or their parent) spends trust nothing refunds.
Teach the integrity frame young and blame-free: "your homework is practice for your brain — a robot doing your push-ups gets the robot strong" lands at seven, before stakes and shame arrive. This is the window logic applied to academic honesty.

What the evidence doesn't say

It doesn't show rampant primary-age AI cheating — the documented pattern at this age is adult-mediated contamination; design for that, not for imagined mini-fraudsters.
It doesn't settle whether AI help harms learning — tool-supported practice vs tool-replaced practice is the live research question; the practical line ("did the thinking happen in your head?") is a defensible placeholder.
It doesn't make homework obsolete — practice still needs reps; it makes unsupervised output as certification obsolete, which was always its weakest use.

In the classroom

Audit each homework's purpose — practice or evidence? Keep the first, move the second in-class.
Normalize the disclosure question: "what helped you?" asked routinely and without heat gets honest answers and teaches tool-transparency as a norm.
Brief parents explicitly — most contamination is loving; one newsletter paragraph ("here's how to help without replacing the thinking") prevents most of it.

How Wiz Kids applies this

The architecture pre-dates the problem but answers it: all certification is in-app performance — skills pass by doing, live, and review gold certifies retention on the machine, under the teacher's model of who-was-at-the-keyboard. Home practice is a bonus lane that adjusts nothing a teacher relies on, so a helpful parent (or chatbot) can contaminate nothing that matters — and the curriculum teaches the helper-never-boss frame directly.

References

Cotton, D., Cotton, P., & Shipway, J. R. (2024). Chatting and cheating: Ensuring academic integrity in the era of ChatGPT. Innovations in Education and Teaching International, 61 — the capability-and-response landscape.
Liang, W., et al. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7) — why detection can't carry integrity (fuller treatment on the detection page).
Messick, S. (1995). American Psychologist, 50(9) — validity: what an assessment is evidence of, the frame this whole page applies.

© 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.