back to the handbook

translating and rephrasing

Reviewed June 2026

Moving one idea across languages, reading levels, and tones is the kind of low-stakes, high-volume work a model handles well. The same sentence can become a parent letter in Spanish, a sixth-grade version of a dense reading, or a calmer reply to a pointed email. The catch is the usual one: each move can quietly change the meaning, so the skill worth teaching is checking whether the new version still says the true thing.

Below are two of these moves to try. The first keeps a text's shape while changing its language. The second rephrases a text for a different reader, which is where this turns into a real accessibility tool. Run each one, then read the output with a critical eye.

translate, and keep the shape

try it: change the language

Pick a target, then run the translation. The third tab asks the model to translate a poem and hold its three-line structure, which a plain translation tool can't do.

Adding "preserve the structure" is what separates a translation tool from a model you can talk to: you can follow up, add context, and steer the result toward what you meant. The same move even crosses into code. Ask for a Python script that prints a line and you get back print("Let's restore humanity to education."), which is really just a translation from plain English into a programming language.

rephrase for the reader

try it: a new form, same idea

Same idea, different reader. Drop the reading level, break a task into steps, or take the heat out of an email. Run each and watch what the model keeps and what it risks dropping.

this is an accessibility tool

For a lot of students these moves decide whether a text is reachable at all. A multilingual learner can meet a hard reading in their stronger language first, then come back to the English with the ideas already in hand. A student with dyslexia or a language-processing difference can read the same content at a level that doesn't spend all their energy on decoding. Breaking a task into small, ordered steps takes weight off working memory, which helps students with ADHD and helps anyone who is tired or new to the material.

This is the reasoning behind Universal Design for Learning. Offer more than one route into the same content and more students arrive, without a separate watered-down track for the kids who needed a hand. A model makes those routes cheap to produce in the moment a student is stuck, rather than weeks ahead on a worksheet nobody ended up needing.

Autistic students and adults increasingly reach for these tools on their own. A blunt message gets softened to the tone a situation expects. An email thick with social subtext gets decoded into what it is actually asking for. An open-ended assignment gets turned into a concrete list of moves. In a Carnegie Mellon study, autistic workers leaned on ChatGPT for workplace questions and preferred its plain, direct answers to softer human advice, and many described using it to get their own thoughts into words other people would read the way they meant them. Rephrasing here works as a translator between communication styles, taking on labor that used to fall entirely on the neurodivergent person.

The handbook's caution still applies, with an extra edge. A model can flatten meaning when it simplifies, soften a translation until it is wrong, or drop the step that mattered. It will also hand over confident social advice that misses badly: in that same study it told an autistic user who wanted to make friends to just walk up to people and start talking. The researchers draw the obvious conclusion, which is that autistic people should help design the tools being sold to them.

Further reading: British Council, on AI tools for neurodivergent learners; the UDL Guidelines (CAST); Carnegie Mellon, on autistic workers and ChatGPT; and the study "I use ChatGPT to humanize my words."

put students in the editor's seat

try with students

A rewrite is easy to get. Judging whether it still tells the truth is the harder, more useful skill, so put students in charge of that call. Hand them a real text and let them be the editor the model needs.

  1. rephrase something real 5 min

    Take a dense paragraph from your current unit. In pairs, students prompt a model to rewrite it one way: a lower reading level, a home language, or a numbered set of steps.

  2. hunt for what changed 10 min

    Set the original beside the rewrite. Each pair marks one thing the new version made clearer and one thing it lost, softened, or got wrong, quoting both. Linger on the spots where pairs disagree about what still counts as true.

  3. fix it by hand 5 min

    Students edit the AI version by hand until it reads clearly and still holds up. What they turn in is that edited version, the one they're willing to stand behind.

  4. share the keeper 5 min

    Each pair reads their fixed version and names which changes helped access and which crossed the line into changing the meaning.

This pairs naturally with make a standard student-friendly, which is the same move aimed at the standards themselves.

back to the handbook