A Guide to Make AI Sound More Human When Writing

Why AI sounds robotic in the first place

Here's the mechanism, plainly: a large language model predicts the most probable next word, over and over, based on a vast corpus of human writing. Most probable means average. And average text, by definition, sounds like nobody in particular.

That's the whole problem in one sentence. Everything else in this guide is really just working out the implications.

The tyranny of the average

Because a model is optimised to sit close to the mean of everything it's read, its output tends to flatten. Sentence length is the easiest place to hear this. Human writers vary their rhythm without thinking about it – a long, winding explanation followed by three words that land hard. AI defaults instead to a metronome. Every sentence roughly the same length. Every paragraph roughly the same size. Nothing that surprises the ear.

You feel this before you can name it. That's the give-away.

Vocabulary that gives it away

A handful of words show up constantly in AI drafts, not because they're wrong, but because they're safe, high-probability choices that avoid the specific, idiosyncratic word a person with actual experience would reach for.

Word	Why it's a give-away
Delve	The single most common signal of AI-generated analysis
Tapestry	Used to describe any complex situation, formulaically
Testament	Signals importance without giving actual evidence
Fostering	The default verb for "building" anything relational
Bespoke	Deployed to sound high-end without being specific
Landscape	A generic stand-in for any industry or field
Pivotal	An overused adjective doing the work a real example should do
Telemetry	Used to sound strategic, usually without cause

None of these words are banned in principle. They're banned in practice because they cluster – and when they cluster, they replace the specific detail a real person would have used instead.

The neutrality trap

Good thinking involves picking a side. AI, left to its own devices, doesn't want to. It reaches for "it's important to remember", "many experts agree", "on the other hand" – hedges that let it avoid committing to anything that might be wrong, or might exclude another view. The result reads as balanced but it also feels like no one is actually there. Like there is no human behind it.

Structural leftovers

Then there's the scaffolding a draft leaves behind: "In this article, we will explore", a tidy "In conclusion" at the end, and connective tissue built entirely from "Moreover", "Furthermore", "Additionally". These are signposts a model uses because it's not confident the logic alone will carry the reader from one idea to the next. A person writing from genuine conviction doesn't need them.

Building the human blueprint

Fixing this starts before the model generates a single word. Left alone, it defaults to an average of averages. Give it something specific to work from instead, and it stops guessing.

The style sample method

Collect ten to twenty pieces of your own best writing. For genuinely deep pattern-matching, aim for somewhere near 70 000 words of source material; below that, the model is still mostly guessing.

Have the AI read this material and extract concrete rules: sentence length, punctuation habits, which words you reach for and which you never use. Then rank those rules – put the non-negotiable ones first in each section, because a style guide, like most writing, is hierarchical. Not every rule carries equal weight, and the model needs to know which ones do.

Give it your actual voice, not a brief

If you want the writing to sound like someone who's lived through the thing they're describing, feed the model a transcript of you talking about it – a recorded meeting, a voice note, a rough personal account – rather than a set of instructions about tone. A transcript carries your vocabulary, your rhythm, your priorities in a way a brief never will. Ask the model to write in first person throughout: I, we, our. It's a small instruction with an outsized effect on how lived-in the final piece feels.

Break the model's habits on purpose

Left alone, AI writing settles into predictable shapes. You can force it out of them:

• Vary sentence length. Follow a long, explanatory sentence with something three words long. Let the short one do the work of a full stop.

• Use fragments deliberately. Real speech is full of them. A sentence that isn't a sentence, used at the right moment, reads as human.

• Avoid triplets. AI has a strong pull toward listing exactly three things. Ask for two, or four – anything but three – and the list stops sounding templated.

• Cut the fluff outright. Borrowing from technical-writing standards like ASD-STE100 – built for aviation manuals, of all things – is a decent way to force every sentence to justify its own existence.

  
  

Prompting techniques that actually help

Once the groundwork is in place, the prompting itself can do more of the work.

Tell it what to avoid, explicitly

A negative style guide – a plain list of banned words and phrases – does more for a draft than almost any positive instruction. Once you have a draft, try what we might call a cliché audit: ask the model to find its own clichés and replace each one with a specific, original piece of reasoning instead.

Category	What to cut
Puffery	"In today's digital landscape," "in this fast-paced world"
Filler	"It is worth noting that," "it goes without saying"
Cliché	"At its core," "tapestry," "seamless," "innovative"
Empty framing	"In summary," "overall," "on the other hand"

Titles built around a colon – "The X Factor: How Y Changes Everything" – are worth banning outright too. It's one of the more recognisable AI packaging habits.

