Why AI English Sounds Translated (and Grammar Isn't the Problem)

TL;DR: AI English sounds translated because the model writes the statistical average of English, and the average of English has the shape of a careful translation: every sentence about the same length, a connector at the head of each paragraph, hedges stacked before every ask, no contractions, a formula at each end. A native reader hears the rhythm before they check the grammar. Asking for "natural" or "native" English changes nothing, because those words describe an effect the model is already aiming at. What works is measuring your own English habits from emails you have sent and handing them to the model as rules. Every new model executes those rules better than the last one did.

You ask ChatGPT to polish an email to a client in London. Every sentence comes back correct. You read it twice, find nothing wrong, and still feel that it sounds translated. You send it. The reply is polite and slightly more distant than last week's.

Nothing in the draft was wrong. Nothing in it was yours, either.

Why Does AI English Sound Translated?

Because the model writes the centre of all the English it has seen, and the centre of English has the shape of a translation. Translated prose is recognisable by structure long before vocabulary: sentences of nearly equal length, a linking word at the top of each paragraph, more cushioning than the ask needs, no contractions, and an opening and closing that could be swapped into any other email without loss. AI produces that structure by default, whether the request came in Japanese, French or English.

The flatness is measurable. Our Average AI baseline, built from 320 samples across five models and four languages, puts English AI output at a consistency score between 52 and 55, and the figure barely moves across languages. In our scale that means moderate, evenly distributed sentence-length variation: the model neither writes the six-word sentence that lands a point nor the thirty-word one that carries context. How we built that baseline explains the method; the short version is that the rhythm is the same everywhere, which is exactly what a reader means when they say a text "sounds translated".

Here is what the average shape looks like on a single email, and what a fluent reader notices in each row:

Every row is a placement or rhythm choice. A grammar checker passes all seven. Your client does not.

What Happens When the Draft Starts in Japanese or French?

The model fixes the sentences and keeps the architecture. Paste a Japanese draft and ask for English, or write English with a Japanese plan in your head, and the model corrects each sentence while preserving the order you gave it: relationship first, circumstances second, apology third, the ask last, gratitude at the end. Correct English arranged in a Japanese order is the most precise definition of "sounds translated" that exists.

Take an account manager in Tokyo writing to a client in London about a delivery that has slipped a week. The Japanese logic of that email is sound: acknowledge the relationship, explain the circumstances, apologise, state the new date, ask for understanding. AI English of the same plan comes out as: "I hope this email finds you well. I am writing to inform you regarding the delivery schedule. Due to unforeseen circumstances in our production process, we regret to inform you that the delivery will be delayed." The date arrives in paragraph two. The ask never quite arrives at all.

The same person, writing in her own English on a good day, sends something closer to: "Quick update on the delivery: it's moving from the 20th to the 27th. The cause was a supplier delay on our side, and we've already switched to a second source. Sorry for the knock-on effect on your planning. Does the 27th still work for your team?" Same facts, same apology, same politeness. The ask sits in the first line and the cushioning is one sentence long.

The French version of the problem is structural in a different way. Our per-language data shows French AI output scoring 75 on sentence complexity and 32 on conciseness, the highest and lowest of the four languages we measured. A French product lead who drafts in French and asks for English gets the vocabulary smoothed and the subordinate clauses kept: an English sentence with a French skeleton, forty words long, the main verb somewhere in the middle. How AI writes differently in Japanese, French and Spanish has the full table.

One thing to be clear about: the fix is not to make either writer sound like a Londoner. The account manager's English is its own thing. She does front-load context, in two short sentences rather than five long ones, and she does apologise once, after the fix is stated. Those are her habits. The translation problem is what the model adds on top of them, and the guide for bilingual professionals covers why that addition hits second-language writers hardest.

Why Doesn't "Write in Natural English" Fix It?

Because "natural", "native", "fluent" and "conversational" are labels for an effect, and the model is already aiming at the effect. Those words describe millions of writers and none of their habits. AI does not know your tone; it knows words that describe tone. Ask for natural English and you get the average of everything the training data labelled natural, which is the same draft with two contractions added.

"Professional yet friendly" has the same problem, and it is the phrase most people reach for. It describes thousands of writers. Your colleague who types the same three words into the same box gets the same email you do. A voice does not live in adjectives. It lives in sentence length and how much it varies, in where the qualifiers sit, in which transitions you use and which you never use, in how you punctuate. We wrote about this at length in Adjectives Are Not Your Voice, and the argument holds with more force for a second language, because the adjectives you would choose for your English are the ones the model already defaults to.

