"Write in a warm, conversational tone" produces something that reads like a model's idea of warm and conversational — which is its own instantly recognisable register, and not yours.
That's the core problem with fixing AI voice. Style descriptions don't transfer style. They select from the model's existing repertoire of registers, all of which readers have now seen a great deal of.
What transfers voice is examples. But before that's useful, it's worth knowing exactly what you're correcting.
The short version
The tells are structural, not vocabulary:
- Uniform section lengths
- Relentless three-item lists
- Every paragraph the same size
- Hedging on everything
- Restating the introduction as a conclusion
- Bolded lead-ins on every bullet point
- No opinions
What actually fixes it: examples of your own writing, plus writing the openings and transitions yourself.
What doesn't: style adjectives, "humanizer" tools, or hunting for banned words.
The tells, properly
Most "AI writing tells" lists are vocabulary lists — delve, tapestry, landscape, testament. Those are real but shallow, and they're the easiest thing to fix. The structural tells are what actually give it away, and they survive any amount of word-swapping.
Uniform section lengths. Every H2 section runs 200 words. Real writing has a 500-word section next to a 60-word one, because some points need room and others don't.
Everything comes in threes. Three benefits, three steps, three reasons. Reality rarely arrives in threes — sometimes there are two things worth saying, sometimes seven.
Paragraph uniformity. Every paragraph three or four lines. Human writing has one-line paragraphs for emphasis and occasional dense ones.
Symmetrical structure. An introduction that previews, sections that deliver, a conclusion that restates. Nothing surprising happens.
Reflexive hedging. "Can often," "may sometimes," "in many cases." Every claim softened, so nothing is actually asserted.
Bolded lead-ins on every single bullet. Useful occasionally, mechanical when every item has one.
No position. The clearest tell of all. It surveys the possibilities and declines to conclude, because that's what consensus-trained output does.
Why style prompts fail
Style adjectives are labels, and the model maps them to registers it already has.
"Conversational" retrieves the model's conversational register. So does everyone else's "conversational." The result is that a million pieces of content all reach for the same handful of voices, which is precisely why the register has become so recognisable.
And adjectives are unfalsifiable. "Warm" and "direct" and "authoritative" don't specify sentence length, paragraph rhythm, whether you use contractions, whether you open with a question, or whether you'd ever write a one-word sentence.
Your voice is those specifics. It isn't a set of adjectives.
What actually works: examples
Paste in two or three pieces of your own writing and instruct the model to match their rhythm, sentence-length variation, and directness.
Why this transfers what descriptions can't: the model can observe actual patterns rather than interpret a label. Your average sentence length. How often you use fragments. Whether you open sections with a statement or a question. Where you put emphasis.
Choose the examples deliberately:
Pick pieces you're genuinely happy with, not just recent ones. Pick things in the same format — reference articles for a reference article, not a LinkedIn post. And pick two or three, since one gets copied too literally while five dilutes into an average.
Build the set once and reuse it. Three of your best pieces, saved somewhere accessible, pasted into every brief. Costs nothing after the first time.
Be explicit about what to match: "Match the rhythm, sentence length variation, paragraph length, and directness of these pieces. Note that sections vary considerably in length and lists are not always three items."
Write these parts yourself
Some things resist transfer, and they're a small proportion of the words.
Openings. The first paragraph is where voice is most visible and where AI is most generic. Write it yourself — it's fifty words and it sets the register for everything after.
Transitions between major sections. The connective tissue where a writer's thinking shows.
Anything expressing an opinion. If you're taking a position, take it in your own words. A model's version of your opinion is a summary of your opinion, which reads differently.
The last line. Endings are hard and AI endings are formulaic.
That's perhaps 15% of the words and it does most of the work on how the piece reads.
Structural editing
Beyond voice, three edits that fix the tells directly.
Vary the section lengths deliberately. Find your longest and shortest sections. If they're within 50 words of each other, cut something and expand something.
Break the threes. Go through every list. Where there are three items, ask whether there are genuinely three — often the third is padding, and occasionally there's a fourth you dropped for symmetry.
Remove the hedging. Search for "can be," "may," "often," "typically," "generally." Not all of them — some hedges are accurate. But where you actually mean the thing, say the thing.
And check the conclusion. If it restates what came before, delete it. Ending on the last real point is stronger than summarising.
