Guide · · 6 min read
Why your AI writing sounds like AI
Readers can tell, and it is not the vocabulary giving you away. It is the structure. Six specific habits and how to break them.
By Manish
People say AI writing sounds generic and then blame the vocabulary. Delve, tapestry, testament, navigate the landscape. Swapping those words out does not fix it, which anyone who has tried a banned-word list has discovered.
The vocabulary was never the problem. What gives AI writing away is shape. Models were trained to produce balanced, complete, evenly weighted prose, and humans writing well do almost none of those things.
The six structural habits
- Every paragraph is the same length. Real writing varies. Sometimes a point deserves one line on its own.
- Nothing is ever just good or bad. Every claim gets a counterweight in the same sentence, so nothing lands and the reader learns you will not commit.
- Rule of three, endlessly. Three benefits, three reasons, three examples, whether or not there are genuinely three.
- Summary sentences that restate the paragraph you just read, adding nothing.
- No specifics. Names, numbers, dates and quotes are what make writing feel witnessed rather than assembled.
- A tidy conclusion that recaps everything. Human writing usually just stops once the point is made.
Read something you know was AI-generated with this list beside you. All six will be present, usually in the first three paragraphs, and once you see the pattern you cannot unsee it.
Why models write this way
It is not a flaw so much as an average. Trained on enormous quantities of text and tuned toward responses people rate highly, models converge on prose that is competent, safe, and offensive to nobody.
Hedging is rewarded because an unhedged claim can be wrong and a hedged one rarely is. Balance is rewarded because it reads as fair. Structure is rewarded because it reads as organised. Each individual tendency is defensible, and together they produce writing with no centre of gravity.
The result reads like it was written by a committee that agreed on everything, which is precisely what it is.
Fixes that work immediately
Cut the last sentence of every paragraph. It is almost always a restatement, and the paragraph gets sharper without it. This single edit does more than any other.
Take out one of the three. If you genuinely have two reasons, say two. Padding to three produces the item nobody believes, and readers can feel which one was invented to fill the slot.
Delete the hedges that are not load-bearing. Some hedges carry real information about uncertainty and should stay. Most are reflexive, and removing them changes nothing except that the sentence now says something.
Vary paragraph length deliberately. Put a one-line paragraph where the point deserves weight.
It works.
The fix that matters most
Give it something only you know. Most flat AI writing is flat because the prompt was flat, and no amount of editing rescues a piece that had nothing specific in it to begin with.
Supply a real anecdote, an actual number, a genuine opinion you are willing to defend. The output changes character immediately, because the model now has material rather than just a topic.
The model is not the bottleneck. It cannot know the thing only you know, and that thing is usually what makes the piece worth reading.
This is also why AI writing about generic topics is worse than AI writing about your specific situation. Ask for a blog post about productivity and you get an average of everything ever written about productivity. Describe how your team actually works and what broke last quarter, and you get something with a point of view.
What to keep the model for
None of this argues against using AI to write. It argues against shipping the first draft, which is a different claim.
Models are genuinely excellent at structuring a mess of notes, at producing a draft you can react to, at cutting length, and at catching the sentence you wrote three different ways. Reacting to a bad draft is far easier than facing a blank page, and that is real value.
They are poor at knowing what matters, at having taste, and at committing to a position. Those are the parts you were always going to have to do.
The detection question
People ask whether AI detectors will catch them. Mostly the wrong question, and worth answering briefly anyway.
Detectors are unreliable in both directions. They flag human writing that is simple and structured, which disproportionately affects non-native English speakers, and they miss AI writing that has been edited even lightly. Building your process around evading them is optimising for the wrong judge.
Readers are the judge that matters, and readers do not run detectors. They simply stop reading when a piece says nothing, and they stop trusting a source that keeps doing it. That is a slower and much more expensive penalty than being flagged.
A test before you publish
Read the piece and ask what it claims that a competitor could not also claim. If the answer is nothing, the piece has no reason to exist, regardless of who or what wrote it.
Then find the one sentence you would be willing to defend in an argument. If there is not one, add one. That sentence is usually the reason someone will remember the piece, and it is always the sentence a model would not have written on its own.
Editing an AI draft is a different job from editing your own
When you edit your own writing you are fixing execution. The thinking already happened, badly or well, and it is on the page. When you edit an AI draft you are frequently supplying the thinking, because the draft is fluent about a topic without having a position on it.
This catches people out. They approach the draft as a copy edit, tidy the sentences, and publish something that reads well and says nothing. The fluency masks the absence, which is exactly what makes AI drafts dangerous to edit casually.
A better sequence: read the draft once and write down, in your own words, what it actually claims. If that list is short or empty, the piece needs a position rather than a polish, and no amount of sentence-level work will supply one.
Structure it yourself first
The single highest-return change to how most people use AI for writing is to write the outline themselves.
Give the model your structure, your section headings, and a line about what each section should establish. The output changes character completely, because you have supplied the part models are weakest at, which is deciding what matters and in what order.
Ask for an outline instead and you will get the same shape every time. Introduction, background, three considerations, challenges, conclusion. It is competent and it is the shape of every article written about anything, which is why nobody remembers any of them.
Voice comes from constraint
People try to get voice by asking for it. Write in a conversational tone, write with personality, write like a human. These produce a performance of voice rather than voice, usually involving more exclamation marks.
Voice comes from constraints that force specific choices. Tell it the reader is a sceptical engineer who has already tried two other approaches. Tell it never to use a metaphor. Tell it the piece must fit in four hundred words. Constraints produce distinctive writing because they close off the safe average.
You do not get voice by asking for personality. You get it by making the generic answer impossible.
When AI writing is genuinely fine
Worth being fair about this, because the criticism above applies to a specific kind of writing and not to all of it.
For functional text where the reader wants information rather than an experience, AI output is frequently as good as a person would produce and considerably faster. Release notes, meeting summaries, documentation, internal updates, first-pass translations. Nobody reads a changelog for its voice.
The problems in this piece apply to writing that has to persuade, to be remembered, or to be attributed to you personally. That is a smaller share of most people's writing than they think, and it is the share worth doing properly.
Try a writing tool that fits your voice
Compare the AI writing tools we have reviewed and find one that sounds like you.