Prompts & Writing
Why AI Writing All Sounds the Same, and How to Break the Pattern
By Jim Vernon, Editor, AI Intelligence International · Published 27 January 2026 · Reviewed against our editorial standards · About the author
Read enough generated text and the sameness becomes physical: the balanced clauses, the tidy triads, the reflexive both-sidesing, the closing paragraph that restates everything you just read.
This is not a mystery. It follows from how models are trained and how text is sampled, and each cause has a corresponding lever.
Key takeaways
- Cause one: averaging: A model trained to predict likely continuations gravitates toward the centre of the distribution.
- Cause two: preference tuning: Models are tuned on human preference data, and raters reward balance, hedging, comprehensiveness and politeness.
- Cause three: format habits: Heavy use of headings, bullets and bold text comes from the same tuning.
- The levers that actually work: Supply source material with texture: interview quotes, real numbers, a specific customer's complaint.
Cause one: averaging
A model trained to predict likely continuations gravitates toward the centre of the distribution. The most probable next sentence is, almost by definition, the least surprising one.
Distinctive writing is improbable writing: an unexpected example, an unusual structure, a claim most people would not make.
The lever is specificity of input. Improbable material forces improbable output, because the model must accommodate facts that do not fit the template.
Cause two: preference tuning
Models are tuned on human preference data, and raters reward balance, hedging, comprehensiveness and politeness. Those preferences produce the characteristic tone: agreeable, complete, and slightly evasive.
It also explains the compulsive summarising and the reluctance to take a position.
The lever is explicit permission: instruct it to take one position, omit caveats, and skip the summary. Without instruction it defaults to the tuned behaviour.
Cause three: format habits
Heavy use of headings, bullets and bold text comes from the same tuning. It reads as helpful in a chat window and as filler in prose.
Ask for continuous prose with no lists when you want writing rather than notes. Ask for lists when you want notes.
Structure is a decision about the reader, and it should be yours.
The levers that actually work
Supply source material with texture: interview quotes, real numbers, a specific customer's complaint. Ban your least favourite vocabulary explicitly. Constrain the structure yourself. Provide a style sample of your own writing.
Ask for a first draft that is deliberately opinionated and then moderate it yourself, rather than asking for balance and trying to add conviction later.
Write the opening and closing by hand. Those two positions carry most of the perceived voice.
What does not work
Telling it to 'sound more human' produces a caricature: contractions, exclamation marks and forced casualness. Humanizer tools that substitute synonyms make text worse and no less detectable.
Asking for creativity without constraints produces randomness rather than distinctiveness. Voice comes from consistent choices, not from variance.
Chasing a detector score is the least productive activity of all, since detectors are unreliable and optimising for them degrades readability.
A quick self-test
Take any paragraph and ask whether a competitor could publish it unchanged. If yes, it contains no proprietary knowledge and should be rewritten or cut.
Then check the ratio of specifics to generalities. Good writing in any field runs heavily toward specifics, and that ratio is something you can measure in a draft in about a minute.
The specific tells readers notice
Balanced tricolons in every paragraph. Openings that restate the question. Transitions like moreover and furthermore doing no work. Conclusions that summarise without adding. Hedged claims that avoid committing to anything checkable.
Individually these are mild. Together they produce the flat, evenly-weighted texture that readers now recognise within two paragraphs, and recognition is the problem — it signals that nobody was accountable for the text.
The cure is asymmetry. Real writing has a strong paragraph and a throwaway one, a long sentence followed by four words, one point argued harder than the rest.
Supply what the model cannot invent
Distinctiveness comes from specifics a model has no access to: what happened in your business last quarter, the objection your customers actually raise, the number you measured, the thing you tried that failed.
Give it three of those before asking for a draft, and instruct it to build the piece around them rather than mentioning them in passing. Output quality tracks input specificity more closely than it tracks prompt sophistication.
If you have nothing specific to supply, the honest conclusion is that the article does not need to exist. Generic content competes with an unlimited supply of the same thing.
A short de-genericising checklist
Delete the first paragraph and see whether anything is lost. Replace every abstract noun phrase with a concrete example. Cut one adjective per sentence. Commit to one claim that could be argued with.
Then read it aloud. Sameness is audible before it is visible, and the sentences you stumble over are usually the ones the model padded.
Finally, check the piece says something a competitor's version would not. If it does not, the editing was cosmetic.
Frequently asked questions
Will future models fix this automatically?
Partly, but the averaging pressure is structural. Distinctiveness will keep coming mainly from the material and constraints you supply.
Do humanizer tools work?
They reduce some surface patterns and typically damage clarity. Editing with a checklist produces better text and better detection outcomes.
Is a personal style guide worth writing?
Yes. Twenty lines covering vocabulary, structure and stance improves every future draft and takes an hour to write.
How much should I edit before publishing?
Enough that the piece contains something only you could have written. That is the publishable threshold, not word count.
Do AI detectors solve this?
No. They are unreliable in both directions and they measure statistical texture rather than whether the writing is any good or any use.
Does a style guide help?
Considerably, if it contains examples rather than adjectives. Two paragraphs of your real writing teach a model more than a page of instructions.