For educators

AI in the Classroom: Ground Rules That Actually Work

Published August 21, 2026 · By the Kidgeni team

Start by correcting the premise

Stanford researchers surveying students across 40 US high schools found that 60-70% reported at least one cheating behaviour in the previous month — and that this rate held steady, or dipped slightly, after ChatGPT arrived. Stanford's faculty lead for AI and education summarised it plainly: there is little evidence the needle has moved for high schoolers overall.

Two things follow. First, the crisis framing is wrong: cheating was already widespread, driven by the same pressures it always was — grades, time, and assignments that feel like compliance rather than thinking. Second, and less comfortably, a policy aimed only at catching AI use is aimed at the wrong problem. The question worth asking about an assignment is not 'can a chatbot do this?' but 'why would a student want to skip doing this?'

Why detection tools are the wrong foundation

AI-detection software produces false positives, and they are not randomly distributed — students writing in a second language and students with more formulaic writing styles get flagged disproportionately. A false accusation costs a relationship you cannot buy back, and the accused student has no way to prove a negative. Use process evidence instead: drafts, version history, in-class writing, and a two-minute conversation about the work. A student who wrote it can talk about it; a student who did not, cannot.

The four-tier scale

The single most effective classroom change is printing the expected AI level directly on each assignment. Students stop guessing, and you stop adjudicating ambiguity after the fact.

• Tier 0 — No AI. This work is the thinking itself. Typically done in class, on paper, or with drafts collected. Say WHY: 'the struggle is the lesson here.'
• Tier 1 — AI for getting unstuck. Brainstorming, topic ideas, explaining a concept a different way. The words submitted must be the student's own.
• Tier 2 — AI as a drafting partner, with disclosure. Outlining, feedback on a draft, rewriting for clarity. Student attaches what they asked and what they changed.
• Tier 3 — AI encouraged. The AI use IS the skill being assessed: prompting, evaluating output, fact-checking it, improving it.

Make disclosure cheap and honest

Students under-report AI use for exactly one reason: they expect punishment. If your disclosure line carries no penalty when the tier permits use, you get accurate information about how your class actually works — which is worth more than a detection score. Make the disclosure a single sentence, make it routine, and never retroactively punish an honest one.

Ground rules worth having in writing

• Check your state guidance first — 34 states plus Puerto Rico had issued official K-12 AI guidance by late 2025, and it usually settles the data questions for you.
• Only district-approved tools with a data agreement touch student information. Personal consumer accounts are not a place for student work.
• Age rules are the platform's, not yours to waive: most general-purpose AI assistants set 13 as a floor, so under-13 classrooms need tools actually built for children.
• Nothing a model outputs is a fact until a human checks it — make verification part of the assignment when AI use is permitted.
• Students see the double standard instantly. If you draft with AI, say so.

For elementary classrooms specifically

Most of the discourse above assumes secondary. For ages 5-12 the calculus is different: the general-purpose assistants are 13+ by their own terms, so the realistic options are tools purpose-built for children with moderation and adult-owned accounts. And the pedagogical goal is different too — you are not managing academic integrity so much as building the first mental model of what these systems are. Our age-banded guide covers what that looks like at each stage.

• Teaching Kids About AI: what to cover at each age
• AI in Education: what's actually happening in schools
• AI for Teachers: where the hours actually come from