Guides · AI at work

How to Make AI Writing Sound Human: Find the Tells and Fix Them

By Josh Croom · October 27, 2026 · 11 min read

To make AI writing sound human, find the patterns that give it away, check every fact it added, and rewrite the first and last lines in your own words. Then give your AI a standing rule against those patterns, because readers trust the sender less once they know a machine wrote most of it.

The thank-you note that almost went out

Thursday, 4:10 PM. A nonprofit director asks AI to draft a year-end thank-you to her donors. The draft looks polished, and she nearly sends it as written. Then one line stops her: “Because of you, every family on our waiting list got help this fall.” Nobody gave the AI that fact, so it came from nowhere, and it sounds just as sure of itself as the true lines around it.

She cuts the line and puts in the real count from her own records. Then she reads the rest out loud and notices a second problem. The note could have come from any charity in the country, because nothing in it names a person, a place or a moment that only her donors would know. Made-up facts and writing that sounds like nobody are the two problems to fix in every AI draft.

What readers notice

Two studies from 2025 measured what AI-heavy writing costs, both for the person who sends it and for the person who gets it.

83%judged the sender sincerewhen AI helped a little with a note of praise
40% to 52%judged the sender sincerewhen AI did a lot of the writing
40%of U.S. desk workers got sloppy AI workin the past month
2 hoursto sort out each caseon average, in the same survey

Sources: Cardon and Coman, International Journal of Business Communication, 2025; Niederhoffer and others, BetterUp Labs and Stanford Social Media Lab, Harvard Business Review, September 2025.

In the first study, roughly 1,100 office workers read a boss's note praising a team. Each reader was told whether AI had helped a little, some or a lot, and then rated the boss. Readers held the person who signed the note responsible for it, so the cost of heavy AI help landed on the boss. In the second, workers who got thin AI work from a coworker said each case took about two hours to sort out, and 42% said they saw the sender as less trustworthy afterward.

Nine patterns that give AI writing away

These tells show up in AI drafts again and again, and once you've read the list, you'll start spotting them everywhere. Each one describes what a sentence does, so a list of banned words won't catch them on its own.

Lines about anyone: a sentence with no name, date or number in it could come from any business, so another company could sign it unchanged.
Setup lines: a sentence like “A note on our hours” does nothing but point at the next one, so delete it and say the thing it points at.
Self-praise: calling your own service friendly or your prices fair asks the reader to take your word for it, when showing the work would let them decide.
Unasked comfort: a line like “Don't worry, we've got you” plants the very worry it means to calm.
Hedges: phrases like “it seems” or “we believe” on every claim make you sound unsure of your own business, so state what you know, and when you don't know something, say that once.
Empty sign-offs: a closing like “Hope this helps!” says nothing at all, and a real next step would do the job better.
Built-to-quote lines: a sentence balanced so neatly that it sounds like a slogan is one nobody would say out loud across a counter.
Hidden actors: a line like “Your request was processed” hides who did what, so name the person and the action instead.
Long dashes and emoji: swap each long dash for a comma, a period or a new sentence, and cut the emoji from anything that isn't a text to a friend.

One note, before and after the edit

Here's a donor thank-you from an AI draft, and the same note after a person checked every fact and rewrote the first and last lines.

The AI draft

“Thank you for your amazing support this year!”
“Because of you, every family on our waiting list got help this fall.”
“We're so grateful for partners like you.”
“Hope this helps you see the impact of your gift!”

After the edit

“Thank you for the gift you sent in March, the week our van broke down.”
“Your gift paid for the diapers we handed out at the spring drive.”
“Maria, who runs our Saturday pantry, asked me to pass on her thanks too.”
“I'll send you photos from the fall drive in November.”

How to edit an AI draft so it sounds like you

Work through these in order on any draft that's going to a customer, a donor or the public. A short post needs a quick pass, and anything with a price, a promise or a legal question gets every step.

Step 1
Learn the tells
Go through the nine patterns above once, slowly, with a recent draft of your own beside them. After that, you'll catch most of them on a first read, the same way you spot a typo in a word you know well.
Step 2
Run the swap test
Put another company's name where yours is and read the draft again. If it still fits, it says nothing about you, and it needs a real detail before it goes anywhere.
Step 3
Cover the logo and read it out loud
Ask whether a regular customer would know the message came from you. If they wouldn't, it's slop, whoever wrote it, and reading out loud is the fastest way to hear it.
Step 4
Circle every fact, name and number
Check each one against your own records, and cut anything you can't check. If nobody can say where a number came from, treat it as made up, because a made-up number can look just as exact as a real one.
Step 5
Ask what it costs if it's wrong
A social post or a team note only needs a quick read. Anything with a price, a promise, the law, health or money goes to a person who knows the answer, and that person checks it before it goes out.
Step 6
Rewrite the first and last lines yourself
Those two lines carry your voice more than any others. Add one real detail, like a name, a date or the next step, because a real detail is the one thing AI can't supply on its own.
Step 7
Have AI check itself
Paste the draft and the list of tells into the check prompt below. Fix every tell it finds, not just the worst one, and then read the result out loud one more time.
Step 8
Stop the tells at the source
Add the standing rule below to the instructions that your AI tool reads before every chat. The rule cuts the tells, but it won't make a draft true, so a person still reads every one.