Let it interview you

Instead of writing the brief yourself, have the model interview you – one question at a time, pulling out the specific opinions, anecdotes, and arguments you'd otherwise leave out because they didn't occur to you unprompted. A related move: feed the model a chunk of your own writing and ask it to reverse-engineer the rules behind it. That gives you a style guide built from what you actually do, not what you think you do.

Get it to critique itself

Ask the model to list five ways a draft still feels generic for your specific reader. It's often better at spotting its own tells than avoiding them in the first place. Switching models partway through can help too – a model like Claude tends to hold a more natural narrative line on a first draft; something more clinical is often sharper at catching logical gaps or padding on a second pass.

Force a position

AI's instinct is to equivocate. Push back on that directly: ask it to pick one path among several and argue it fully, rather than laying out all sides. You can also ask for a turn partway through the piece – a point where the reader's understanding of the topic genuinely shifts, the way a good magic trick shows you something ordinary before the reveal. It mirrors how an actual expert walks someone through a discovery, rather than just stating the conclusion up front.

The human pass

None of the above replaces an actual edit. AI gets you a serviceable first draft. The read-through afterwards is where it becomes yours or stays generic.

Cut the padding

AI drafts tend to over-explain and then repeat themselves in a tidy wrap-up at the end. If a point's been made once, it doesn't need making again three paragraphs later. Cutting these restatements alone often takes a piece down by a third, sometimes more, without losing a single idea.

Read it aloud

This is the fastest test there is. If you run out of breath partway through a sentence, so will your reader – break it. If a paragraph sounds the same as the one before it, your ear will tell you before your eyes do.

Look at the shape of the page

An editor can often spot AI-written text within about five seconds, purely from the shape of the paragraphs – they tend to be strikingly uniform in length. To break that:

• Let a colloquialism through, if it's one you'd actually use.

• State a measured opinion – "this is the weak part of the argument" – rather than hedging it.

• Vary your punctuation. A colon or a full stop can do a different job than the comma the model defaulted to.

• Drop a single-sentence paragraph after a long block.

  
  

Name your sources

This is the one people skip, and it matters more than any word on the banned list. "Experts say" and "research shows" are exactly the kind of vague attribution that gives a draft away. Say who. Say what study, what year, what the actual number was.

Two real examples worth knowing, since they get cited constantly and rarely with a source attached: Raptive, the publishing platform, surveyed 3,000 US adults and found that trust in an article dropped by close to half the moment readers merely suspected it was AI-written – even when the piece had actually been written by a person. And separately, McKinsey's 2025 research on AI adoption found that document generation and formulaic content creation are among the task categories most likely to be automated as AI use grows across US work hours. Both are specific, both are attributable, and both do more for an argument than "studies show" ever will.

Making it stick

Getting one draft right is one thing. Keeping your writing recognisably yours over months of using these tools is a different problem, and it needs a system rather than a one-off effort.

A before-and-after, for reference

	Raw draft	After the pass	What changed
Opener	"In today's fast-paced world, social media is important for growth."	"Do you want your brand to get noticed? Showing up is the first step."	Filler swapped for a direct question
Evidence	"Many experts agree social media helps businesses reach audiences."	"Raptive's 2025 survey found trust in a brand halves when its content feels AI-made."	Vague attribution replaced with a named, specific source
Voice	"It is necessary to leverage digital platforms."	"I posted daily for thirty days. Our reach doubled."	First person, concrete claim, no jargon
Rhythm	"AI tools can assist with writing. They are fast and efficient."	"We use AI as a starting point. The final call is always ours."	Sentence length varied, generic claim replaced with a specific stance

A scorecard worth running on every draft

• Active voice in at least nine sentences out of ten.

• Zero banned words – no delve, tapestry, testament, telemetry.

• Named evidence – no vague attributions left in.

• No leftover scaffolding – no "in this piece" opener, no "in conclusion" close.

• Genuine sentence variety – not just alternating short and long on a pattern.

• First person, used naturally, not bolted on.

• No triplets – check any list defaults to three and break it.

• A real stance – somewhere in the piece, you've actually committed to an opinion.

  
  

Treat your banned list as a living thing

Every time a draft comes back sounding corny or over-familiar, that's useful information – note the specific phrase and add it to your list. Keep a running record of how your instructions to the model change over time, too. Models update, and a style guide that worked six months ago can quietly stop working without you noticing, so it's worth re-checking your own scorecard against fresh drafts periodically rather than assuming the fit still holds.

The tools get better every year. The thing that doesn't come from the tools is judgement about which sentence earns its place – and that part stays yours, no matter how good the drafting gets.

---