How do I make AI writing sound human?

By Chanelle Segerius-Bruce · Published 21 September 2026

This guide is for anyone using AI to write their marketing copy who wants it to sound like a person wrote it. Emails, social posts, opt-in pages, sales pages. I've used Claude for copywriting since 2023, back when you needed a VPN to reach it from South Africa and there was no paid plan to buy. So everything here comes out of three years of real client copy. Below: what gives AI writing away, how I built my AI Ban List to catch it, why that still isn't enough, and the edit I run on everything before it goes out.

Does AI writing sound robotic?

Out of the box, yes.

A language model writes by predicting the most likely next word. Most likely means safest. So when you ask it for a launch email with no guidance, you get the average of every launch email it has ever seen, polished smooth. It sounds competent. And there's nobody home.

Your reader notices it immediately, even if they can't pinpoint why. It might be the word "seamless", or a neat little run of three benefits. And once people realise something was written with AI, that's what they focus on instead of your message. They're busy spotting the AI, so they stop reading what you actually wrote.

Most people blame the tool. They switch from ChatGPT to Claude, or they spend an evening rewriting the prompt. Lately I've noticed a lot of people moaning about Claude. I haven't had any problems with it. I think it's because they haven't set up their AI brain, or AI operating system, and they haven't trained it on their voice before asking it to write.

You can swap the tool. You can rewrite the prompt. It still reads as AI. The words are what give you away, and after the words, the rhythm.

Where I'm writing from

I started using Claude in 2023 on a VPN. It was free back then, and when you hit the session limit you waited, because there was no paid plan to upgrade to. I used it for the copy I was already writing for launches: webinar invites, show-up emails, post-webinar sales emails and sales pages.

When I started, I had one-to-one mentoring from a professional copywriter in the US who has ghostwritten New York Times bestsellers. He taught me to wrangle Claude into writing like someone else. That's still what brand voice training comes down to, and my own methods have come out of the three years of testing and client copy since.

The volume grew. In my work with one client I was writing up to around 130 emails a month with Claude: the invites and show-up reminders for live webinars, and the sales emails after them. Different presenters ran the webinars, so each presenter got their own brand voice in Claude, and the company got one too. I couldn't have written that volume by hand and kept it sounding like several different people. With AI you can, once you've trained it. There's a copywriter on that team now and I manage the copy instead of writing all of it. I wrote more about how that email volume fits into a launch in what's working in online course launches this year.

Alongside that: registration pages, opt-in pages, social posts, and a sales page for a client event that I wrote using a sales page skill I built in Claude.

A lot of that early copy needed heavy editing. That editing is where everything in this post came from.

What gives AI writing away

AI's tells come in a few kinds. Words are the easiest to spot. Rhythm is the one that catches people out.

The words

These are the ones everybody knows: delve, tapestry, seamless, unlock. They show up because they fit almost any sentence without committing to anything.

You can measure this. Researchers who analysed more than 15 million biomedical abstracts found a sharp jump in certain "style words" after ChatGPT arrived, and estimated that at least 13.5% of 2024 abstracts had been through a language model (Kobak and colleagues, 2024, later published in Science Advances). So AI's favourite words can be counted. They also change, which matters more than people realise. I'll come back to that.

The phrases

"In today's fast-paced world." "Let's be honest." "Now more than ever." "At the end of the day." Stock openers and closers that take up a line and say nothing.

Some are newer. In June this year I started noticing everyone suddenly had a moat, and was doing everything quietly. I didn't know everyone did things so quietly. I thought we were all trying to let the world know about what we do. LOL

The structures

Structures are harder to catch, because every word can be fine and the sentence still reads as generated.

• The flip. "It's not just a strategy, it's a mindset." Wikipedia's editors list it in their field guide to AI writing, and it's number one on my list.
• Everything in threes. Three benefits in a row, or three adjectives stacked on one noun. People write in twos and fours and ones as well.
• The em dash. Dropped in where a comma or a full stop would do. More on this one below.
• Announcing a count. "Two things make this work." It makes writing sound assembled. No detector I've tried flags it.

The em dash, and what to use instead

If you're not sure what an em dash is, it's the long dash: —. A hyphen is the short one, -, and it joins words together, as in "well-known" or "follow-up email". There's also the en dash, –, which sits in between and is mostly used for ranges, like 9–5.

In most sentences an em dash does the same job as a comma or a pair of brackets, and a full stop often does it better. Take this sentence:

• With an em dash: The webinar starts at 7pm — bring your questions.
• With a spaced hyphen or en dash: The webinar starts at 7pm - bring your questions. (Or 7pm – bring.) Same job as the em dash, same effect on the reader.
• With a comma: The webinar starts at 7pm, so bring your questions.
• With brackets: The webinar starts at 7pm (bring your questions).
• With a full stop: The webinar starts at 7pm. Bring your questions.

