Writing for readers who are not you

The register problem

Most professional documents fail their readers not because they are wrong but because they are pitched at the writer. A discharge letter written for a doctor. A tenancy notice written for a lawyer. A privacy policy written for a regulator. The reader is a patient, a tenant, a customer.

Moving a document between registers is genuinely one of the strongest things these tools do, and the reason is the one from the first lesson: the truth is in the text. All the facts are in the document you paste. Nothing needs to be retrieved. The job is rephrasing, and rephrasing is what the technology is.

Rewrite this discharge letter for a reader with no medical training and a reading age of about nine. Keep every instruction and every date exactly as they are. Where a medical term must stay, explain it once in brackets.

Two details in that prompt do the work. Naming a reading age is far more actionable than "simple". And instructing that dates and instructions stay exact tells you where the danger is — simplification loses precision, and precision is what a discharge letter is for.

Useful targets

• A reading age of nine is what GOV.UK aims at for public content. It sounds patronising until you see that it mostly means shorter sentences and common words, not a simplified message.
• One idea per sentence, one topic per paragraph. More useful than any readability score.
• Numbers as digits, not words, and dates written out — "3 March 2026", not "03/03/26", which means two different days on two continents.
• Active voice with a named actor. "You must reply by 3 March" survives translation and stress; "a response is required" does not.

Free checkers exist and are worth ten seconds: LanguageTool for grammar and clarity, and the built-in readability statistics in Word or LibreOffice.

Translation: strong, uneven, checkable

For the major world languages, machine translation is now good enough for real work, and AI models add something dedicated translators historically lacked: you can give them context. This is a letter to a patient's family; keep it formal but gentle; this term is a legal term of art and must not be paraphrased. That instruction genuinely changes the output.

The unevenness is the part to hold onto. Quality tracks how much of a language was on the internet. Spanish, French, German, Mandarin, Portuguese and Arabic are strong. Hindi, Bengali, Vietnamese and Indonesian are decent and slipping on idiom. Many African and indigenous languages, and most regional variants, are weak — and the output is equally fluent in all cases, so the reader cannot tell which they got. That is the same lesson as invention: confidence is uniform, accuracy is not.

The back-translation check

If you cannot read the target language, one check gets you most of the way:

1. Translate your document into the target language.
2. Take the result to a different tool and translate it back into your language.
3. Read the round trip against your original.

Meaning that survives two machine passes is usually sound. What this catches, and it catches it reliably: reversed negations ("you may not appeal" becoming "you may appeal"), altered numbers, dropped clauses, and terms of art turned into ordinary words. What it does not catch: tone, formality level, and the specific register that tells a reader whether they are being addressed with respect. For anything with consequences — consent, legal notice, safety instructions, dosage — a competent human speaker reads it before it is used. The machine draft makes that person's job twenty minutes instead of two hours; it does not replace them.

The free path is real here: DeepL's free tier, Google Translate, and any chat model. Using two different ones for the round trip is the point, not an inconvenience.

Accessibility, which is the same skill

The same move covers work most people never think to ask for.

• Alt text for images in documents and on the intranet: describe the image, ask for one sentence naming what matters, then check it says the thing you would have said.
• Plain-language versions alongside formal ones, rather than instead of them. Publish both and the formal document keeps its legal precision.
• Reading a dense document aloud through a screen reader or text-to-speech, which many colleagues rely on and which exposes bad structure to everyone else.
• Making your own writing easier for you. Colleagues with dyslexia, and anyone working in their third language, describe this as the single most useful thing on the list — not because it writes for them, but because it removes a tax they were paying that nobody else could see.

That last point is worth saying without decoration. A great deal of professional judgment goes unheard because it arrives in imperfect prose. If a tool closes that gap, the gain is not efficiency. It is that the right person gets listened to.