7 min read

How to Make AI-Written Content Stop Sounding Like AI

The specific tells, why asking for your voice does nothing, and the three-step method — voice profile, blind test, banned list — that removes them.

Published 25 August 2026

The complaint is always the same and it is always correct: it sounds like AI. What almost nobody can do is say which part sounds like AI, which is why the usual fix — asking the model to “make it sound more human” — produces something that sounds like AI trying to sound human, which is worse.

The tells are specific and they are learnable. So is the method for removing them, and it is not a prompt trick.

The tells, specifically

Read anything a model wrote unprompted and you will find most of these. They are worth naming because a named tell can be banned and a vague feeling cannot.

The tricolon, everywhere. Three-item lists used as a rhythm rather than because there are three things. Faster, cheaper and more reliable.
The “it’s not X, it’s Y” pivot. Deployed at least once per piece, usually as the closing line of a section.
Symmetrical paragraphs. Every paragraph three to four sentences, every sentence a similar length. Real writing is lumpy.
Hedged claims that commit to nothing. Can help you, may improve, is often considered.
The restated topic sentence. The paragraph says the thing, then explains the thing, then says the thing again in different words.
Abstract nouns doing a verb’s work. The implementation of a strategy rather than doing it.
The summarising close. A final paragraph that adds nothing and exists because the model learned that pieces end with one.
Words nobody says out loud. Delve, leverage, robust, seamless, landscape, tapestry, navigate the complexities of.

Two things about this list. It is generic — every one applies to any subject. And it is only half the problem, because removing all of it gets you writing that is not identifiably AI and is also not identifiably you. Neutral is a different failure from robotic, and it is harder to notice.

Why “write in my voice” does nothing

Because the model has no idea what your voice is, and neither do you in any form it can use.

Voice feels obvious from the inside and is almost impossible to state from memory. Asked to describe it, people produce adjectives — conversational, punchy, warm — and adjectives are descriptions rather than instructions. Every model has a generic interpretation of conversational and it will give you that, which is precisely the thing you were trying to avoid. A description is not a position, and it cannot be followed.

What works is extraction rather than description: taking the voice out of writing you have already done, in the form of rules concrete enough to check against.

Extracting a voice profile from what you have already written

Gather ten to fifteen pieces you wrote yourself and were happy with. Emails count. Then ask for the analysis in a form you can argue with:

Voice profile extraction

Below are 12 pieces I wrote myself. Analyse them and produce a voice profile
with these sections, using evidence from the text rather than adjectives:

SENTENCE LENGTH: average, range, and where I deliberately go short.
OPENINGS: how I typically start a piece. Quote three.
STRUCTURE: how I sequence an argument. Do I lead with the claim or build to it?
VOCABULARY: words I use often that most writers in this field do not. Words I
never use.
PUNCTUATION: my actual habits — dashes, semicolons, parentheses, one-line
paragraphs.
STANCE: how strongly I commit to a claim. Do I hedge, and where?
HUMOUR: present or not, and what kind.
RHYTHM: what I do that a competent but generic writer would not.

For each section, quote at least one line from the source as evidence. If a
section has no clear pattern, say so instead of inventing one.

PIECES: [paste]

The last instruction is the one that makes the output usable. Without it you get a confident profile of a writer who does not exist, because the model will manufacture a pattern out of twelve samples that share nothing.

Read the profile and correct it. It will get two or three things wrong, and the corrections are the most valuable part — they are the rules you knew but had never articulated. Then save the corrected version into a Project, a Gem, or Custom Instructions, so it applies by default rather than being pasted into every request. Pasting it each time is the dilution problem waiting to happen.

Free · no card

Twenty prompts that survive a long conversation

The free guide is fifteen pages, and every prompt in it is printed in full.

A voice profile is a hypothesis. This is how you find out whether it works.

A-003 · Blind voice testAI for Content Creators, Ch. 4

Here are four captions. Three are mine, one you wrote using my voice
profile. Without me telling you which, identify the tells that would
give away a non-human or non-me author, and turn each tell into a
new voice rule.

