Within a sentence, AI writing that doesn’t sound like AI is often easy to recognise. The email opens with “In today’s fast-paced business environment” and you’ve already filed it as noise before you consciously decided to.
That is the problem with most machine-generated drafts. They are grammatically clean and completely dead. And the thing everyone reaches for first, a longer and cleverer prompt, barely helps. This is not a prompting problem. It is a context problem.
This post covers what actually makes a draft read as machine-written, the specific patterns to cut, how to give a model enough of your voice that it stops guessing, and the edit pass that takes a passable draft to something your clients would believe you typed yourself. No detector tools. No tricks. Just the mechanics.
Key Takeaways
- AI writing that doesn’t sound like AI starts with context, not prompts. A model with no samples of your voice defaults to the internet average.
- The tells are structural, not just vocabulary. Three-part lists, hedged claims, symmetrical paragraphs and tidy summaries give a draft away faster than any single word.
- Reading a draft aloud catches most of the slop in one pass. If you stumble or feel embarrassed saying it, cut the sentence.
- Specifics kill the machine tone faster than anything else. Swap every abstract noun for a number, a name, a place or a time.
- AI content detectors are unreliable in both directions, and Google has said it rewards helpful content regardless of how it was produced.
- A one-page voice file with five samples of your real writing does more for tone than a 400-word prompt retyped from scratch each session.
- Dictate the first draft. Speaking at 150 words a minute produces your natural rhythm, so the model tidies instead of inventing.
- Hand the AI the structure and the grunt work. Keep the opinions, the stories and the numbers, because that part nobody can generate.
What Makes AI Writing Sound Like AI in the First Place
A language model is trained on an enormous pile of text and asked to produce the most probable next word. That single design fact explains almost every complaint you have about the output.
Probable means average. The average business email, the average LinkedIn post, the average blog introduction. If you ask for something with no other information, you get the middle of the distribution back. Nobody writes like the middle of the distribution, which is exactly why it reads as artificial.
Then there is the stakes problem. You’ve lost a client. You’ve had a quote rejected. You’ve written an email at 11 pm knowing the next reply decides whether payroll clears. That experience shows up in how you write: what you refuse to promise, which word you choose when you’re being careful, where you get blunt. A model has never had a bad month. So it hedges everywhere, promises softly, and never commits to anything sharp enough to be wrong.
The third and biggest cause is the one most people never fix. The model doesn’t know anything about your business. It doesn’t know your average job value, your two best clients, the thing your competitor does that annoys you, or the phrase your team uses for a job that goes sideways. Faced with a gap, it fills the gap with something generic, because generic is the only safe guess available.
Why better prompts stop working
Most founders respond to bad output by writing longer prompts. “Write in a warm, conversational, professional tone. Avoid corporate jargon. Be authentic.”
Those instructions are adjectives, and adjectives are interpreted at the average too. “Warm and conversational” produces the average of every piece of writing that has ever been described as warm and conversational. You are asking the model to guess at your voice using a word you’d struggle to define yourself.
Rules beat adjectives. “Paragraphs never run more than two sentences. Never use an exclamation mark. Always name the actual number instead of saying significant.” Those are checkable. The model can follow them. “Be authentic” it cannot follow, because it has no idea what your authentic sounds like.
That gap between what you asked for and what you got is the same gap that shows up everywhere else AI disappoints people, and it is why so many small businesses conclude the tools don’t work. Usually the tool is fine. It’s flying blind. I’ve written about that pattern in more detail in why AI automation projects fail in small businesses.
The Tells I Look for in Every Draft
Once you know the patterns, you cannot unsee them. Here is the list I run through on anything a model has touched, grouped by how obvious they are.
The structural tells
The three-part list. Everything comes in threes. “Faster, cheaper, and more reliable.” “Plan, build, and refine.” One or two are fine. When every second sentence resolves into a neat triple, the rhythm becomes hypnotic in the worst way. Real writing is lopsided. Sometimes you list four things. Sometimes one.
