Differentiation Was Never Buried

I had a good lesson recently. Year 9, English Literature, A Midsummer Night's Dream. The students were in the text; asking questions about it, arguing with it, writing about it without the usual negotiation over whether writing was strictly necessary. There was that quality of attention in the room that teachers recognise and rarely feel entitled to claim. They were thinking. I noticed it, said nothing about it, and was quietly pleased.

Then I read the feedback form. Most of it was fine. And then, almost as an afterthought, the line: "Differentiation could have been made more explicit. I want to be precise about what I felt. Not outrage. Something flatter; a door closing on a room I'd actually liked the look of. Because I have come to understand something about that sentence: it could follow any lesson. Any lesson, any teacher, any room. There is no observable threshold at which differentiation becomes sufficient. There is no version of a lesson that could not be told, afterwards, that more was possible. The demand has no upper limit, no fixed form, and no clear definition of done.

That is not a high standard. That is no standard at all.

This is what I want to argue: differentiation has survived not because it works, but because it cannot be proven to have failed. It is unfalsifiable as a demand; and that makes it uniquely resistant to criticism, and uniquely corrosive to the teachers it is applied to.

Differentiation did not begin this way. It began as a humane objection to sameness; the recognition that students arrive in classrooms with different knowledge, confidence, language and need, and that a single pace, a single text, a single task cannot serve all of them equally. That instinct was sound. A classroom is not a level surface. Pretending otherwise has always excluded students quietly and at scale.

The problem was not the principle. The problem was what happened when institutions tried to make the principle visible.

In many schools, differentiation drifted from an act of professional attention into a performance of planning. It was no longer enough to know your students and adapt accordingly. Teachers were expected to prove, in advance, that adaptation had occurred. Difference had to be labelled. Support had to be colour-coded. Challenge had to be documented before the lesson had happened. Responsive teaching — which is alive, contingent, dependent on reading the room as it moves — was asked to present itself as a folder that could be handed to an observer.

This is where the unfalsifiability entered.

Because responsiveness, by its nature, cannot be fully planned. A student who needed scaffolding on Monday may need to be left alone on Wednesday. A student who normally flies may be tired, lost, somewhere else entirely. The lesson changes as it moves, and good teaching changes with it — a question redirected, a student left to sit with an idea a moment longer, a scaffold removed, an explanation tried a different way. None of this leaves a trace in the planning document. None of it can be pointed at.

So the system learned to ask for the things that could be pointed at. And in doing so, it created a standard that was permanently, structurally out of reach. The observable evidence of differentiation — the tiered worksheet, the colour-coded task, the documented pathway — was always only a representation of responsiveness, never responsiveness itself. But the representation became the measure. And representations can always be more thorough, more visible, more explicit.

Hence the feedback that follows good lessons as readily as poor ones.

I have been teaching in classrooms for sixteen years in total — thirteen as an ESL language teacher before I qualified formally. That earlier career gave me something I would call a feel for learner difference. Watching students at different levels move through the same language, adjusting explanations mid-sentence, reading faces, working out in real time who had understood and who was performing understanding; that is adaptive teaching. It is not something I do when observed. It is how I teach.

In my Year 9 lesson, that responsiveness was present throughout. Shakespeare is difficult language. These are fourteen-year-olds, many working in a language that is not their first. The adaptive work of holding that lesson together; a question simplified here, a student redirected there, the text unpacked and then allowed to become difficult again — was considerable. It just didn't look like three worksheets.

The feedback could not see it. The feedback was not designed to see it. And this is the real cost of an unfalsifiable standard: it doesn't only fail to reward good teaching. It actively misdirects attention away from it. If the measure is visible, documentable differentiation, then the teacher learns to produce visible, documentable differentiation. The eye moves from the student to the evidence. The question shifts from what does this person need right now to what will this look like in the folder. These are not the same question. They do not produce the same teaching. AI will make this worse.

