AI Differentiation: How to Adapt One Lesson for Every Learner

What is differentiating with AI?

Differentiating with AI means using an AI tool to adapt one lesson to the spread of learners in a real room — scaffolded versions of the same text, language supports, extensions, accommodation-ready formats — drafted in minutes and reviewed by the teacher before any of it reaches a student.

Here's the room this guide is written for. Second period, 7th-grade ELA, 29 students. Two are newcomer multilingual students, both Spanish speakers around WIDA level 2, arrived in October and November. Five read near a 4th-grade level on the fall benchmark. Seven finished last unit's novel days early and have been coasting since. The other fifteen sit somewhere in the middle. This week the class starts an argument-writing task built on an 1,100-word article about the 1911 Triangle Shirtwaist Factory fire — who bears responsibility for the 146 deaths? — under CCSS RI.7.1: cite textual evidence to support analysis.

One article, one standard, twenty-nine different distances between student and text. Closing those distances is the work teachers most often name when they describe what AI is for: in Gallup and the Walton Family Foundation's 2025 teacher survey, 28% of teachers reported using AI to modify materials to meet students' needs. The prompt mechanics underneath everything below — pasting the full standard, naming your room's constraints — are the same ones in our AI lesson planning guide. This article assumes the lesson exists. The question is how everyone gets into it.

Why "a version for every level" burns teachers out

The classic advice — a below-level, on-level, and advanced version of each lesson — fails on arithmetic before it fails on pedagogy. Tripling every artifact across a week of lessons is a production load that collapses under its own weight within a few weeks — a big part of why differentiation gets praised in every PD session and practiced in far fewer classrooms. AI compresses the production time, and the temptation is to use the speed to finally build the parallel tracks. It's worth resisting, because parallel tracks were the wrong target all along: they quietly sort students into different lessons, and the students holding the "easy" packet know exactly what it means.

The sturdier frame is Universal Design for Learning, CAST's framework for designing instruction "to improve and optimize teaching and learning for all people based on scientific insights into how humans learn." UDL's central move is to locate the barrier in the material rather than the student — differentiate the entry points and supports around one common lesson, not the destination. For second period: everyone reads the Triangle fire article and everyone writes the argument paragraph. What varies is the glossary in the margin, the frames under the prompt, the depth of the questions.

What that looks like group by group is the rest of this article.

How do you scaffold a text down without gutting it?

The instinct with the five students reading near 4th-grade level is to ask AI to "rewrite this at a 4th-grade reading level." Sometimes that's the right call. Often it's the wrong first move, for a reason hiding inside the assignment: the task requires students to cite the article's evidence, and a rewrite replaces the evidence. When AI simplifies the sentence about the owners' "negligence," the word negligence — the term the strongest claims will hinge on, and exactly the academic vocabulary RI.7.1's assessments assume — is usually the first casualty. A leveled rewrite can strip the very language the standard requires.

Two more cautions before the prompts. AI readability claims are estimates, not measurements — no general chatbot is computing a Lexile score, so "now at a 4th-grade level" means "simpler than before, probably." And AI has no idea why each of your five students reads below benchmark; a decoding gap, a vocabulary gap, and an attention gap want different scaffolds, and that diagnosis stays with you.

The scaffolds that keep students in the real text:

Margin glossary: "Here's the 1,100-word article my 7th graders are reading on the Triangle Shirtwaist Factory fire [paste]. Identify the 10 words most likely to stop a reader at about a 4th-grade level — include 'negligence' and 'garment' — and build a margin glossary: each word, a plain-English definition of 10 words or fewer, and the sentence where it appears. Do not change the article itself."

Chunk and check: "Break the same article into five chunks at natural topic shifts. Above each chunk, add one line telling the reader what the section will explain ('This part explains why the exit doors were locked'). After each chunk, add one literal check-in question answerable from that chunk alone."

Sentence frames: "My students are writing an argument paragraph on who bears responsibility for the Triangle fire deaths, citing the article's evidence (RI.7.1). Write two sets of frames: a base set for claim-evidence-reasoning ('The ___ were most responsible because the article says ___, which shows ___') and a stretch set that forces counter-argument ('Although some blame ___, the stronger evidence points to ___ because ___')."

Each of these changes the support around the text without touching the text, so the five students cite the same sentences as everyone else.

What about the two newcomers?

The strongest guidance here predates AI. WIDA's framework for multilingual learners is explicit that "unit language and content learning goals are not differentiated" — what differs is the scaffolding each student needs to reach them. The failure mode AI makes newly easy is the opposite: a separate, simplified, fully translated packet that removes a newcomer from the class's work entirely. It takes thirty seconds to generate, and it builds a second classroom inside your classroom.