My guess is AI loves the em dash because it's all over the polished writing it learnt from, and it's an easy way to bolt one thought onto another without deciding how they connect. I've dropped it entirely. Readers link it with AI now, and a comma or a full stop nearly always does the job. The full stop is my favourite, because it gives you short sentences, and short sentences sound like a person talking.

The smoothness

My ban list calls this "smooth but empty", and it's the biggest tell of 2026. Copy that's fluent and plausible, with no opinion and no specific detail. Every sentence roughly the same length, arriving at the same pace. You can strip out every banned word and still have it.

How I built my AI Ban List

My list built up slowly, from the copy I write for my own business and for my clients.

Whenever a word kept turning up in drafts, and then in everybody else's posts, it went on the list. And the words keep evolving. New ones show up as a pattern, and I notice them because I write a lot of copy for myself and my clients, and I'm constantly reading social media posts and marketing emails. Once you start to recognise a pattern, you can't unsee it. The classics like "delve" were joined from around mid-2024 by a newer set: fostering, showcasing, enhance, robust. So I started dating the tiers. My list has a section headed as current offenders from mid-2024 onwards, because a list without dates goes stale without you noticing.

The clearest proof I have of how fast the list changes is sitting in my own workspace. In April 2025 I had Claude put together a creative brief for my brand. It recommended "leverage", "transform", "streamline" and "innovative" as my top vocabulary. All four are banned on my list now. What read as confident business language a year and a half ago reads as AI today.

Some of what I banned was mine. Nine phrases on the list used to be my own signature lines, things I said for years, like "let me know if this resonates" and "stop trading time for money". They spread so far through coaching copy that they now read as generic, so they had to go too.

Then there's what I call the Consensus Middle Rule. AI tends to pick the most common, safe word for any context. The rule tells it to resist that and reach for the accurate word that isn't the obvious one.

The structures took me longest. They're much harder to police than words, even in my own copy. On one of my own sales pages, a fresh read found five rule-of-three lists I'd missed completely. Now a fresh pair of eyes checks everything before I call it done.

And it's a working document. If a word is on there that you genuinely say, take it off. I say "game-changer", and so does Chantal, my developer, and it stays in my copy. The list is there to catch the words AI puts in your mouth. The ones you already use are yours to keep.

Why a clean ban-list pass still isn't enough

I learnt this one the hard way, in July.

I'd written a guide for this blog about using AI to make sense of personality test results. It passed my ban list. It passed a voice check against my brand voice document. Two fresh reads went over it. Then I put it through an AI detector and it came back 100% AI.

Not one banned word in it. But the sentences were tidy and all about the same length, arriving at an even pace, and that smoothness is what detectors react to. We rewrote it for rhythm, with short sentences next to long ones and more of how I actually talk, and it dropped to about 2 to 3% AI.

Then I caught a bigger problem. The writing had invented a neat storyline, that my personality tests all contradicted each other. They don't. They cross over a lot, and that overlap was the whole point of the post. The false certainty read beautifully and it wasn't true. No detector or word list would have found it. I found it because I knew what my tests said.

So now the detector is its own step, after the ban list and the voice check, never instead of them. I use an open-source detector that runs offline on my laptop, so nothing I write gets uploaded anywhere, and I aim for 90% human or better. I also check rhythm: a small script measures how much my sentence length swings and compares it with blog posts I wrote by hand years ago.

One thing to be clear on. I treat the detector as a smoothness check. It isn't a way to hide that AI helped, and a good score doesn't make copy good. Detectors aren't proof either way. OpenAI withdrew its own AI text classifier in 2023, citing its low accuracy; in OpenAI's own tests it caught only 26% of AI-written text (OpenAI). A high score tells me where the copy has gone flat. That's all I ask of it.

How I make AI copy sound human, step by step

This is my actual process, in the order I do it.

1. Train it on your voice before you ask it to write anything. I think of it like onboarding a new copywriter. You wouldn't expect a new hire to sound like you on day one, so don't expect it of Claude. Give it your story, your phrases, how you talk, who you're talking to. When I started there were no Projects and no Skills, only a chat window, and brand voice was still what made the difference. It still is.

2. Keep your voice separate from your offers. One brand voice, then a separate document for each offer, because each offer can have a different ideal client. Mush them together and the copy goes vague.

3. Write from transcripts. I highly recommend this one. Brain dump by talking instead of typing. You can use voice-to-text in the Apple Notes app, or a tool like Wispr Flow. I use Talk to Text in ClickUp Brain, ClickUp's AI voice-to-text add-on, and a lot of mine get recorded on the beach with the dogs. Transcripts from Zoom calls and podcast interviews work just as well, because they capture how you actually talk. Your content should carry your own point of view and start from your voice note or transcript. AI helps you shape it from there. I think of Claude as my formatter. The ideas come from me. The moment you ask it for ideas instead of organising yours, the content stops sounding like you. My rule of thumb is 90% you, 10% technology.

