Every teacher knows the arithmetic. Thirty students, one lesson, and at least four different starting points. Differentiation is the right answer, and it is very hard to do at scale, which is why "differentiated instruction AI" is one of the things teachers search for most. The interest is well placed. In Gallup's 2026 study of teachers, among those who received guidance on using AI for different tasks, encouragement was most common for using it to modify student materials to meet individual needs, at 58% (Gallup).
But "AI can make three versions" is not the same as "AI can differentiate." Producing versions of a text is the easy part, and knowing which student needs which version is the hard part, which no tool does for you. This guide covers what AI can do, what it cannot, and how to use it without losing the point of differentiation. The short answer is that AI is a good assistant for preparing materials and supports, and a poor substitute for knowing your students.
What does differentiation actually mean?
Differentiation means adjusting something about the teaching so that more students can reach the same goal. You can adjust the content, which is what students read or work with, and the process, which is how they make sense of it, for example with a partner, a model or a graphic organiser. You can adjust the product, meaning how they show what they know, the support they get, such as hints, sentence frames, worked examples and extra time, and the environment, including grouping, seating, noise and timing.
AI is most useful for content and support, because those are the parts that involve producing many versions of text. It is much less useful for process and environment, which depend on knowing your students in the room. A tool can draft a word bank in seconds, but it cannot tell you that two particular students work better apart.
What is AI good at here?
The strongest use is adapting texts. The same passage can be rewritten at different reading levels, or given a glossary, and our guide to adjusting the reading level of a text with AI covers how to do that without losing meaning. AI is also good at tiered questions, meaning the same skill at three levels of challenge, and at sentence frames and word banks that give students who find writing hard some structure to start from.
It can produce worked examples, either complete models to study or partly completed ones for students to finish, and it can draft choice boards, which offer several ways for students to show learning at a comparable level of effort. For students who finish early, it can suggest extension tasks that are more interesting than another page of the same problems. In every case the value is speed on first drafts, which you then improve.
What can't it do?
It does not know your students. It cannot tell who is anxious, who is bored, or who hides confusion behind a quick "yes." It cannot diagnose what a student needs, since that is your professional judgement, made with specialists where appropriate. Decisions about students with identified needs belong with your learning support team and the student's plan, and an AI draft should never replace them. And it will not tell you when it got the level wrong. It hands over a confident draft either way, so the checking is always yours.
What does this look like for one lesson?
Take a grade 6 lesson where the goal is that students can explain how energy moves through a food web. A single well-built prompt can produce three versions of the same task.
Create three versions of a task on food webs for grade 6, all with the same goal: explain how energy moves through a food web. Version A (support): a partly completed food web with a word bank and two sentence frames. Version B (core): students draw a food web for a pond from a list of organisms and write three sentences explaining the energy flow. Version C (stretch): students predict what happens to the web if one organism disappears, and justify the prediction. Keep the same key vocabulary in all three: producer, consumer, decomposer, energy.
Three design choices make this work. The goal is the same in all three, so no student is working toward a smaller aim. The key vocabulary is the same, so students can discuss the topic together afterwards. And the difference lies in how much structure each student gets, not in how interesting the thinking is.
Which prompts help most?
A handful of prompts cover most needs. To adapt a text, ask the AI to rewrite a passage at a grade 4 reading level, keeping every fact and every term in a list you give and defining each term the first time it appears. To tier questions, ask for six questions on a skill, two basic, two medium and two challenge, in that order. To add frames, ask for three sentence starters that would help a student explain an idea in writing. To make a model, ask for a worked example with each step labelled, followed by a similar problem with one step missing. To offer choice, ask for four ways a student could show they understand an idea, each needing about the same effort. Save the ones that work for you and reuse them.
Does more support always help?
A small 2026 study is a good caution against assuming it does. Researchers compared four ways of presenting a text to 14 primary-school learners with special educational needs and disabilities: plain text, text split into segments, segments with pictograms, and segments with pictograms and keyword labels. Some learners seemed to benefit from segmentation and pictograms, while others seemed to be slowed by the extra visuals. The authors concluded that no single scaffold is universally optimal, and that supports need to be adjustable (arXiv preprint, not yet peer reviewed, and a very small sample).
The practical lesson is simple. Offer the support, watch how each student responds, and remove it when it gets in the way. A fifth scaffold on the page is not automatically an improvement, and a student who is slowed by extra visuals is being made to work harder for the same content.
What should you check before handing it out?
Start with the goal. Every version should lead to the same understanding, so check that the support version has not quietly lowered the target. Make sure the support is removable, meaning students can work up toward the core version instead of being stuck below it. Watch for "support" versions that water down the thinking as well as the reading, because the reasoning should stay. Check that the "stretch" version is a harder question and not just more of the same, since extra problems are not a challenge. Verify any worked example yourself, and look at the wording on the page: labels such as "low group" are not respectful and students notice them.
How do you protect student information?
Do not paste names, diagnoses, IEP details or other identifying information into a general AI tool. Describe the need in general terms instead, for example "a student who needs shorter sentences and a word bank." Our student data privacy checklist explains why this matters and what to ask your school about the tools you use.
How do you keep the workload manageable?
You do not need three versions of everything. A sustainable pattern is to make the core task as you normally would, then prepare one support option that adds structure and one stretch option that adds challenge, and to do this for the one or two lessons a week where the range of needs is widest. Keep what worked and reuse it next year. For worksheets in particular, see our guide to making better worksheets with AI, which shows how to build support and stretch versions from a single page.
Frequently asked questions
Can AI write an IEP or support plan? It should not. Those documents involve sensitive data and professional responsibility, so they belong with you, your colleagues and the student's team.
How do I know which students get which version? From what you know of them, and from short checks such as exit tickets. Let students move between versions as they learn, and avoid fixed labels.
Is differentiated work more work for me? It can be, which is why it is worth using AI to draft supports and then reusing them. Focus on the lessons where the range is widest.
Does differentiation lower standards? Not if the goal stays the same. The support changes how students get there, and the stretch keeps stronger students thinking.
Related guides
- Adjusting the reading level of a text with AI
- Making better worksheets with AI
- Free AI tools for teachers
- Education News: latest research, a live feed of new education studies