The balance sentence. “While automation saves time, it’s important to remember that human oversight matters.” Every claim gets immediately counterweighted so nothing is ever quite asserted. This is the single most common reason a draft feels like it says nothing.
The self-summary. Each section ends by restating what the section just said. A human writer finishes a point and moves on. A model closes the loop every time, because closure is probable.
Symmetrical paragraphs. Four paragraphs, every three sentences, each roughly the same length. Human paragraphs are ragged. Some run six sentences. Some run two words.
The tidy bow. The final paragraph reassures you that with the right approach, success is achievable. Nobody writes their own ending like that unless they’ve run out of things to say.
The vocabulary tells
There is a set of words that appear constantly in machine drafts and almost never in speech. Furthermore. Moreover. Myriad. Plethora. Utilise. Meticulous. Boasts. Testament to. Stands as. In conclusion. None of these is wrong. They are just not words you would say out loud to a customer standing in front of you.
The em dash deserves its own mention. Models love the long dash and use it as connective tissue between half-formed clauses. If a draft has four of them per page and you have never once typed one deliberately, that’s a fingerprint.
The other vocabulary tell is unearned enthusiasm. Words like powerful, exciting and incredible attached to things that are merely fine. Enthusiasm has to be paid for with a specific reason; otherwise, it reads as a sales brochure written by someone who has not used the product.
The content tells
Zero specifics. No names, no dates, no dollar figures, no place names, no times of day. This is the loudest tell of all, and it is the one worth fixing first. “Many businesses struggle with slow lead follow-up” is a machine sentence. “You quoted 40 jobs last month and you’re still the only one who knows which six are worth chasing” is a human one, because a human had to have been there.
No opinion. The draft describes and never argues. It presents both sides of things that only have one side. If you cannot tell what the writer thinks after reading 800 words, a person did not write it, or a person wrote it without caring.
Advice that fits anyone. If you could swap out your industry for a different one and the whole piece still works unchanged, it is generic by construction. Specificity is not decoration. It is the proof that you know something.
The Fastest Route to AI Writing That Doesn’t Sound Like AI: Give It Your Voice
Every fix above is downstream of one thing. The model needs raw material from you, and most people never give it any.
Here is what actually moves the output, in order of impact.
Build a voice file once, reuse it forever
Collect five to ten pieces of writing you’re genuinely happy with. Sent emails work better than polished marketing copy, because the emails are how you actually sound. A proposal you wrote at speed. A text you sent a client explaining a delay. A LinkedIn post that got real replies.
Put them in one document. Above them, write the rules, not the adjectives:
- Sentence length: mostly short. Two sentences per paragraph maximum in emails.
- Never say: solutions, in today’s world, I hope this email finds you well.
- Always name the number instead of saying a lot or significant.
- Contractions on. It’s, don’t, you’re.
- No exclamation marks.
- Open with the point. Never with context-setting.
That document is now the thing you hand the model at the start, every time. Not a paragraph you retype from memory in each new chat, which is what most people do and why the quality swings wildly session to session. Consistency comes from a file that doesn’t change, not from a prompt you improvise.
This is the same principle behind giving any AI system real business context rather than starting fresh in every conversation, which I’ve covered in how small businesses should approach AI adoption.
Dictate the first draft
You speak at around 150 words a minute. You type at 40 if you’re quick. That difference matters less than the other thing dictation gives you, which is your actual rhythm.
Open a voice note, talk for two minutes about what you want to say, no structure, no editing, just talk like you’re explaining it to someone in the passenger seat. Then hand the model the transcript and say: tidy this into an email, keep my phrasing wherever you can, don’t add anything I didn’t say.
The output is now shaped by you and cleaned by the machine, which is the correct division of labour. The reverse, where the machine generates and you edit, means you spend your time deleting things that were never yours.
Feed it the raw material, not a summary
If you’re writing about a client win, don’t describe the win in a sentence and ask for 600 words. Paste the actual call transcript, your notes, the invoice amount, and the two things the client said that stuck with you. The model’s job is arrangement, not invention. Every fact you don’t supply is a fact it will make up, and made-up facts are always generic ones.