4. Load the ban list before it writes. Put your ban list in the instructions of a Claude Project or Skill (or ChatGPT's custom instructions, or a Gemini Gem) so it applies to every draft automatically.

5. Read it out loud. This is the two-minute edit I use on everything. Where you stumble, there's an AI-ism, so cut it. Where you naturally pause because something's interesting, that's probably your real opening, so move it up. Then ask of every line: would I actually say this? Your mouth knows before your brain does.

6. Don't polish it back and forth in the same thread. Asking the same conversation to "make it sound more human" five times creates an echo chamber that polishes the same wrong thing. Step away, ramble what you meant into a fresh voice note, and bring that back instead.

7. Check the judgement yourself. In July I had twelve emails drafted for my own list. The first draft was too harsh. One of them told people they could leave before I'd given them a reason to stay. The copy was competent. It was also wrong for the relationship, and only I could see that. AI writes the words. You decide whether they're right for the people reading them.

8. Run the detector last. Only after all of the above. If it flags a section, that's where the rhythm has gone smooth, so rewrite that part.

If you want something to start with today, this is a short version of the kind of instruction I give Claude before it writes. It's nowhere near the full list, but it helps:

What the edit looks like on an email

This is an illustration I've written for this post, not a client's email. It's the opening of a show-up email for a live masterclass, the kind I've written hundreds of.

The default AI version:

In today's fast-paced world, showing up for your business has never been more important. Tonight's masterclass is designed to help you unlock the strategies, tools and mindset you need to transform your launch. It's not just a webinar, it's a turning point.

Every tell is in there. The stock opener, the run of three, "unlock" and "transform", the flip at the end. Even the rhythm is flat, with each sentence landing at much the same length and pace as the one before, which is the part most people never notice until someone points it out.

After the edit:

We go live at 7pm. Bring the launch you're planning, even if it's only a date in your diary so far. I'll show you the first email I'd send and you can ask me about yours in the chat. Same link as this morning.

Shorter. A time and one clear instruction, and it sounds like a person sending a message. Read both out loud and you'll hear it.

Does it matter whether you use Claude, ChatGPT or Gemini?

A bit, yes. I've tested all three extensively, and for copy that sounds genuinely human, Claude is still the winner. But only after a lot of brand voice training, which is something I set up for my clients. Without it, Claude falls back to the safe middle like the others.

ChatGPT is a close second. Gemini comes third. It has its own uses, but I wouldn't use it for copywriting.

Whichever one you pick, the setup around it does most of the work: your voice, your raw material, your ban list, and your ear at the end. The instructions in my ban list paste into Claude, ChatGPT or Gemini, because the tells are the same in all three.

FAQ

Why does AI writing all sound the same?

Because the models lean toward the same safe, average phrasing, and lots of people are using the same tools with the same prompts. I'm seeing it in programme and challenge names too, which are starting to share the same words. The more of your own material and opinions you give it, the less it sounds like everyone else.

Can people tell if you used AI to write something?

Often, yes, even when they can't say why. Readers pick up on the patterns in this post, the smoothness most of all. But nobody can prove it from the text alone. Detectors get it wrong in both directions, and Wikipedia's own guide warns that plenty of human writing shows the same signs. Better to write something so specific to you that the question never comes up.

Does Google penalise AI-written content?

Not for being AI-assisted. Google's guidance says generative AI can be useful for research and for adding structure to original content, but using it to churn out lots of pages without adding value for readers may break its spam policy on scaled content (Google Search Central). In my experience, the same things that make copy sound human, like real examples and a clear point of view, are what make it worth reading in the first place.

Do AI humaniser tools work?

They can shuffle words and vary sentence length. What they can't do is add a detail they don't know, or catch a claim that isn't true about your business, which is exactly what went wrong in my personality test guide. I'd rather fix copy at the source than run it through a second tool and hope.

Should AI write your copy for you?

I don't believe in handing your thinking to AI. You should know your topic well enough to have opinions about it, and AI makes it easier to get them out of your head and into an email. I use it for volume and structure. The ideas and the final say stay with me.

Key takeaways

• AI writing sounds robotic by default because it chooses the safest word and the smoothest rhythm.
• A ban list catches the words and structures. It won't catch smoothness or a claim that isn't true.
• Train your voice and feed it your own raw material before it writes. Then edit by reading aloud.
• Use an AI detector last, as a smoothness check, and never as a replacement for your own judgement.