Here's the check prompt at work. You paste the draft and your list of tells, and the AI points at the lines that still sound like a machine.

Read this draft one sentence at a time and count every phrase that matches my list of tells. Then tell me which sentences still sound like a machine wrote them, and why.

Sentence 1 praises the reader with no detail. Sentence 2 states a fact you didn't give me. The last line is a sign-off that says nothing.

Good. Cut sentence 2, and suggest a real next step for the last line.

Sample

A sample chat using the check prompt.

Two prompts to copy

The first prompt checks a draft you already have. The second goes into your standing instructions, so the tells stop showing up in the first place.

Your list of words is personal. Start with the ones that make you wince in your own drafts, and add one every time you catch the AI using a word you'd never say out loud. If your team runs the same AI job every week, save these lines inside it. Our guide to ChatGPT prompt templates shows how to set that up once, so everyone on the team works from the same rule.

To have AI check a draft: Read this draft one sentence at a time: (paste). Count every phrase that matches this list: (paste the nine tells). Then tell me which sentences still sound like a machine wrote them, and why.
For your standing instructions: Never write in these patterns: (paste the nine tells). Never use these words: (your list). Keep my facts and my meaning exactly as I give them. Never add a fact, a number or a claim I didn't give you. If something is missing, ask me.

Should you use an AI humanizer or an AI detector?

Tools that promise to humanize AI text or to detect it aim at the wrong reader. Detectors have a poor record. OpenAI pulled its own AI text classifier on July 20, 2023, because of its low rate of accuracy. In OpenAI's own tests, it caught 26% of AI-written text and wrongly flagged 9% of human writing as AI. In 2023, a Stanford team tested seven popular detectors on essays written by people who learned English as a second language, and on average the detectors wrongly labeled 61% of those essays as AI-written.

Slipping past a detector doesn't tell you whether a person will believe you. Run the swap test and the read-aloud test instead, because those measure what your reader notices. Google's search guidance points the same way. It says appropriate use of AI or automation isn't against its guidelines, and it treats content made mainly to manipulate search rankings as spam.

What AI should never write for you

A thank-you to a longtime donor, an apology to an upset customer and hard news for a client all depend on the reader believing you meant every word. Write those yourself, even if they come out shorter and plainer than an AI draft would.

Never trust a number just because it sounds exact, and trace each one back to your own records. Keep card numbers, home addresses and anything private out of AI tools completely.

Check what's already out there

Once a quarter, pull your last two emails, your last two posts and the top of your home page, and count the tells in each. Two or more on one piece means it fails. Rewrite the worst one the way you'd say it to a customer standing in front of you, and you'll hear the difference right away. If the worst one is the text you send when nobody picks up, our speed to lead guide has a sample worth starting from.

Talk it through with us

We run Google Ads and build websites for small businesses, nonprofits and online stores, and every call and form our ads bring in shows up in your portal with what it cost. If you'd like to talk about what your ads and pages say and who they reach, book a call with Miguel. He'll ask what you sell and how customers find you today.

Questions people ask

How do you make AI writing sound more human?

Start by finding the patterns that give AI away, such as lines that could fit any business, hedges on every claim and sign-offs that say nothing. Check every fact it added, rewrite the first and last lines yourself, and add one real detail only you would know.

Can people tell when writing is AI-generated?

Software often can't, which is why OpenAI pulled its own detector for low accuracy. Readers do react once they know, though. In a 2025 study, readers were told how much AI helped write a manager's note. When AI helped a little, 83% judged the manager sincere, and when AI did a lot of the writing, only 40% to 52% did.

Do AI detectors work?

Not reliably. OpenAI pulled its own detector in 2023 for low accuracy, after it caught only 26% of AI-written text and wrongly flagged 9% of human writing. Write for the person who will read it, since a detector can't tell you whether that person will believe you.

Does Google penalize AI-written content?

Google's guidance says appropriate use of AI or automation isn't against its guidelines. It treats content made mainly to manipulate search rankings as spam.

What should I tell AI so it writes like me?

Give it a standing rule: never write in the patterns on your list, never use the words you'd never say, keep your facts exactly as given and ask when something is missing. Then paste one message you wrote yourself as an example.

See how we build websites for small businesses and nonprofits

How to Make AI Writing Sound Human: Find the Tells and Fix Them

The thank-you note that almost went out

What readers notice

Nine patterns that give AI writing away

One note, before and after the edit

How to edit an AI draft so it sounds like you

Learn the tells

Run the swap test

Cover the logo and read it out loud

Circle every fact, name and number

Ask what it costs if it's wrong

Rewrite the first and last lines yourself

Have AI check itself

Stop the tells at the source