How to Write With AI Without Sounding Like It

Someone reads your draft and tells you it sounds like AI. So you do the obvious thing, the first thing every guide on how to write with AI tells you. You cut “delve”. You delete “moreover” and “furthermore”. You strip out “landscape” and “tapestry” and “in today’s fast-paced world”. Then you read it back.

It still sounds like AI.

Vocabulary is only the symptom, and most advice on how to write with AI stops there. What readers are reacting to is the shape of the piece: how the argument is weighted, how much it is willing to commit, where it chooses to spend its attention. A model writes by averaging, and averaged prose comes out smooth, symmetrical and careful. Swapping the words on top of that skeleton changes very little, because the skeleton is what gave it away.

Writing that works is lumpy. It gives four paragraphs to the one thing the writer genuinely cares about and a single line to everything else, because real interest is disproportionate. It says at least one thing that could turn out to be wrong. A good deal of what gets called AI-sounding is ordinary unclear writing arriving at unusual speed and fluency, and the repairs are the same mechanical fixes that work when writing is not your job.

The Tell Is the Shape

Ask a model for a thousand words on almost anything and watch what comes back. Four sections of roughly equal length. Three examples in each. Every claim followed by a polite qualification. A closing paragraph that restates the four sections in the order you just read them.

Nothing in that is wrong, exactly. It is the average of everything ever written on the subject, and an average has no preferences. The reason it feels hollow is that a person writing about something they care about cannot keep it up. They overrun badly on the bit that interests them. They dismiss an entire school of thought in half a sentence because they find it boring. They pick a side and accept the cost of picking it.

Editors and heavy readers detect this in about fifteen seconds and usually cannot say why beyond “it reads like AI”. What they are responding to is evenness. Once you know that, the checks below stop being mysterious, and most of them you can run on a draft in five minutes.

Eight Tells You Can Check For Yourself

• Negative parallelism. The “it is not X, it is Y” construction, and its cousin “not just A, but B”. One of these is a legitimate rhetorical device. Three in the same article is a signature, and trained readers spot it instantly. This is the most recognisable tell there is. Search your draft for “not just” and “isn’t about” and count the hits.
• Even weighting. Every section the same length, every point given the same emphasis, as though the writer had no favourites. Look at your section lengths. If they sit within twenty percent of each other, you have written a report rather than an argument.
• The hedge reflex. Every claim trailed by its own qualification, so the piece never quite says anything you could disagree with. “Of course, results will vary.” “That said, context matters.” Delete most of them. Keep the qualifications that carry real information.
• Inflated significance. “Plays a vital role.” “Stands as a testament to.” “Underscores the importance of.” These phrases add weight while adding nothing, which is exactly why an averaging machine reaches for them.
• Stacked connectives and asides. “However”, “moreover”, “additionally”, “it is worth noting”. One “however” in a whole article is plenty. “It is worth noting” can always be cut, and the sentence is better afterwards.
• Tricolon as the default rhythm. Three-item lists everywhere, for everything, whether or not there happen to be three things. Some ideas have two parts. Some have one. Some have nine.
• Summary paragraphs that do not advance. A closing section that repeats the argument you just read, in the same order, slightly compressed. If the final paragraph could be deleted with no loss, delete it and write a better one.
• The absence of specifics. No names, no numbers, no dates, no places. A model drifts toward the general because the general is safer, and the general is precisely what makes prose forgettable. “A major retailer” is a fingerprint. “Marks and Spencer, in 2022” is writing.

Run the first and the last of those over any draft you are unsure about. Between them they catch more than the other six combined.

What AI Detectors Are Actually Worth

A fair question at this point is whether you can just run the draft through a detector and see what it says. The honest answer is that detectors are unreliable in both directions, and the errors land unevenly.

In 2023 a Stanford group led by Weixin Liang ran seven GPT detectors over 91 TOEFL essays written by non-native English speakers. The average false positive rate was 61.22 percent. Eighty-nine of the 91 human-written essays were flagged as AI-generated by at least one detector, and 18 of them were flagged by all seven. The paper, published in Patterns, argued the tools should be treated with caution in any evaluative setting.

