How to Make AI Writing Sound Human (Without a Banned-Words List)

Almost every guide to this problem hands you a list of forbidden words. Delete “delve.” Never write “in today's fast-paced world.” Ban the em dash. That advice removes the fingerprints and leaves the body: the writing still reads as machine-made, because word choice was never the actual tell. This is the diagnosis, the three fixes that work in the order they matter, and the prompts we use to do it.

The real reason it sounds like AI

A language model produces the most probable continuation of your prompt. Ask it for “a blog post about customer retention” and the most probable output is the statistical centre of every article ever written on customer retention — true, tidy, and completely interchangeable with a thousand other pages. Nothing in it could only have come from you.

AI writing does not sound artificial because of the words. It sounds artificial because it contains no information that had to come from a specific person.

That reframing changes what you do about it. If the problem were vocabulary, a find-and-replace would solve it. Because the problem is generic content, the fix is to put things in the prompt that the model could not have guessed: your numbers, your customer's exact words, the thing that went wrong last quarter, the opinion you hold that your competitors would not put in writing.

First, learn the tells (so you can hear them)

You cannot edit what you cannot hear. These are the patterns that make a reader think “this was generated” before they can say why. Read your draft hunting for these specifically:

• The rule of three, every time.“Clear, concise, and compelling.” “Strategy, execution, and results.” Humans use tricolons occasionally for rhythm. Models use them as a default sentence shape. Three or four in one piece is a signature.
• “It's not just X — it's Y.”Along with “more than a Z, it's a W.” This construction has become the single most recognisable machine sentence in English. Cut every instance.
• Throat-clearing openers.“In today's rapidly evolving landscape.” “In an era where.” “Businesses of all sizes are increasingly.” These say nothing and exist only to reach the real first sentence.
• The summary sandwich. Tell them what you will say, say it, then tell them what you said. Models do this at every level — per article, per section, sometimes per paragraph. It doubles length and halves signal.
• Uniform rhythm. Every sentence between fifteen and twenty-five words, every paragraph three to four sentences. Human prose lurches. Short sentence. Then a long one that runs a clause past where a model would have stopped, because the thought was not finished.
• Vocabulary nobody says aloud. Delve, leverage, robust, seamless, tapestry, landscape, testament, underscore, navigate the complexities, unlock the potential. Not forbidden words — just words that appear in machine drafts at rates no human matches.
• Symmetrical hedging.Every claim balanced by its counter-claim, so the piece never actually says anything. “While X has benefits, it also presents challenges.” A person with real experience commits.

Why banned-word lists fail

Give a model a list of forbidden words and it will obey — by finding the second-most-probable phrasing, which is the same generic sentence wearing a different coat. “Delve into” becomes “explore in depth.” “Seamless” becomes “effortless.” You have moved down the probability ranking, not away from the average.

Worse, long prohibition lists eat the model's attention. Every token spent on “never say X” is a token not spent on what you actually want. Keep the ban list to the handful of constructions above, and spend the rest of your prompt on the three fixes below.

Fix 1: Give it your voice as data, not as an adjective

“Write in a friendly, professional tone” is not an instruction — it is a mood. Every model already thinks it writes in a friendly, professional tone. The instruction that works is a sample plus a derived rule set.

Do this once, keep the output forever. Find three things you wrote that sound like you — a long email to a customer, a post you were happy with, a page off your own site. Then run this:

Below are three samples of my writing. Do not compliment them and do not rewrite them.

Analyse them and produce a VOICE SPEC I can paste into future prompts, covering:
1. Average sentence length and how much it varies
2. Paragraph length
3. Contractions: which ones I use, which I avoid
4. Punctuation habits (dashes, colons, semicolons, parentheses, sentence fragments)
5. Words and phrases I use repeatedly
6. Words and constructions that appear ZERO times in my samples
7. How I open a piece and how I close one
8. Where I am direct vs where I hedge
9. Whether I use humour, and what kind
10. Person and address: I / we / you, and how often

Write the spec as blunt imperative rules ("Use X." "Never Y.") — not description.
Quote my actual words as evidence for each rule.

SAMPLE 1: [paste]
SAMPLE 2: [paste]
SAMPLE 3: [paste]

The output is typically a page of rules like “open with a concrete situation, never a definition” and “never uses the word utilise.” Save it as a text file. Paste it at the top of every writing prompt from then on — or, if your tool supports persistent instructions, put it there so it applies automatically. This one artefact does more for human-sounding output than any prompt engineering trick.

