Teachers search for this constantly: how to shrink a text's Lexile level, or how to make the same article readable for a class where students read at four different levels. AI tools make the rewrite trivial. The harder part is knowing whether the result suits your students, and that is where a shortcut can go wrong. A passage can arrive at "grade 5" on paper and still be a wall for the particular ten-year-olds in front of you.
A 2026 study from the University of Tsukuba, published in Reading and Writing (summary), is a useful reminder. The researchers had 41 university students who were learning English read 30 passages. They measured difficulty two ways: objectively, using readability formulas and a large language model, and subjectively, using each reader's own nine-point difficulty rating. The two lined up closely, since passages scored harder by the formulas felt harder and were read more slowly. But each reader's own rating also explained their eye movements beyond what the text score could. In other words, a level number tells you about the text, and only the reader can tell you about the reader. That is the idea behind this guide: use AI to make the draft, then use your students to check it.
What does a reading level number actually measure?
Readability scores such as Flesch-Kincaid and Lexile are calculated mostly from surface features, namely how long the sentences are and how common the words are. That makes them useful for a rough match between a text and a reader, and it also explains their blind spots. They do not measure background knowledge, so a text about the water cycle is easier if you have watched a kettle boil. They do not measure whether the vocabulary that matters is familiar, and a short sentence with an unfamiliar key term can still stop a reader.
They also miss structure, since short sentences can be badly organised, and interest, since students read harder texts about things they care about. A multilingual learner may find idioms and long noun phrases harder than the formula suggests. So treat the number as a starting point. It is a reasonable first filter and a poor final judge.
Should you simplify the text or scaffold it?
Before you prompt an AI, decide which job you want done, because they are different. Simplifying means rewriting so that the text is easier to read, with shorter sentences and simpler words, and it is right when reading is the obstacle and the content is not. Scaffolding means keeping the original text and adding support, such as a glossary, a summary beside each paragraph, questions after each section and sentence starters. It is right when you want students to meet the real vocabulary and ideas.
Over-simplifying can take away the very words and ideas students need to learn. For core texts, scaffolding is often the better choice, because students still read what everyone else reads, with help along the way.
Which prompts work?
To simplify while keeping the content, name the facts and terms to preserve and forbid additions.
Rewrite the passage below for students reading at about a grade 5 level. Keep every fact and every key term (list them: photosynthesis, chlorophyll, glucose), and define each key term in a short phrase the first time it appears. Use sentences under 15 words where you can. Do not add facts that are not in the original. Passage: [paste]
To scaffold without rewriting, ask for the supports and tell the AI to leave the passage alone.
For the passage below, create: (1) a glossary of the 8 hardest words with student-friendly definitions, (2) a one-sentence summary for each paragraph, and (3) four questions, from literal to inferential. Do not change the passage. Passage: [paste]
To make several versions for a mixed class, keep the structure the same across all of them.
Give me three versions of this passage: one at grade 4, one at grade 6, and the original. Keep the same facts and headings in all three so students can discuss the same content.
Using the same headings across versions lets students of different levels work in one discussion, which matters as much as the reading level itself.
What should you check before you hand it out?
AI rewrites fail in predictable ways, and it takes about two minutes to catch most of them. First ask whether it is still true. Simplification can quietly change a claim, so "some plants" becomes "plants," or a cause becomes a correlation, and that matters most in science and history. Next check that the key terms are still there, because if students need a word for the test it should appear with a definition. Compare the facts in the original and the rewrite to see whether anything was dropped or added.
Read it aloud for tone. Some rewrites become choppy and childish, which older students resent, and a grade 9 student reading below level still needs a text that respects grade 9. Look for idioms and local references that will trip up multilingual learners. If you plan to confirm the level with a readability formula, remember that it is a rough guide, since long technical words can push a score up even when the passage is simple in structure, and the reverse is also true.
How do you check it with real readers?
This is the step the research supports most. Before you use a version with a whole class, try it with two or three students at the target level. Ask for a difficulty rating on a scale of 1 to 5. Ask them to retell the passage in their own words, which shows understanding better than a rating does. Ask them to underline any word or sentence they had to re-read, and notice how long it took, because a much longer reading time than expected suggests the text is harder than its label says. Two minutes with two students will tell you more than any score.
What extra support do multilingual learners need?
Adjusting the reading level is not the same as supporting a language learner. Helpful additions include a glossary with words in both languages where you can provide one, sentence frames for responses such as "This paragraph is mainly about...", diagrams beside the key ideas, and a chance to talk about the text in pairs before writing. You can ask an AI to draft these, and it helps to check them with someone who knows the language, because translations from AI can contain mistakes.
What should you be careful about pasting?
Do not paste student work, names or other identifying details into a general AI tool. Public articles, textbook passages you have the right to use, and text you have written yourself are fine to adapt, but check copyright and your school's rules before rewriting a published text and sharing it widely.
What does the whole workflow look like?
Choose the goal first, simplify or scaffold. Prompt with the facts and key terms to preserve, then check truth, key terms and tone. Try the result with two students and ask for a rating and a retell, then adjust and keep the version that worked for next year. Over a term you build a small library of texts you have tested with real readers, which is worth more than a folder of untested rewrites.
Frequently asked questions
Is it fine to use Lexile scores for grouping students? They are a rough guide, and they work better as one input than as a label. Combine them with what you see when students read.
How much simplification is too much? If key vocabulary disappears or the ideas get thinner, you have gone too far. Aim to change how the text is expressed, not what it teaches.
Can I ask the AI what reading level a text is? It will give an answer, but it can be wrong. Use a readability tool for a rough check and your students for the real one.
What about texts in other languages? The same method applies, but have a fluent speaker check the result, because AI mistakes are harder to spot if you do not read the language.