If you teach

Where it genuinely helps

Teaching has an unusually high ratio of preparation to delivery, and much of the preparation is transformation of material you already understand. That is the good territory.

• Differentiation. One worksheet, three versions: one scaffolded, one standard, one extended. This is a genuine hour saved every time, and the subject content is yours.
• Practice items. Twenty more questions of the same type, with different numbers and contexts. Check them — generated questions in science and history contain factual errors more often than you would like.
• Reading level. Rewriting a source text for a lower reading age while keeping the substance. Genuinely good at this.
• Letters home, including translation into the languages your families actually speak. Check the translation backwards, as the translation lesson explains.
• Making a rubric explicit. You know what a good answer looks like; getting it into words a pupil can use is hard, and this helps.
• Lesson plan skeletons you then fill with the things only you know about this class.

Feedback: the direction that works

The productive pattern is narrow and worth stating exactly.

You read the work and form the judgement, in three or four bullet points of your own. Structure fine, evidence thin in paragraph two, misuses "significant", strong conclusion.

The model turns your bullets into two paragraphs of feedback in a supportive register, at the right reading level.

Every judgement came from you. What was outsourced is phrasing, which is the part that takes the time and carries no professional content. This is the same division as the dictation lesson, and it is the safest configuration available.

The reverse — giving it the pupil's work and asking what is wrong — hands over the judgement, produces generic comments dressed as specific ones, and puts a child's work into an outside system.

Do not hand over grading

A grade is a decision about a person, with consequences for that person. It also depends on criteria you apply in a particular context, moderated against a cohort you know.

Models grade consistently, which is easily mistaken for grading validly. Consistency without validity is the most dangerous combination there is, because it produces stable, defensible-looking numbers that are systematically off in ways nobody can see from the numbers.

Detection tools: the honest position

This section matters more than the rest of the lesson.

AI-detection tools produce false positives, and published work has repeatedly found that they flag writing by non-native English speakers at substantially higher rates than writing by native speakers. Simpler, more formulaic prose reads as machine-written to a detector, and that describes a great deal of competent second-language writing, and also the writing of some autistic and dyslexic students.

Several institutions have turned these tools off for exactly this reason. Some vendors publish accuracy figures; those figures come from the vendor and are typically measured on text unlike your pupils'.

So: a detector score is not evidence, and must never be the basis of an accusation. Treating it as one means the burden of proof falls hardest on the students least able to contest it, which is not a marginal unfairness — it is a systematic one aimed at a group you can identify in advance.

What works instead:

• Process evidence. Version history in a shared document, drafts, notes, a planning sheet. Ask for the process at the outset rather than demanding it under suspicion.
• A conversation. Ask the pupil to talk you through the argument. Five minutes settles it, and it is fair in a way a score is not.
• Assessment design. Tasks anchored in class discussion, in this week's local example, in the pupil's own data or observation. Not to defeat the tool, but because those are better assessments anyway.

Pupils using it, and the position worth taking

They are using it. A ban you cannot enforce teaches concealment.

The workable position is to teach the use explicitly — including everything in this course about checking and about work you can defend — and to set some tasks where it genuinely does not help. Require the working, the sources, the intermediate drafts. Ask for the thinking rather than only the artefact.

Pupil data, and age limits

Pupil work is personal data. In many contexts — a safeguarding note, a special educational needs record, anything about health or family circumstances — it is the more sensitive kind, with stricter rules.

Most consumer AI products set a minimum age in their terms, commonly 13 or 18, sometimes with parental consent provisions. A teacher pasting a class set of essays into a personal account is processing children's personal data through a service the school has not assessed and the parents have not been told about. That is a matter for your school's data protection lead before it is a matter of pedagogy.

The free path

LibreOffice for everything you produce. A local model through Ollama for anything containing a pupil's name or work — and differentiation, rewriting and feedback phrasing are all tasks where a small local model performs close to a large one. Whisper locally for transcribing your own spoken notes after a lesson.

And the free open educational resource libraries that already exist. A generated worksheet is not automatically better than a good one somebody has already written and tested with three hundred children.