OpenAI's own custom instructions guidance frames the feature as two questions: what should the model know about you, and how should it respond. Most people answer the second with a mood. The answer that changes the draft is a rule: open with the ask, one hedge maximum, contractions on, no closing formula. A rule can be followed. A mood can only be approximated, and the approximation is the average.

Will a Better Model Stop Sounding Translated?

No, and the reason is structural rather than temporary. Model capability and user-data conditioning are separate layers. Labs improve the first with every release and cannot supply the second, because the second is built from emails you sent and the lab has never seen them. A stronger model follows "natural English" more faithfully, which produces a smoother version of the same average shape.

Our own per-model data makes the point in numbers. Across the 80-sample English slice of the baseline, the five models we measured spread 61 points on expressiveness and sit nearly on top of each other on two other axes. Switching from one model to another changes the energy of the draft. It does not change the placement of the ask, the hedge count or the rhythm, because none of the models has any information about yours. What the Average AI baseline hides breaks the numbers down by model.

The same separation works in your favour once you have a profile. The better the model gets at following nuanced instructions, the better it executes your profile too. "Open with the update, one hedge per ask, at least one sentence under eight words per paragraph" is a nuanced instruction, and each generation carries it out more accurately than the last. Give a vague adjective to a stronger model and you get a smoother average. Give it a measured profile and you get a more accurate you. Which one you receive depends entirely on what you handed over, a point we made in Why Better Models Still Don't Know Your Voice.

How Do You Give AI Your English Instead of the Average?

Measure your habits from English you have already sent, write them down as rules, and put the rules where the model reads them on every request: ChatGPT custom instructions or a Project, a Claude Project, a Gemini Gem. Three steps, and the first two need no software.

Step 1: collect ten sent emails in one register. Client-facing, internal, or first contact; pick one. Choose the ones you were content to send, not the ones you rewrote for an hour, because the profile should capture your normal Tuesday and not your best day. Keep them separate from anything you wrote in Japanese or French. Your English is not a translation of your other languages and should not be measured as one.

Step 2: count, do not describe. For each email, note:

• Shortest sentence, longest sentence, and the typical length. Most people are surprised by how short the shortest is.
• Where the ask sits: sentence one, or after a paragraph of context.
• Hedges before the ask: which phrases, and how many. "Could you" and "I was wondering whether it might be possible" are not the same hedge.
• Contractions: present or absent.
• Your actual openers and closers, including "none".
• Connectors you use ("also", "so", "and") and connectors that never appear in your writing.
• The apology pattern: how often, how concrete, before or after the fix.

Step 3: write the counts as rules. A finished block for one register looks like this:

English email, client-facing:
- Open with the update or the ask in sentence one. Context comes after.
- Vary sentence length: at least one sentence under 8 words per paragraph, none over 30.
- One hedge per ask, maximum. Use "could you". Never "I was wondering if".
- Contractions on: it's, we're, I'll.
- Connectors: "also", "so", "and". Never "in addition" or a linking adverb to open a paragraph.
- No opening formula. Close with the next step and my first name.
- Apologise once, concretely, after the fix is stated. Never "we regret to inform".

That block is not a description of a mood. Every line can be checked against the output, and every line came from something you actually wrote.

A five-minute test tells you whether an AI draft passed. Put the draft next to one of your own emails. Underline every sentence under eight words in each. Circle the ask in each. Count the hedges before it. If the AI draft has no short sentences, the ask in paragraph three and three hedges in front of it, you are reading the average, and the average sounds translated.

MyWritingTwin does the counting for you. A Writing DNA analysis measures your sentence rhythm, hedge placement, connectors and formality from your own samples, per language and against a per-language Average AI baseline, and turns the result into a Style Profile you carry to ChatGPT, Claude and Gemini. Your English profile and your Japanese profile are built separately, from separate samples, because the register shifts that matter in Japanese business email are different habits from the ones that matter in your English, and a model should not be allowed to average them. If you write in more than one language, Your Writing Voice Changes When You Switch Languages explains why one profile is never enough.

---

If you want to see the shape of your own English against the average before writing a single rule, the free Writing DNA Snapshot measures a few of your samples against the per-language AI baseline and shows you where the model would have flattened you. Submit English samples and Japanese or French samples separately, and you get two charts, not one average of both.