Teachers: user of the tool, referee of its use

A teacher meets generative AI twice: as a user preparing lessons, and as the person deciding what counts as cheating in row three. The two roles have different rules, and most of the trouble comes from applying one role's rules to the other's situation.

One number frames both. UNESCO's survey of governments found only seven countries with "frameworks or training programmes on AI for teachers" developed or in development. The realistic assumption is that no one is going to train you, and the judgement calls land on you anyway.

The user role: your preparation time

This is the strong use case, and UNESCO endorses it in concrete terms: "teacher-AI co-designing of lesson plans, course packages, or entire curricula", and using the tools to "inspire new ideas, generate multi-perspective examples, develop lesson plans and presentations, summarize existing materials".

The practical shape: generate three ways to explain the concept that did not land on Tuesday, a set of practice problems at graded difficulty, a quiz from your own notes, a counter-example for the misconception half the class holds. Concrete examples beat adjectives here — paste in an anonymised sample of what your students actually got wrong and ask for material aimed at that.

Two hard limits sit inside the user role.

Every fact you did not check is a fact you might teach. Models produce confident errors, and a wrong date on a generated worksheet is multiplied by the size of the class. UNESCO's own rule covers the teacher as much as the learner: "never accept the information provided by the GenAI at face value and should always critically assess it". Checking a worksheet is faster than writing one; that is the honest version of the time saving.

Student data does not go in. Names, grades, diagnoses, family circumstances — a consumer chat tool is not the place for any of it, and minors' data is the strictest category most teachers will ever handle. Anonymise before pasting, or describe the pattern instead of quoting the student.

The referee role: their learning

The student page gives students the sorting test — does the use build the ability or replace it? — and the referee role is mostly holding assessment to the same test from the other side.

The workable moves are assessment-shaped rather than surveillance-shaped: retrieval checks in the room, asking a student to explain the essay they submitted, staging drafts so the process is visible. UNESCO's classroom example makes the pattern explicit: "Human teachers should ask learners to compare AI tools' art techniques with their own artwork" — put the tool's output and the student's side by side and make the comparison itself the exercise.

What does not work is a detector score treated as a verdict; the university page covers why that instrument cannot carry an accusation on its own.

For younger classes, one line of the guidance is a hard floor rather than a judgement call: for independent conversations with GenAI platforms, "The minimum threshold should be 13 years of age". Below that, tool use in your classroom means mediated use — through you, not beside you.

The role nobody assigned: keeping your own judgement

Delegating lesson drafting is fine. Delegating the judgement of whether an explanation is right, at what level, for which misconception — that is the skill your job rests on, and it decays without use. UNESCO frames teacher development around exactly this: "Define the value orientation, knowledge and skills that teachers need in order to understand and use GenAI systems effectively and ethically". Until someone defines it for you, the working version is: stay the author of what is true and what is assessed; let the tool draft the scaffolding around it.

What goes wrong

Teaching the hallucination. A generated error, unreviewed, delivered with your authority to thirty students who have no reason to doubt it.

Pasting the class into the chat. Real names and real grades in a consumer tool, from the person legally responsible for exactly that data.

Feedback that pretends to be you. Generated comments on student work, unread and undisclosed, hollow out the one channel where students assume a human read their work. Use it to draft; read and own what goes out.

The detector verdict. A percentage from a classifier, presented to a 16-year-old as proof.

Banning in the classroom, using in the staff room. Students notice, and the inconsistency costs more authority than either policy alone.

How to check it worked

Once a term, run your own next assignment through a chat tool before setting it. If the output would earn a decent grade, the assignment measures access to a tool, and it needs redesign before it goes out — better to discover that than to mark thirty variations of the same generated answer. And spot-check the user role too: pick one generated worksheet and verify every factual claim on it once, cold. The error rate you find is the review budget the tool actually costs.