The better use is targeted bilingual support around the same article:

Newcomer support: "Two of my 7th graders are newcomer Spanish speakers around WIDA level 2. For this article [paste], build a two-column English–Spanish glossary of the 15 highest-value words and phrases, and reword the three discussion questions in simpler English sentence structure — same content, same rigor. They are reading the same article as the class; do not summarize or replace it."

AI translation is good but not reliable, and a wrong cognate on a printed glossary is an error two students will carry for the whole unit. If your building has a bilingual aide, an ELL teacher, or a fluent colleague, ask for ninety seconds on the glossary before it goes to the copier.

What do early finishers get, if not more pages?

The seven students who finished early are the group differentiation guides usually wave at last, and the AI failure mode with them is specific: ask for an "extension" and most tools generate more — more questions, a bonus essay, a poster. A longer packet is not a harder task, and strong students recognize busywork the moment it lands on their desk.

A real extension raises the cognitive demand on the same deliverable. For second period, the useful historical wrinkle is that the factory's owners were tried for manslaughter and acquitted — which turns a clean argument into a contested one.

Extension: "Seven of my 7th graders can already argue who was responsible for the Triangle fire. Using the historical fact that owners Harris and Blanck were tried for manslaughter and acquitted, write an extension task: they must revise their argument paragraph to address the acquittal as counter-evidence, plus three coaching questions to push the revision ('What would the defense have said about the locked doors?'). Do not add new comprehension questions — the deliverable is the same paragraph, made harder to write well."

Same rubric, same paragraph, so your grading load stays flat. What to do with feedback on 29 versions of that paragraph is its own workflow — that's covered in our guide to AI grading and feedback.

Can AI help with IEP and 504 accommodations?

Separate two jobs. Writing an IEP is a legal process owned by the student's team; AI has no role in drafting it, and identifying student information should never enter a general AI tool at all. Implementing accommodations that already exist — chunked assignments, simplified directions, guided notes — is materials production, and that is squarely AI's lane. Teachers seem to feel the distinction: in the 2025 Gallup–Walton Family Foundation survey, 57% of teachers agreed AI will improve the accessibility of learning materials for students with disabilities — 65% among special education teachers.

The technique that keeps this safe: prompt about the accommodation type, never the student. "A student who…" becomes "an accommodation that…" — the materials come out the same, with none of the exposure. (The full student-privacy checklist, including what counts as identifying, is in our AI for teachers guide.)

Accommodation materials: "One accommodation I implement is 'assignments broken into chunks with a visual checklist.' Convert this two-day argument-writing task [paste] into that format: no more than seven steps per day, each step under 15 words with its own checkbox, and a 'done when' line for each ('Done when: my claim names who was responsible'). Keep the task itself unchanged."

One limit worth stating plainly: whether the checklist version actually satisfies the accommodation as written is a judgment that stays with you and the case manager. An IEP meeting will not accept "the AI formatted it" as evidence of implementation — you sign off on the fit, exactly as if you had made it by hand.

Leveled questioning: one text, four depths

The cheapest differentiation in this article is asking better-graded questions of a single text. Instead of separate worksheets, one ladder:

Question ladder: "Write four questions about this article [paste], one at each Webb's Depth of Knowledge level, all answerable from the text: DOK 1 (recall a stated fact), DOK 2 (explain a relationship the article implies), DOK 3 (evaluate which piece of evidence is strongest and defend the choice), DOK 4 (connect the fire's aftermath to a broader principle about regulation). Label each, and note what a strong answer must include."

In second period the ladder runs one discussion at four depths: the newcomers and striving readers take DOK 1–2 with glossaries in hand, the middle of the room argues DOK 3, the extension group takes DOK 4, and anyone can reach up a rung. The same structure works for the end-of-class check; that version, and how to read what comes back, is in our AI exit tickets guide.

Second period, differentiated

Here's the full inventory the week requires. The article, untouched. A margin glossary and a chunked copy for five students. A bilingual glossary and reworded questions for two. Sentence frames in two strengths, on the back table for anyone. A counter-evidence revision task for seven. One question ladder for the discussion, and a checklist version of the task for one accommodation on file. Producing that by hand is most of a weekend; with the prompts above it's under an hour of generating, plus the part that doesn't compress — reading every line before students see any of it. All 29 students work in the same article, on the same paragraph, against the same standard. The five striving readers cite the same sentences as everyone else; the seven early finishers rewrite their paragraphs instead of lengthening them; and the checklist goes into the accommodation folder for the case manager to review.