Give it something to react to
The single most useful instruction I use is not “write this”. It’s “here’s my draft, tell me where it’s flabby”. A model is far better at critique than at creation, because critique is bounded by something real. You still own the words. It finds the three sentences that say nothing.
The Edit Pass That Turns a Generic Draft Into AI Writing That Doesn’t Sound Like AI
Assume every draft needs this pass. It takes about five minutes on an email and twenty on a long article. Do it in this order.
1. Read it aloud. Actually out loud. Not in your head. This catches more than every other step combined. Where you stumble, the sentence is badly built. Where you run out of breath, it’s too long. Where you feel a flicker of embarrassment saying it, that sentence is the machine talking. Cut it. Nielsen Norman Group’s research on how people read on the web found that readers scan rather than read in full, which means the sentences that make you stumble are the ones nobody will finish.
2. Delete the first paragraph. Nine times out of ten it’s throat-clearing. The real opening is your second paragraph. Try it on any draft you have sitting around and see how often it improves.
3. Hunt the abstract nouns. Circle every word like efficiency, growth, quality, results, performance. Each one is a hole where a specific used to be. Replace it: not “improved efficiency” but “cut the Monday admin block from three hours to forty minutes”. If you can’t replace it with a specific, you may not have anything to say in that sentence.
4. Break the rhythm deliberately. Find two paragraphs in a row that are the same length and shorten one to a single line. Add a three-word sentence somewhere. Start one sentence with “And” or “But”. The variation is what makes it sound like breathing rather than a metronome.
5. Cut every self-summary. Any sentence that begins “In other words”, “Ultimately”, “The key takeaway here is” or “By doing this, you can”. Delete the sentence. The paragraph above it already made the point.
6. Add one thing only you could know. A number from your own business. The name of the suburb. What the client said on the phone. The detail that would be wrong if someone else wrote it. One per section is enough.
7. The mate test. Would you send this, exactly as written, to someone you know well? If the answer is “not quite, I’d loosen it up first”, then loosen it up first. That instinct is correct, and you should follow it every time.
What to Hand the AI, and What You Should Keep
Not everything needs your voice on it. A lot of writing in a business is plumbing, and plumbing should be automated without guilt.
Hand it over freely: meeting notes into action items, a long transcript into a summary, a proposal reformatted for a different client, ten subject line variations, turning a rough set of bullet points into a readable update, translating your technical explanation into plain language for a customer, first drafts of routine confirmations and scheduling emails.
Keep for yourself: the opinion, the story, the bad news, the apology, the price conversation, anything with a legal edge, and the specific numbers out of your own business. Those are the four or five paragraphs a year that actually build a relationship, and they are exactly the paragraphs a model has no material for.
The middle ground is where most of the value sits. On a database reactivation build for a finance broker, the messages that got replies did not read like marketing. They read like someone typing on a Tuesday afternoon: short, slightly informal, one specific reference each. The AI wrote and sent them at volume. The voice, the offer and the reason to reach out were the client’s. From 319 contacts the team had already written off, that recovered $49,000. The system did the work. The human supplied the thing that made anyone reply.
That is the split worth internalising. AI is very good at doing a known thing many times. It is poor at deciding what the thing should be. If you’re thinking about where else that split applies in your operation, AI agents for small business walks through the same logic applied to work that isn’t writing.
A note on detectors
Founders ask me constantly whether they’ll get caught. The honest answer is that AI content detectors are unreliable in both directions. They flag human writing that happens to be clean and structured, and they miss machine writing that has been edited properly. Plenty of well-documented cases exist of students and professionals being wrongly accused by them.
More usefully, Google’s Search Central team has stated directly that its focus is the quality of content rather than how it was produced. Content made primarily to game rankings is the problem, not automation as such.
So stop optimising to beat a detector. Optimise so a human reads the whole thing. Those two goals point in the same direction anyway, and only one of them is measurable by whether people reply.
