Differentiation with AI: a workflow that keeps teacher judgement

76 percent of teachers in English state schools used AI for day-to-day work in February 2026, up from 53 percent a year earlier (NEU survey of 9,408 teachers, published April 2026). 61 percent used it to make resources, 41 percent to plan lessons. Differentiation with AI is where most of that use ends up. It is Sunday at 20:15. A teacher has 26 kids in grade 5 and one fractions lesson on Tuesday. A kid who finishes in eight minutes and starts drawing. A kid who copies the first answer from the neighbour. A kid who reads the word problem three times and never gets to the fraction. The model can write three versions of the task in under a minute. It cannot decide which kid gets which one. The versions are the easy part. The judgement is the whole job.

Start from one objective, not from "make it easier"

We talk about differentiation as a time problem, a class-size problem, a materials problem. It is a judgement problem: who needs what, this Tuesday. The workflow below keeps that judgement with the teacher. It begins with the objective for the lesson, pasted in word for word. For the fractions lesson above, the grade 5 objective in England's mathematics programme of study (DfE, updated 28 September 2021) reads: compare and order fractions whose denominators are all multiples of the same number. A grade 5 class in Denmark meets the same idea in the same year.

The prompt then asks for three versions of one task set: easier, core and harder, with each word defined. Easier is fewer denominators, smaller numbers, one picture per question, same objective. Core is the objective as written, eight questions. Harder is not more questions. It is a reason to compare: a wrong answer to explain, or two fractions where the obvious rule fails.

Paste the objective. Never ask the model to recall it. From memory it produces plausible wording that does not exist in the document.

Differentiation with AI: the four checks before you print

The three versions arrive. Before anything reaches the printer, four checks. With practice, about six minutes per lesson.

1. Reading load. Count the words on the easier version. It is often the longest of the three, because the model explains more when told to simplify. When I taught in the Danish public school system, the kids who needed the easier maths sheet were usually the same kids who needed the shortest sentences. If the easier version has more text than the core, send it back with a word limit.

2. Does the easier version still meet the objective? The common failure is a quiet drop to an earlier goal. The easier sheet becomes "shade three quarters of this shape", which is a grade 3 task. The kid is busy and calm, and not working on this week's goal. One test: could a kid finish this sheet without ever comparing two fractions? If yes, it is a different lesson.

3. Which misconceptions does the harder version surface? For this objective, two are well known: a bigger denominator means a bigger fraction (so 3/8 is judged larger than 3/4), and same numerator means equal fractions. A useful harder version puts both in front of the kid and asks them to argue. If harder is only core with twelve questions instead of eight, ask again.

4. Are the numbers and the progression right? A systematic review of 66 studies on generative AI in maths education (Frontiers in Education, 26 August 2026) found AI's maths errors are "often plausible", and that aligning output with curriculum objectives is a recurring difficulty. Check every answer on the key yourself. Then check the jump: harder should not slide into denominators that are not multiples of each other, because that is next year's objective.

AI gets the same three things wrong

Generic tasks. Pizza slices, chocolate bars and a birthday cake, identical to every worksheet on the internet. The fix is context the kid cares about, and that is a teacher decision, not a model default.

Wrong progression. The model treats "harder" as "more" and "easier" as "an earlier year". Neither is differentiation. All three versions sit on the same objective with a different route through it.

Invented objectives. Asked for a goal code, the model invents one that reads correctly. Only the pasted text is safe.

We see the same pattern in our own tool. On GradeAid, 47.6 percent of worksheets are edited by the teacher after generation (our logs, checked 14 September 2026). We say that number out loud because it is the honest baseline for any generator, ours included.

Three versions is not three groups

Three sheets should not become three fixed groups. The EEF Teaching and Learning Toolkit entry on within-class attainment grouping (last updated August 2021) reports an average of two months additional progress, based on 22 studies, and rates the evidence as very limited. Its conditions matter more than its headline: groups should be flexible "so that pupils know ability is not fixed", and disadvantaged pupils are more likely to be placed with lower prior attainers because of lower teacher expectations.

So the teacher decides who gets which version for this lesson, and decides again next lesson. Version names go on the sheets, kids' names do not. I think letting the model assign the versions is the worst thing a school can do with this workflow, because it turns a Tuesday decision into a label. The teacher in control is the whole point.

The best evidence that teachers hold the line comes from the EEF Teacher Choices trial evaluated by NFER (published 12 December 2024). 259 grades 7 and 8 science teachers in 68 English schools were randomised. The ChatGPT group spent 56.2 minutes a week on planning against 81.5 in the comparison group. An expert panel found no difference in the quality of the resources the two groups used. Teachers used it for questions, quizzes and tailoring materials to specific groups. Quality was kept. Not raised. It was kept because the teacher stayed in charge of the material.

The evidence on training is thinner. A German study of 100 teachers who took a three-hour course on AI for differentiation (Frontiers in Education, 16 September 2026) found that 78 percent rated the content useful for future lesson planning, while their stated intention to use AI did not change. Three hours moves attitudes, not habits. The four checks above are the habit.

AI can write three versions of a fractions task in a minute; only the teacher knows which of the 26 kids should get which one, and why.

• Paste the objective word for word and ask for easier, core and harder on that objective, with each word defined.
• Check reading load, objective, misconceptions and numbers before printing, about six minutes per lesson.
• Assign versions yourself, per lesson, and change the assignment as kids move.

What a school can do

1. Agree one prompt template per subject, with a blank for the pasted objective and the definitions of easier, core and harder written in. About 30 minutes once, in a subject meeting.

2. Put the four checks on an A5 card next to the staffroom printer. About 10 minutes once.

3. Make the assignment of versions a per-lesson teacher decision and say so in writing. No fixed lists that last a term. No extra time, but the head needs to say it once.

4. Keep a shared folder of versions that worked, with the objective in the file name. About 2 minutes per lesson.

5. Write a one-page AI policy for staff. In the NEU survey (April 2026), 49 percent of teachers had no AI policy at all. About 2 hours once, then 20 minutes a term.

Where GradeAid fits

GradeAid is a Danish platform for teachers. The teacher types in the topic and the level and gets a lesson plan, a worksheet, a quiz, a presentation and screen-free activities in seconds, which you can also differentiate. The material can be printed. The worksheet is our weakest format (47.6 percent are edited after generation, our logs, 14 September 2026), and GradeAid does not decide which kid gets which version, that stays with the teacher.

If you work with differentiation in a school or a municipality, we would like to hear from you. Write to us at founders@gradeaid.ai.