A detector will tell you almost nothing useful about your own draft, and what it does tell you is as likely to be wrong as right. Your ear is the better instrument. The rest of this is about training it.

Where the Model Is Genuinely Good

None of this is an argument against using the tools. I use them every working day, and the places where they earn their keep are specific enough to name.

They are good at structure. Hand over your raw argument in a mess and ask for three ways to order it. One of the three is usually better than the order you had.

They are good at the dull parts. Background, definitions, the paragraph explaining a term some readers will not know. These have to be accurate and clear and nobody will ever admire them.

They are very good at arguing against you. “Here is my position. Give me the strongest case that I am wrong, and tell me which paragraph is weakest.” That is the single highest-value request I make of a model, and the answer is frequently uncomfortable.

They are good at openings. Twenty alternative first lines takes ten seconds, and you will then write the twenty-first yourself, better, having seen what the obvious twenty look like.

What they are bad at is the sentence that carries your judgement. The line where you say what you actually think, in the words you would use, is the one piece of the article the reader came for. Write that one by hand, every time.

How to Write With AI: Bring the Raw Material

The parts a model cannot produce are the parts only you have. What you watched happen in a meeting. The figure you looked up yourself. The decision you got wrong in 2019 and what it cost you. The name of the town.

Most people use these tools back to front. They ask for content and then try to make the content sound like them. The better move is to turn up with the raw material and use the model to arrange it. Paste in your rough notes, the real numbers, the half-remembered example, the thing your co-founder said in the car, and ask for an order.

Specificity is also the fastest cure for the eighth tell on that list. A paragraph with a date and a place in it cannot have been averaged out of anything, because averages do not have dates.

Write It Badly First

The order of operations is not a detail here, it is the whole technique.

Write your argument yourself, in bad prose. Ugly sentences, wrong words, a paragraph that is really just a note to yourself in square brackets. Get the thinking down in whatever state it arrives. Then improve it, and use the model to help you improve it.

Reverse that order, asking for a draft and then editing what comes back, and you end up with a piece organised around the model’s instincts. You can polish that for a week and it will keep its original shape, because editing cannot repair an argument you never made.

It helps enormously to know what the finished thing is meant to do before you start, which is the same reason so many projects get started and then quietly abandoned. With no definition of done, there is nothing to write toward and no point at which you are allowed to stop.

Cut a Third, Then Read It Aloud

Model prose is padded by construction. It restates, it qualifies, it announces what it is about to say, and it eases you into paragraphs that would land harder starting at their second sentence. Almost every draft I have worked on reads better at roughly sixty-five percent of its length.

Do the cut as a separate pass with a number attached. Take the word count, multiply by 0.65, and hit it. Arbitrary targets work here because the padding is spread evenly through the whole thing, so wherever you cut, you will find some.

Then read it aloud. Out loud, at speaking pace, the whole way through. It is the fastest test in existence and it costs four minutes. Every place you stumble is a place the rhythm is wrong. Every sentence you catch yourself rushing is too long. Every paragraph that bores you to say will bore a stranger to read. Most usefully, the evenness that no amount of word-swapping could fix becomes audible straight away, because your voice has to do something with the emphasis and an averaged sentence gives it nothing to do.

Say One Thing That Could Be Wrong

Here is the deepest reason unedited model output feels empty, and it survives every fix above if you skip it.

The prose risks nothing. Ask for a piece on hiring and you get a balanced survey of hiring considerations. Nobody could argue with any of it, which is why nobody needed to read it, because the reason anyone reads a person rather than a summary is to find out what that person concluded.

So before you publish, go and find the sentence a reasonable reader could disagree with. If there is no such sentence, you have written a summary. Go back and put your actual position in, the one you would defend out loud in a room with people who know the subject.

This is also the mechanism by which writing does anything for you professionally. A reader who disagrees with a clear position still walks away knowing you have judgement, and demonstrated judgement in public is the closest available substitute for a track record when you do not yet have one, which is why writing is such an effective way to find your first clients without a network. Careful, balanced, unobjectionable prose teaches a reader nothing about you whatsoever.

The model will never supply that part, and no guide on how to write with AI will change it. It has no stake in the answer. You do, and the difference is audible.