AI Differentiation for Mixed-Ability ESL Classrooms

Walk into almost any ESL classroom and you’ll find the same quiet problem: one lesson, ten different reading levels. A single passage that challenges your strongest student bores your weakest one into silence, and a worksheet pitched at the middle of the class leaves both ends underserved. For years the only real fix was brute-force teacher labor — rewriting the same passage three times, building three vocabulary lists, drafting three sets of comprehension questions. Most teachers didn’t have the hours, so differentiation quietly got skipped. AI has changed the math on that trade-off, not by replacing the teacher’s judgment about what students need, but by collapsing the time it takes to produce leveled materials from hours to minutes.

Why Differentiation Is the Hardest Part of ESL Teaching

Differentiated instruction sounds simple in teacher training: assess where each student is, then adjust content to meet them there. In practice, ESL classrooms make this unusually hard because “level” isn’t one variable. A student might read fluently but struggle with listening, or handle vocabulary well but freeze on grammar-heavy questions. Multiply that by a class of twenty or thirty students from different first-language backgrounds and different prior schooling, and true differentiation becomes a full-time curriculum-writing job layered on top of an already full-time teaching job.

Most teachers respond by picking a middle-ground text and hoping stronger students self-extend while weaker ones lean on peers or context clues. It’s a reasonable survival strategy, but it isn’t differentiation — it’s compromise. The gap between what differentiation research recommends and what a single teacher can produce by hand is exactly where AI tools have found their footing in language classrooms.

What AI Actually Does Well for Differentiation

Large language models are genuinely good at one specific skill that differentiation depends on: rewriting the same content at different levels of linguistic complexity while preserving the core meaning. Give a model a 200-word passage about renewable energy and ask it to produce an A2-level version, a B1-level version, and a B2-level version, and it will reliably simplify sentence structure, swap in higher-frequency vocabulary, and shorten sentences for the lower levels — all while keeping the same topic and most of the same factual content. That’s the exact task that used to eat a teacher’s Sunday afternoon.

Reading Passages at Multiple Levels

The clearest win is leveled reading. Instead of choosing one article for the whole class, a teacher can take a single authentic or teacher-written text and generate two or three parallel versions matched to CEFR bands. Because the underlying content stays the same, the whole class can still discuss the same topic together afterward — students just accessed it through text pitched at their own level. That preserves the social and communicative parts of the lesson while removing the reading-level bottleneck.

A classroom whiteboard with educational posters, job charts, and student name tags

Vocabulary Scaffolding

The same principle applies to vocabulary lists. A model can pull the key terms out of a passage and generate three tiers of support: a glossed list with first-language cognates or simple synonyms for beginners, a list with English definitions and example sentences for intermediate students, and a list of collocations and nuanced usage notes for advanced students. This turns one vocabulary-building exercise into three, without three separate research sessions.

Adjusting Question Complexity, Not Just Vocabulary

The part teachers often miss when they first try this is that differentiation isn’t only about simplifying words — it’s also about the cognitive demand of the questions themselves. A beginner-level comprehension task should stick to literal recall (“What did the article say happened first?”), while an advanced version can ask for inference, evaluation, or opinion (“Why might the author have chosen to present the information in this order?”). AI tools can generate both question types from the same source text if you specify the thinking skill you want, not just the reading level.

students in classroom with teacher presenting

A Practical Workflow: From One Lesson to Three Levels

The workflow that holds up best in real classrooms is simple enough to run in under fifteen minutes once you’re used to it. Start with a single source text or lesson objective, then use AI to branch it into leveled versions before you ever touch a photocopier.

Write or choose one strong source passage at your class’s middle level — this becomes the anchor text everyone discusses together.
Ask the AI tool to rewrite it at one level below and one level above, specifying CEFR band or grade level explicitly rather than vague terms like “easier.”
Generate matching comprehension questions for each version, specifying literal recall for the lower level and inference or evaluation for the higher level.
Read all three versions yourself before printing — check that facts, names, and numbers weren’t altered in the simplification process.
Distribute by table group or individually, but keep the discussion phase whole-class so everyone engages with the same ideas.

That last step matters more than it looks. The point of AI-assisted differentiation isn’t to split your class into three silent, isolated tracks — it’s to remove the reading-level barrier so the whole class can still do the collaborative, communicative work that actually builds language proficiency.

Children in a Classroom. In the back of a classroom, are children about 11 years old with a female teacher talking about the

Real Classroom Use Cases: TOEIC and IELTS Prep

Exam prep classes are where mixed ability shows up most painfully, because students are often grouped by target score rather than true current level, and a single class can span a 200-point TOEIC gap or a full IELTS band. Differentiated AI-generated materials are particularly useful here because exam-style tasks have a rigid, predictable format that models handle reliably.

Leveling Practice Passages for IELTS Reading

For IELTS Academic Reading practice, you can take one authentic-style passage and generate a slightly shortened, lower-vocabulary version for students still working toward Band 5.5, alongside the full-length original for students targeting Band 7 or above. Both groups practice the same question types — True/False/Not Given, matching headings, summary completion — so the exam skill being drilled stays constant even though the input text difficulty doesn’t.

A woman in a grey sweater using a tablet at a wooden table

Custom TOEIC Vocabulary Drills by Level

For TOEIC classes, ask the AI to generate business-context vocabulary drills at two tiers: high-frequency Part 5 vocabulary for students below 600, and more idiomatic collocations and phrasal verbs for students pushing toward 850+. Because TOEIC vocabulary is narrowly business-and-workplace themed, models are especially reliable here — they rarely drift into unrelated topics the way general-purpose vocabulary generation sometimes can.

Guardrails: Keeping AI-Differentiated Content Accurate

None of this works if the leveled versions quietly change facts. When a model simplifies a sentence, it sometimes drops a qualifier, flattens a date range, or swaps a number for a rounder one — the kind of small edit that doesn’t look wrong until a sharp-eyed student in the advanced group notices the two versions of the passage disagree. The fix is a five-minute read-through of every generated version before it reaches students, comparing it against the original for factual drift. This is also the moment to check that idioms or cultural references were simplified sensibly rather than replaced with something that doesn’t map onto your students’ background knowledge.

It’s also worth being explicit with the AI tool about tone. Simplification prompts that just say “make this easier” often produce flat, choppy sentences that read like a decoding exercise rather than real English. Asking instead for “natural sentences a B1 learner would encounter in everyday reading” tends to produce more usable, less robotic output.

A classroom with supplies, desk, and a chalkboard

Where AI Differentiation Breaks Down

AI differentiation is strongest with text-based skills — reading, vocabulary, and written comprehension questions. It’s noticeably weaker for differentiating speaking and listening tasks, where pacing, pronunciation feedback, and real-time interaction matter more than text complexity. It also can’t judge classroom dynamics: it doesn’t know that pairing your two weakest readers together will stall the activity, or that a particular student needs a confidence win more than an accuracy challenge this week. Those calls stay entirely with the teacher. AI is best understood here as a drafting tool that removes the mechanical bottleneck of rewriting content three times, not as a system that decides what any individual student needs.

Getting Started This Week

You don’t need a new platform or a training session to try this. Take one reading passage from your next lesson, open whatever AI chat tool your school already permits, and ask for a version one CEFR level down and one level up, plus matching questions at appropriate cognitive demand for each. Read all three before class. That single fifteen-minute experiment is usually enough for teachers to see whether the workflow fits their classroom — and most who try it once keep doing it, because the alternative was never writing three versions by hand in the first place; it was writing one and hoping it worked for everyone.