Do AI detectors work?
Not reliably, and it's worth understanding why before anyone builds a process around them.
The false positive problem is real. Clear, well-structured human writing is frequently flagged as AI-generated, because clarity and structure are exactly what these tools measure. Non-native English speakers are disproportionately affected, since more conventional phrasing scores as more machine-like.
There's no independent verification of accuracy claims. Detector vendors publish their own figures. No neutral benchmark exists.
And there's an obvious conflict of interest in this whole category: detector companies profit from the belief that AI content is detectable and risky, and "humanizer" companies profit from selling the antidote. Treat claims from either side accordingly.
What this means practically: don't optimise for passing a detector. It's an unreliable target, and chasing it distorts your writing — people add errors and awkwardness deliberately to score better, which makes the piece worse for the actual reader.
Optimise for a reader instead. A human who reads a lot will spot the structural tells long before any tool does, and fixing those makes the piece genuinely better rather than merely differently-scored.
Can AI learn my writing style?
Partially, and it's worth being precise about what "partially" means.
What transfers well: sentence rhythm, paragraph length, formality level, whether you use contractions and fragments, structural habits.
What transfers poorly: the judgement about what's worth saying, what to leave out, when to be blunt, and which example illuminates a point. That's not style — it's thinking, and it looks like style from outside.
What doesn't transfer at all: your actual opinions, your experience, and your knowledge of the specific situation.
The practical implication: the more examples you provide, the closer the surface gets. But the closer it gets on surface, the more the remaining gap is the substantive part — which is why "make it sound like me" has a ceiling, and why the pieces worth publishing are the ones where you supplied something only you had.
Frequently asked questions
What are the signs of AI writing?
The reliable tells are structural rather than lexical. Uniform section and paragraph lengths, lists that always contain three items, symmetrical structure with a conclusion restating the introduction, reflexive hedging on every claim, bolded lead-ins on every bullet, and no clear position. Vocabulary tells like "delve" and "tapestry" are real but shallow — they're the easiest thing to fix and the least diagnostic, since the structural patterns survive any amount of word substitution.
Do AI detectors work?
Not reliably. False positives are a genuine problem — clear, well-structured human writing is frequently flagged, and non-native English speakers are disproportionately affected because conventional phrasing scores as more machine-like. Accuracy claims come from detector vendors with no independent benchmark, and the category has an obvious conflict of interest on both sides. Don't optimise for passing one; a well-read human spots the structural tells long before any tool does.
Can AI learn my writing style?
Partially. Sentence rhythm, paragraph length, formality, and structural habits transfer reasonably well when you provide examples of your own work rather than style adjectives. What doesn't transfer is the judgement about what's worth saying and what to leave out — which reads like style from outside but is actually thinking. That gap is why the pieces worth publishing are the ones where you supplied something the model couldn't have.
What to do next
Take a piece you're proud of and measure two things: your longest section and your shortest.
Then do the same to your last AI-assisted piece.
If the first has a wide spread and the second doesn't, you've found the most visible tell in your own work — and varying section length deliberately in the brief fixes it before you're editing rather than after.
Free: The content brief template.
Related guides
- AI content production without the slop — the full workflow
- The content brief — where voice examples belong
- Email copywriting — voice in a different medium
- Does AI content rank? — whether any of this affects search
Join the Newsletter
Get practical marketing tactics delivered straight to your inbox.
Written by
Muhammad Basim
Related Articles
What Marketing Automation Actually Costs at Scale
Nobody's automation bill jumps because they added more automations. It jumps because they added steps to workflows that were already running — a filter here, an enrichment lookup there, a second notification — and the billing model charges for every one of them, on every run. That's the mechanic behind almost every "why did this […]
One Workflow, Three Tools: A Build Comparison
Feature tables don't tell you what a tool is like to use. The only thing that does is building the same thing twice and noticing where you got annoyed. So here's one realistic workflow — lead routing with enrichment and conditional assignment — specified once and mapped across Zapier, Make, and n8n. What changes between […]
GoHighLevel: 7 Automation Workflows to Build First
GoHighLevel gives you a hundred things you could automate, which is exactly why most accounts end up with forty half-built workflows and no measurable result. Seven are worth building first. They're the ones tied directly to revenue — catching leads before they cool, recovering appointments that would otherwise vanish, and asking for reviews at the […]