Using AI in teaching without losing the assessment

Where it earns its place for teachers

The honest list is narrower than the marketing and still worth having: first drafts of schemes of work, differentiating a text three ways, generating practice questions, rephrasing a letter home, summarising a long document, and producing the scaffolding around a lesson that takes an hour and teaches nobody anything.

What it is not good at is knowing your class. The output is a starting point that a teacher edits, and the editing is where the professional judgement goes in. A draft taken unedited is usually worse than what the teacher would have written.

Why detection is not the answer

AI-detection tools report false positives, and they do it unevenly — pupils writing in a second language and pupils with a plain, formulaic style get flagged more often. An accusation of cheating based on a probability score is not a position a school wants to defend to a parent.

The more durable response is to change what is being assessed rather than to police the tool. Work produced unsupervised at home now carries an irreducible uncertainty about authorship, and no software is going to remove it.

Setting work that still tells you something

• Weight in-class, supervised and oral work more heavily where the point is to know what a pupil can do alone.
• Assess process as well as product — drafts, notes, a short conversation about the choices made.
• Set tasks anchored to things a model does not have: this class’s trip, this week’s discussion, this pupil’s own data or experience.
• Where AI use is permitted, require it to be declared and ask the pupil to say what they changed and why. That is a better assessment of understanding than the essay was.