Differentiated Instruction with AI — Hand Over the Copying, Keep the Judgement

It is 4:40 on a Thursday and a Form 3 English teacher is sorting a set of paragraph responses into piles on the staff room table. One pile states a claim but never supports it. One pile lists evidence with no claim anywhere. Two students have written something she would happily read aloud. She taps the thin pile and says, "I know they need more than this, but I've got one free period tomorrow."

Differentiation stalls in Hong Kong classrooms not because teachers disagree with it, but because two very different jobs get crushed into the same evening: working out where each student is stuck, and turning that into three sets of materials. AI can only take the second job, and it has to take it cleanly for the first one to get any air.

Two jobs hiding inside lesson prep

Pull an evening of preparation apart and you find two kinds of labour that happen to share a desk.

The judgement half

Two students both score four out of ten. One has not grasped what a claim is; the other knows exactly and ran out of time. The remedies point in opposite directions. Telling them apart depends on what you saw in class, what he asked during recess, the fact that she was absent for the modelling lesson. None of that is on the page, and none of it fits in a prompt box.

The copying half

Applying "paragraphs of three lines or fewer, verbs in bold, one instruction per sentence" to a worksheet is mechanical. Forty minutes of teacher time and forty seconds of machine time produce the same artefact. The only difference is that the forty minutes could have been spent on those three piles.

There is a quick test for whether a task is safe to hand over: can you write the method as instructions someone else could follow and reproduce? If yes, delegate it. If you cannot get it into words, you are still making a judgement, and delegating it just returns a very tidy mistake.

Write one objective you can see

Most failed attempts at AI differentiation happen right here. Ask for "Form 3 English, argument paragraphs, three difficulty levels" and you get three worksheets with different question counts and identical cognitive demand.

An objective you can see looks like this: from a short opinion text, identify the writer's claim, name one piece of evidence used to support it, and explain why that evidence is or is not sufficient. Three observable actions - identify, name, judge. The doors get cut from those actions, not from the number of questions.

The check that takes ten seconds

Read the objective and ask what a successful student's page looks like. If you can answer concretely - "the claim underlined, the evidence circled, two sentences underneath" - the objective is solid. If the answer is "they'd understand it", it is not, and no amount of prompt tuning will save the materials.

Three doors into the same objective

Not levels. Doors. What separates the three versions is where the student picks the work up and how many handrails are left in place, not how much they do.

The first door supplies the text with the claim and evidence already marked, and asks the student to explain why the evidence works. Maximum support, but the required action is reasoning, not locating. The second door supplies the text with one prompt - "find the sentence the rest of the paragraph is defending" - and leaves the rest to the student. The third door reverses everything: it gives a short paragraph with strong evidence and a weak claim, and asks the student to rewrite the claim so the evidence actually supports it, then explain what changed.

The third door contains less writing than the first. It is also the hardest, because it runs the reasoning backwards. That inversion is the part people miss.

Reading "harder" as "more questions" is the most expensive mistake in differentiation: the reward for finishing early becomes a longer punishment.

One rubric, all three doors

If the output from the third door cannot be marked with the rubric you wrote for the first, it is not another door into the same objective - it is a different objective. Run this check before printing anything: put the rubric next to all three sets of expected output. If door three does not fit, rewrite door three, not the rubric.

Adaptation is layout and language, not difficulty

Adaptation and tiering get conflated constantly, and the cost is that SEN students quietly receive the easiest version all year.

A Primary 6 non-Chinese-speaking student in General Studies usually does not struggle with how a council works. He struggles with three words - motion, vote, representative. What he needs is not thinner content; it is a ten-word bilingual glossary in the margin and "compare and explain" split into two sentences: write two differences, then write which one matters more and why. The intellectual demand has not moved at all.

Dyslexia adaptations live in the layout - short paragraphs, bolded verbs, taller answer boxes. ADHD adaptations live in the pacing - one task per page, each with a visible finish line. All of that is expressible as rules, and anything expressible as rules can be applied across a whole unit in one pass, consistently. Consistency is itself part of the adaptation: the student should not have to relearn the page every Tuesday.

What must not move is the direction. Which student needs which adaptation is a judgement made by the subject teacher and the SENCO against an individual learning plan. Give an AI "these three students need reduced visual load, these two need bilingual key terms" and it performs reliably. Give it "help me support the SEN students in my class" and you get a bland simplification.

After the materials reach the room

Generating is maybe a third of the work. Handing out and recording decide whether any of it survives past week three.

Put all three doors in one packet or one digital task and let students turn the page themselves. No allocation list. Number them however you like in your own planning, but the student-facing labels should be neutral - name them after the text or the scenario, never one, two, three.

Record against actions, not doors. "Claim identified, evidence named, judgement missing" tells you what tomorrow's ten minutes are for. "He did door one" tells you nothing and hardens into a label. Review the column every fortnight: three consecutive misses on the same action is no longer a differentiation problem, it is a student who needs fifteen minutes alone with you.

Common misreadings

• The claim: once three doors exist, the teacher can circulate calmly in the middle of the room. What actually happens: you talk less and move far more, and the questioning gets harder. Ten minutes in, twenty minutes working, ten minutes converging is the shape that holds up on a first attempt.
• The claim: AI-generated adaptations are ready to print. What actually happens: layout rules hold, vocabulary drifts. Two minutes per sheet, ten minutes per unit, reading only the instruction lines. That check does not compress.
• The claim: given a choice, everyone picks the easiest door. What actually happens: for the first lesson or two, yes. Once the third door's output gets read out at the end of the lesson and carries no extra workload, the distribution spreads on its own.

Three things to try on Monday

1. Take one lesson from next week and rewrite its objective as a single sentence containing two or three observable actions. Put it on the top line of the plan.
2. Generate all three doors from that sentence, then hold the marking rubric against the expected output of each before you print anything. Rewrite whichever door the rubric cannot reach.
3. Add one column to your class list holding only ticks and crosses for the actions in that objective. Fill it for two weeks before drawing any conclusions.