Three real pieces, one generated, presented without labels. You are not asking which is which — you are asking what gives the odd one away, and the answer comes back as rules rather than as a verdict.

Run this every time you materially change the profile. It takes five minutes and it is the only part of this process that produces evidence rather than opinion.

The banned list

The blind test finds tells specific to you. This catches the generic ones, plus the ones peculiar to your field.

A-007 · Banned-words listAI for Content Creators, Ch. 4

Based on my niche and voice profile, build a banned list: the
clichés, filler phrases and AI tells I never want to appear in my
content. Include niche-specific ones, not just generic AI
vocabulary. Aim for 30 entries, grouped by why each is banned.

“Not just generic AI vocabulary” is the operative clause. Every field has its own set of phrases that mark writing as unconsidered — game-changer in tech, bespoke in agencies, journey in almost everything — and those do more damage than delve because your readers are the people who notice them.

Put the banned list in the same standing layer as the voice profile. Grouped by reason rather than as a flat list, so you can lift a ban deliberately when a word is genuinely the right one.

Before and after

Same brief, same model. First with nothing but the topic:

In today’s fast-paced digital landscape, small businesses face unprecedented challenges when it comes to managing their online presence. Leveraging AI tools can help streamline workflows, reduce manual effort, and free up valuable time. It’s not just about working faster — it’s about working smarter. By implementing a robust content strategy, businesses can navigate these complexities and unlock their full potential.

Then with a voice profile, a banned list, and a brief that names the audience and the claim:

Most small businesses do not have a content problem. They have a Tuesday problem: the person who writes the posts is also the person doing the invoicing, and invoicing wins. AI does not fix that by writing faster. It fixes it by making the writing a forty-minute job on Tuesday morning instead of a two-hour job you keep moving to Wednesday.

The second one is not better because it is livelier. It is better because it makes a claim someone could disagree with, and the first one does not say anything at all. That is the real tell underneath all the surface ones — generic writing avoids commitment, and avoiding commitment is what makes it sound generated.

The editing pass that does the rest

Even with a profile and a banned list, a draft arrives at about eighty per cent. The last twenty is an editing pass, and it is the same three moves every time.

Cut the first paragraph. Models warm up. The real opening is almost always the second or third paragraph, and the first one is throat-clearing that restates the title. Delete it and read what is left — if it still makes sense, it was scaffolding.

Break the rhythm deliberately. Find two consecutive paragraphs of similar length and shorten one hard. Three words is fine. The unevenness is what reads as a person thinking rather than a template being filled, and no amount of voice profiling produces it reliably.

Add one thing only you know. A number from your own account, a client who said something specific, the version of this that went wrong in 2024. This is the single highest-value edit available and it is the one thing a model structurally cannot do, because the material is not in its training data or in your prompt. A piece with one genuinely specific detail reads as first-hand even when the surrounding prose is unremarkable; a piece with none reads as synthesised even when every sentence is good.

If you only do one of the three, do the last one.

What this does not fix

Three honest limits.

It will not give you judgement. The model can match your voice and still be wrong about what is worth saying, and choosing between four competent drafts is a decision you have to make.

It will not survive being ignored. A voice profile in a Project you stopped opening is a voice profile you are not using, and drift back to default is quiet.

And it will not manufacture a position you do not have. If your angle is a description rather than a claim, no amount of voice work will fix it — the writing will sound like you and still say nothing. That is an editorial problem, and it is upstream of everything in this post.

The three prompts marked A-003 and A-007 are printed exactly as they appear in AI for Content Creators, which treats this as the central problem rather than a caveat — voice extraction is chapter four, and the 100-prompt library is organised by the stage of the work. If the underlying skill is what you are after rather than the content workflow, the Prompt Engineering Masterclass covers why a constraint changes an output at all.

The same discipline applies to images, incidentally, where the failure mode is not tone but inconsistency — getting the same result twice from an image model is that problem.

Related guides