How to Use AI for Writing Without Sounding Like AI

Most readers cannot explain what makes a piece of writing feel machine-produced, but a lot of them can feel it. The reaction is rarely about grammar — AI-assisted prose is usually cleaner than the average human draft. It is about texture. The sentences are correct, evenly weighted, and oddly frictionless, and after three paragraphs the reader has learned nothing they could not have guessed.

That is a fixable problem, and fixing it does not mean abandoning the tools. It means understanding what the tells actually are and building an editing pass that removes them. Below is a description of the common patterns, followed by a workflow that has held up well as the underlying models have changed.

Why the tells exist

Language models produce text by repeatedly choosing likely continuations. Averaged over a whole article, that produces prose gravitating toward the middle of everything: middling sentence length, conventional structure, safe claims, familiar transitions. Nothing is wrong; nothing is surprising either. Add the fact that models are tuned to be helpful and inoffensive, and you get writing that hedges, balances, and summarises itself constantly.

So the tells are not random quirks to memorise. They are symptoms of one underlying trait: statistical averageness. Once you see that, the edits become obvious.

The common tells

Structural tells

• The three-part everything. Three examples, three bullets, three adjectives, three sections. Human writers are lumpier — sometimes two, sometimes seven.
• Uniform paragraph length. Four sentences, four sentences, four sentences. Real writing has one-sentence paragraphs where the point deserves the emphasis.
• The restated conclusion. A final section that adds nothing and simply recapitulates what was already said, often opening with a signal that it is about to conclude.
• Symmetry that reality does not have. Every advantage matched with a disadvantage of equal weight, even when the evidence is lopsided. Genuine analysis is often unbalanced because the world is.

Sentence-level tells

• The it-is-not-X-it-is-Y construction, used as a rhythm device rather than to make a real distinction. Once per article is fine. Four times is a fingerprint.
• Stacked pairs. Clear and concise, robust and scalable, thoughtful and deliberate. Two words where one would do, chosen for cadence rather than meaning.
• Heavy connective scaffolding. Moreover, furthermore, additionally, in conclusion — signposting so dense the reader is told where they are going more than they are taken there.
• Reflexive hedging. Every claim wrapped in it is important to note, generally speaking, in many cases. Some hedging is honest. Constant hedging means nothing has been asserted.
• Even rhythm. Very few short sentences, very few long winding ones. The music is flat.

Content tells

These matter most. The strongest signal is not vocabulary but absence of specificity. Machine-averaged text tends to lack proper nouns, numbers with sources, dates, named people, and concrete scenes. It explains that a strategy can improve outcomes without saying whose outcomes, by how much, measured how. It also tends toward what might be called the encyclopaedia voice — the tone of an entity that has read everything and experienced nothing. No preferences, no mistakes, no stakes.

A related tell is confident vagueness about verifiable things: a statistic with no source, a study with no author, a quotation attributed to no one in particular. Readers who check will find nothing there, which costs far more trust than clumsy prose ever would.

An editing workflow that works

The goal is not to disguise AI use. It is to produce writing worth reading, which happens to be the same thing. Six passes, roughly in this order.

Pass 1: Own the structure before you generate

Decide your argument, your sections, and your key examples yourself — even roughly, even badly. If the tool supplies the outline, everything downstream inherits its averageness. Ten minutes of your own thinking here does more for the final piece than any amount of prompt engineering.

Pass 2: Inject specifics you actually have

Go through the draft and replace every generic claim with something concrete from your own knowledge: the real number, the real client situation, the thing that went wrong last spring. This single pass fixes more of the AI feel than all the others combined, because it supplies exactly what the model cannot — first-hand knowledge.

Pass 3: Verify anything checkable

Every statistic, date, name, quotation, citation, and technical claim gets checked against a source you would be comfortable naming. If you cannot verify it, cut it or mark it as uncertain in plain language. Do this before you polish the prose, so you are not fine-tuning sentences you are about to delete.

Pass 4: Break the rhythm

Read the draft aloud, or have it read to you. Then deliberately vary it: cut one paragraph to a single sentence, merge two short ones into a longer argument, delete most of the connective scaffolding. If a moreover can be removed without confusing the reader, remove it. Aim for uneven paragraph lengths and at least a few sentences that are noticeably short or noticeably long.

Pass 5: Cut the throat-clearing

• Delete sentences that announce what the next sentence will do.
• Delete the concluding restatement unless it adds a new implication.
• Delete hedges around claims you are actually confident about.
• Delete adjective pairs down to the more accurate word.
• Delete any sentence that would still be true if the topic were something else entirely.

Most drafts survive losing fifteen to twenty percent of their words. Expect the piece to get better as it gets shorter.

Pass 6: Take a position

Find the place where the draft balances two views and ask yourself which one you believe. Say so, and say why. A stated view with reasoning attached is the clearest possible signal that a mind was involved. If you genuinely do not have a view, say that instead — that is also a position, and an honest one.

A note on detection tools

Software that claims to identify AI-generated text is unreliable in both directions. It flags human writing — particularly from non-native English speakers and from people trained in formal or technical registers — and it misses edited machine text routinely. Do not use these tools as a target to optimise against, and be cautious about any process that treats their output as evidence. Writing well and being honest about your process is a better strategy than chasing a score.

The practical takeaway

The fastest route to writing that does not sound machine-made is to put things in it that only you could have put there: your examples, your numbers, your judgment, your position. Use AI for what it is good at — structure, first drafts, tightening, catching errors — and spend the time you save on specificity and verification.

If you want one rule, use this one: before you publish anything, find at least three sentences that could not have been written by someone who had not done your particular work. If you cannot find them, the draft is not finished yet.