There is a version of the AI-in-education story that goes like this: differentiation was always the right idea, but the workload made it impossible; AI removes the workload and so restores the possibility. A teacher can now generate three differentiated versions of a task in seconds — simplified texts, vocabulary banks, tiered questions, extension prompts. The labour that once made differentiation unmanageable has been absorbed by the machine. That version of the story is superficially plausible and I think it is wrong.

The workload was never the deepest problem. The deepest problem was the unfalsifiable demand — the expectation of visible, documented, inspectable differentiation that moves its target whenever you approach it. AI does not dissolve that demand. It hardens it. If tiered resources are now effortless to produce, the expectation of them will calcify. The observer's question: ‘Where is the differentiation?’ becomes cheaper to ask and more difficult to escape. The folder gets thicker. The eye stays on the document. The student remains, in some important sense, beside the point.

There is also a subtler problem. AI can produce variation in materials. It cannot know what the variation is for. A simplified text may remove exactly the difficulty a student needed to encounter. An extension task may produce more activity without producing deeper thought. A vocabulary bank may support one student and become a crutch for another. The resource is not the judgment. The resource is a guess about what might be needed, and the judgment is what happens when that guess meets an actual person. AI can accelerate the guessing. It cannot do the meeting.

The fantasy that differentiation was always quietly selling was that the right resource, prepared with sufficient care, could solve learner difference before the lesson began. AI makes that fantasy faster and cheaper. It does not make it more true. And if the fantasy becomes easier to indulge, the unfalsifiable standard becomes easier to enforce — because now there is no longer even a workload argument to hide behind.

An unfalsifiable standard does a specific kind of damage. It is not simply that it is unfair, though it is. It is that it cannot be engaged with honestly. You cannot improve in response to feedback that could have been given regardless of what you did. You cannot learn from a standard that offers no picture of sufficiency. What you can do — what the system trains you to do — is produce more visible evidence and hope that this time it is enough. This is not professional development. It is performance management dressed as pedagogy.

The teacher who received my feedback did not need to be told to care more about learner difference. She was already caring about it, in the room, with the students, in the lesson the observer had just watched. What she needed, what the system is structurally unable to give, was a definition of enough. A description of what differentiation looks like when it is working, that is specific enough to be achieved and honest enough to be believed.

That definition does not currently exist. And as long as it doesn't, the feedback will keep coming. It will follow good lessons and poor ones with equal confidence. It will arrive after Shakespeare and after worksheets and after AI-generated tiered tasks and after whatever comes next. It will always be available, because its target always moves.

I am not arguing that learner difference doesn't matter. It does. Students are not the same. Some need more language before they can think clearly in it. Some need the task broken into smaller steps before it stops feeling impossible. Some need to be pushed past the first safe answer into something harder and more true. Noticing this; really noticing it, in the room, with the actual students; is part of what teaching is.

What I am arguing is that this noticing cannot be reduced to a document, cannot be produced in advance, and cannot be made explicit in any form that will finally satisfy an unfalsifiable demand. The teacher who adjusts her question because she can see a student is lost, who leaves another alone because the struggle is useful, who slows the whole class down because something hasn't landed; that teacher is doing the work. The fact that none of it appears in the folder is not a failure of documentation. It is a feature of teaching.

The system that cannot see this will keep asking for more. AI will make the asking easier. Neither development makes the teaching better.

After my lesson observation, I sat with the feedback for a while. I thought about what I could have done differently, because that is what you do. I could have prepared a tiered worksheet. I could have labelled the tasks. I could have made the differentiation explicit, in the precise sense that the feedback meant; visible, pointed-at, present in the room as an object rather than as a practice.

I don't think the lesson would have been better. I think it would have looked more like what the form was looking for. Those are not the same thing. And until the system understands that, the feedback will keep arriving, unchanged, at the bottom of the form; ready to follow the next good lesson into the folder where it will finally, briefly, count as evidence. Differentiation survives because it operates as an endlessly movable standard, and AI may make that standard easier to impose.