The Real Fix Is Upstream
Here is the uncomfortable part. Every technique in this post is a workaround for the same missing piece.
You are re-explaining your business to a blank chat window several times a day. Who you are, who you sell to, what you charge, how you talk, what happened last week. Then you edit generic output back into something that sounds like you, and tomorrow you start from zero and do it again.
That is not an AI problem. That is a system problem, and it is the same one that has you answering questions your team should be able to answer themselves. The knowledge lives in your head, so everything routes through your head, including your own tone of voice.
The fix is to get the context out of your head once and into something the AI reads before every task. Your positioning, your rules, your real numbers, your actual writing. That is how you get AI writing that doesn’t sound like AI: a brain for the business, with a workforce that acts on it. When that exists, good writing stops being a five-minute editing job you do forty times a week and becomes the default output. Same for the rest of the work you’re currently doing by hand.
You do not need to build all of it to feel the difference. Start with the voice file described above. One page, ten minutes, used every time. It is the smallest version of the same idea, and it will do more for your writing this week than any prompt you copy off the internet.
If you’d rather talk it through than work it out alone, book a 30-minute Discovery Call. No pitch deck. I’ll ask what’s currently eating your week, tell you where AI genuinely helps and where it will waste your time, and you can take the answers away and do it yourself if that’s the right call.
Frequently Asked Questions
Can Google detect AI writing and will it hurt my rankings?
Google has said publicly that it assesses content quality rather than production method. Automation used to mass-produce low-value pages aimed at rankings is against its guidelines. A well-researched, specific, genuinely useful article that a model helped draft is not. In practice, the risk is not detection. It is publishing thin, generic content that nobody finishes reading and nobody links to.
Do AI content detectors actually work?
Not reliably. They produce false positives on human writing that is clean and well-structured, and false negatives on machine writing that has been properly edited. There have been many documented cases of people wrongly accused by them. I would not make a business decision on a detector score. Read the draft aloud instead. Your ear is a better instrument than the tool.
How do I train AI to write in my tone of voice?
Give it samples rather than adjectives. Collect five to ten pieces of your own writing you’re happy with, ideally sent emails rather than polished marketing copy, and store them in one document alongside a short list of concrete rules: sentence length, words you never use, whether you use contractions. Hand that same file over at the start of every task instead of retyping instructions from memory.
What are the most obvious signs a piece was written by AI?
Constant three-part lists, sentences that immediately counterweight every claim, paragraphs that are all the same length, a summary sentence at the end of every section, and a complete absence of names, numbers, dates or places. Vocabulary matters less than structure. A draft with no specifics in it was almost certainly written by something that has never been in the room.
Should I let AI write my emails to clients?
Routine ones, yes. Confirmations, scheduling, follow-ups, sending information someone asked for. Those are plumbing. Keep the ones that carry weight: bad news, apologies, price conversations, anything with a legal edge, and anything where the relationship is the point. Those are a handful of emails a year, and they are exactly the ones a model has no material to write from.
How long does it take to set up a voice profile?
About ten minutes for a usable first version. Copy five sent emails into a document, write six rules underneath them about how you want things phrased, and save it somewhere you can grab quickly. Refine it whenever you catch yourself correcting the same thing twice. The value is in reusing the identical file, not in making it perfect on day one.
Is it faster to write it myself than to edit AI writing?
For short, high-stakes pieces, often yes, and you should just write them. Editing is faster for long or repetitive work: articles, summaries, proposals built from a template, anything where the structure is known. The trick is to reverse the order. Dictate two minutes of rough thinking first and have the AI tidy it, rather than generating from scratch and deleting what was never yours.
About Octavius
Titus Mulquiney is the founder of Octavius AI, where he builds AI brains and AI workforces for founder-led businesses stuck running everything out of their own head. Twenty years in marketing, ex-Sony product manager, ex-GM Zeal NZ. Based in Auckland, working with operators across NZ, Australia, and the US. Connect on LinkedIn.