How to Differentiate One Reading with MagicSchool AI

November 16, 2026

Differentiation is the highest-value use of MagicSchool AI, and the one most likely to blow up in class. Levelling a text is easy. Levelling it without introducing false facts is the part that needs a review pass.

This walkthrough turns one article into three versions and one question set, in about twelve minutes.

Why differentiate at all in 2026?

Class reading spread is real. In a typical middle school class, reading comprehension spans several grade levels, and handing everyone the same article means part of the room cannot access it and part of the room is bored.

Three versions is usually enough. More than that and you cannot manage the group work.

How do you level one text three ways?

Start with a clean source article in the public domain or one you are licensed to adapt. Then:

  1. Open the text leveller tool and paste the whole article.
  2. Set grade levels explicitly: one at grade level, one two grades below, one three to four below.
  3. Ask it to preserve the key vocabulary and change sentence structure, not just word length.
  4. Generate each version separately rather than in one call — one call tends to produce three versions that differ only in vocabulary.
  5. Read each version against the original.

Point 3 matters. If the simplified version drops technical terms, the students reading it cannot participate in the same discussion as everyone else. Preserve the terms and simplify the syntax around them.

How do you check accuracy before it reaches students?

Read the lowest-level version line by line against the source. Simplification is where meaning breaks.

Common failures:

  • A causal statement becomes a sequence. “Because the enzyme denatured, the reaction stopped” turns into “the enzyme changed and then the reaction stopped,” which is now wrong about mechanism.
  • Quantities get rounded or dropped. “A 2.5 percent increase” becomes “a small increase,” which is fine, or “a 25 percent increase,” which is not.
  • Nuance disappears. “Suggests” becomes “shows.”

That third one is the most damaging in science and social studies texts. An AI-levelled passage that upgrades hedging to certainty teaches students the wrong thing about how claims work.

Budget fifteen minutes for review on a 700-word article. It is non-negotiable; the failure mode is a parent email.

How do you build scaffolded questions?

Generate one question set, then assign different subsets rather than generating separate sets per level. A shared question pool keeps the class discussion coherent.

A structure that works:

Student group Questions assigned
Reading below grade level Recall and vocabulary questions
At grade level Inference and main-idea questions
Above grade level Evaluation and evidence-based argument questions

Ask the generator for questions tagged by cognitive level, then sort them yourself. The tool can label levels, but the labelling is inconsistent enough that checking is faster than fixing.

Ensure at least one question in the set is answerable from every version of the text. Otherwise the group work splits into students who cannot contribute.

What about IEP language and parent communication?

Drafting present-level statements and progress notes is where these tools save the most real hours, but the output must be reviewed by someone who knows the student and the legal requirements.

Use it to produce a first structure — the categories and phrasing conventions — then rewrite with actual observations. Generated statements read as fluent and generic, which is exactly the kind of note that gets challenged.

Never submit generated text about a specific student without a person confirming every claim in it.

Where it falls short

Text levelling is uneven across subjects. Narrative and expository prose simplify well; primary source documents, poetry, and dense technical text do not.

It cannot see your students’ IEPs, so its accommodations are generic. And it produces questions in a recognisably AI voice, which students who have used chatbots notice.

Quick Summary

  • Generate each reading level in a separate call, preserving key vocabulary while simplifying syntax.
  • Review the lowest version line by line; simplification breaks causality and turns hedges into claims.
  • Generate one question pool and assign subsets, not three separate sets.
  • Budget about 15 minutes of review per 700-word article.
  • Use it to structure IEP notes, never to submit unreviewed statements about a student.