Fix 2: Inject the facts only you have

This is the highest-leverage step and the one almost nobody does, because it takes five minutes of thinking instead of five seconds of typing. Before you ask for a draft, write down the raw material the model cannot invent:

1. 1Numbers.Prices, timelines, quantities, durations. “A quick turnaround” is machine filler. “Back to you inside a business day, or we tell you why not” is a claim.
2. 2Verbatim customer language. Pull the actual sentences people use in emails, reviews, and calls. Their phrasing is unpredictable in exactly the way generated text is not, and it doubles as search-query research.
3. 3A specific incident. One thing that happened — what broke, what it cost, what you changed. Nothing marks writing as human faster than a story with consequences attached.
4. 4An opinion with a downside. Something you believe that costs you to say: who your product is wrong for, the popular tactic you think is a waste, the thing you got wrong. Models default to safe symmetry; a real position cannot be generated.
5. 5The constraint. What this piece must not do. Not sell. Not mention the upsell. Not run over 600 words. Constraints produce shape, and shape reads as intent.

Feed those five as the body of the prompt and ask the model to write fromthem rather than about the topic. The difference is not subtle — it is the difference between a page that could sit on any competitor's site and one that could not.

Fix 3: Constrain the shape, then run a rewrite pass

Even with voice and facts supplied, first drafts drift toward the average. Do not argue with the draft in chat — run a dedicated second pass whose only job is de-genericising. This is the exact pass we run on our own copy:

Rewrite the draft below. Keep every fact, number and claim exactly as written —
you may not add information that is not already present.

Apply these edits:
1. DELETE any sentence that would still be true for a competitor. If removing a
   sentence loses no information, it goes.
2. DELETE all preamble. The piece starts at the first sentence carrying content.
3. NO tricolons ("clear, concise and compelling"). Max one in the whole piece.
4. NO "it's not just X, it's Y" or "more than a X, it's a Y". Zero instances.
5. NO section that restates the section above it. Say things once.
6. VARY sentence length hard: some under 8 words, some over 30. No two adjacent
   paragraphs with the same number of sentences.
7. REPLACE every abstraction with the concrete thing it stands for. "Improved
   efficiency" -> the actual task that got faster.
8. BAN: delve, leverage, robust, seamless, tapestry, landscape, testament,
   underscore, navigate the complexities, unlock, elevate, empower.
9. END on the last real sentence. No summary paragraph, no call to reflect.

Then list, separately, the three weakest sentences that survived and why.

DRAFT:
[paste]

The last instruction matters more than it looks. Asking for a self-critique after the rewrite surfaces the lines you should cut by hand, and it stops the model from declaring victory. If you want the longer version of why prompts written as procedures outperform prompts written as wishes, that is the whole argument in our prompt-structure guide.

The five-minute human pass no prompt replaces

Everything above gets you to a draft that reads as written by a person. This gets it the rest of the way:

• Read it out loud. The only reliable detector. Anywhere your voice flattens or you skim ahead, the reader will too. This catches things no checklist does.
• Delete the first paragraph. Then check whether anything was lost. Roughly half the time the piece is better and starts faster.
• Add one thing only you know. A detail, a number, a name, a small admission. One per piece is enough to change how the whole thing reads.
• Cut the closing summary. Models end by recapping. People end on the last thing worth saying and stop.
• Fix one thing to be slightly wrong for a machine. A fragment. A one-sentence paragraph. A joke that only lands if you know the industry. Perfect evenness is the tell.

An honest note on AI detectors

People arrive at this problem because a detector flagged their draft, or because they fear one will. Two things are worth knowing.

First, detectors are unreliable in both directions. They score how predictable text is, so careful, plain, well-structured human writing gets flagged routinely — a known and repeatedly documented failure that has affected non-native English writers in particular. A tool that misfires on clear human prose is not a standard worth writing toward.

Second, and more practically: nobody buys because your page passed a detector. They buy because the page told them something specific they did not already know. Optimise for the reader who might pay you and the detector question mostly stops mattering — the same fixes that make writing pass as human are the ones that make it worth reading.

The whole thing, as a checklist

• Extract a voice spec from three real samples. Save it. Reuse it every time.
• Before drafting, write down numbers, verbatim customer language, one incident, one opinion with a downside, and the constraint.
• Ask the model to write from those, not about the topic.
• Run the de-genericiser as a separate pass, then read its self-critique.
• Read aloud, delete the opener, add one thing only you know, cut the closing summary.
• Never publish a number, quote or source you have not personally checked.

Do that and the “does this sound like AI” question resolves itself, because the page now contains things no model could have produced without you. If you want the operating layer underneath — how the documentation, prompts and review steps fit together as a system rather than a habit — that is the company-on-AI-agents write-up, and which model to run it on matters far less than most people think.