You asked an AI tool for a blog post and got 800 words that sound like every LinkedIn post you scrolled past this morning. Your product is "seamless" now, a word no customer has ever said out loud.
So you typed "make it sound more human." Back came a cheerful "Absolutely!" and the same post with a rocket emoji on top.
You asked it to drop the dashes. It apologized, with a dash.
5 hours later you're rewriting every line yourself, the exact job you were trying to hand off.
You came here to humanize AI content, so let's get you out of that loop. My own loop ran for 18 months, with Claude projects that wrote like a polite stranger, all lovely manners and no opinions.
As of October 2026, my AI clone has published 750+ videos, every script drafted in a single Claude project that sounds like me. Not one of them opens with a rocket emoji.
Asking for "more human" after the draft is written is the slow way around. The first 3 of my 7 steps happen before the AI writes a word.
What humanizing means for your reader
Humanizing AI content means editing an AI draft until it sounds like one real person with firsthand experience wrote it. You’re writing for one judge, the reader who’s deciding whether to hand you an email address or a credit card.
A 2025 Raptive study of 3,000 U.S. adults found people rated content 48% less trustworthy once they suspected AI had written it. Those ratings fell even for pieces a human wrote, proof that suspicion alone does the damage.
Your job is to make AI content sound human to the reader with a credit card in hand.
Why every AI draft sounds the same
A language model builds a sentence one word at a time and keeps picking a likely next word.
Likely is another word for average.
Your draft about email marketing comes out as the paragraph a million people would nod along to. It’s polite and correct. It could also sit on any of your competitors’ websites.
I saw that paragraph on repeat in 2023, as president of an AI writing company that pushed out 50 million+ AI-written words a month.
That averaging also gives every model the same tics, with the long dash as the famous one. On November 13, 2025, Sam Altman announced what he called a “small-but-happy win.” ChatGPT would finally follow a no-em-dash rule set in custom instructions.
The fix for one punctuation mark took nearly 3 years to arrive after ChatGPT launched in November 2022.
Pew Research Center’s August 2026 analysis of roughly 490,000 English-language webpages found em dashes appear about twice as often as they did in 2023. Pew also found that negative parallelism, the sentence that knocks one idea down to set up the next, had nearly tripled, though it’s still fairly rare.
Dashes and that sentence shape are the famous tells. Your draft is probably carrying a few quieter ones.
9 AI tells to cut from every draft
Your reader starts to feel the machine once several tells stack up in one paragraph.
| Tell | What it looks like | What to do about it |
|---|---|---|
| 1. Em dashes everywhere | A long dash in every other sentence | Swap most of them for a period or a comma |
| 2. Negative parallelism | “It’s not just a tool, it’s a partner” | Delete the first half and say the second half plainly |
| 3. Groups of 3 | “Faster, smarter and stronger” | Keep the one item you can prove |
| 4. AI vocabulary | Delve, tapestry, testament, interplay, landscape | Use the word you’d say out loud |
| 5. Throat-clearing openers | “In today’s fast-paced world” | Open on a fact or a story |
| 6. Filler hedges | “It’s important to note that” | Cut the phrase and state the point |
| 7. Tacked-on -ing endings | “…highlighting the importance of trust” | End the sentence at the fact |
| 8. Same-size sentences | Every sentence runs about 20 words | Put a 4-word sentence next to a 25-word one |
| 9. No person on the page | No “I” and no opinion anywhere | Add a story or number only you have |
Go easy on the delete key before you strip out every dash. Human writers use em dashes too, a point Pew makes in its August 2026 analysis. A single dash in something you wrote yourself is fine with me.
How to humanize AI content in 7 steps
Steps 1 to 3 are the setup, done while the page is still blank. A draft that starts in your voice needs far less fixing later.
1. Teach the model your voice first
Open a project in Claude or ChatGPT and load it with your own writing. Mine holds my best 20 published pieces and a page of positions I’d defend in public.
A model learns more from those 20 pieces than from any prompt I could write.
Pick the pieces that got a reaction, such as the email people replied to or the post a client quoted back to you. Polished brochure copy teaches the model your brochure voice.
Mix the topics while you’re choosing. A model copies whatever its examples have in common, so 5 pieces about pricing will drag pricing into a draft about hiring.
Match the format as well. A project full of blog posts hands you emails that read like blog posts, which is why I keep one project per content type.
Then ask the model to list 10 patterns it sees in your samples, such as how you open and which words you repeat. Reading that list feels a bit like getting your own report card. You’ll also catch anything it has misread about you while it’s still cheap to fix.
Anything that has to sound like me gets drafted in Claude. I ranked the other options in my guide to the best AI for writing.
2. Write your standing instructions once
Standing instructions are the rules your model reads before every draft. ChatGPT calls them custom instructions. Mine set a sentence length range, an opening rule, a banned-word list and a line describing the reader.
My dash rule sits in there as well.
Write each rule as the thing you want and give the reason. Anthropic’s prompting guide says a model follows an instruction better once it knows why. A rule such as “Use a period where you’d reach for a dash, since my readers read dashes as AI” hands the model a reason to work with.
Ban the pattern along with the word. A lone ban on “game-changer” gets you “groundbreaking” in its place, which is whack-a-mole with a thesaurus.
Build the banned list from your own drafts. Scroll back over the last few things AI wrote for you and note every word that made you wince.
Start a fresh chat for every new piece too. The 2024 instruction drift study (Li et al.) found chatbots sliding away from their instructions within 8 rounds of conversation. The thread you’ve been revising in all afternoon has likely drifted from your rules.
You can copy this starter and paste your banned list underneath it.
Write the way I talk, with my samples as your guide. Open with a fact or a story. Mix short sentences with long ones and use plain words I’d say out loud. Leave out em dashes, groups of 3, “it’s not X, it’s Y” sentences and filler such as “it’s important to note,” because my readers spot all of these as AI. Use only the facts and stories I give you and ask for anything that’s missing.
3. Brief it like a writer you hired
A new freelancer with a 6-word brief would hand you back somebody else’s voice. A model does the same.
Tell the model who’s reading and the single point you want to land. Then give it your opinion on the topic, the one a competitor would argue with.
A model left alone gives you both sides of everything, which is how a draft ends up sounding like a committee wrote it.
Tell it what your reader already knows as well. Models love explaining the basics, so a line such as “they use ChatGPT every day, skip the definitions” removes the warm-up before it’s written.
For the best brief of all, flip it around. Ask the model to interview you with 5 questions before it writes a word. Your answers arrive in your own phrasing, which gives the draft raw material your competitors can’t get.
4. Add something only you have
Drop in a number from your own dashboard or a customer’s exact words. This is the step a tool can’t do for you.
Better still, put it in the brief, before the draft exists. A story the model has from the start shapes the whole piece.
Stuck for material? Your sent folder is full of it. So are your sales call transcripts.
Then run the competitor test on any sentence you’re unsure about. A line your closest competitor could publish word for word needs a number or a name in it.
Google calls this experience, the first E in E-E-A-T.
On July 15, 2026, my trained system wrote a Facebook post that pulled 336K views in about 2 days. Every specific in that post came from me.
5. Cut the 9 tells
Run the draft against the 9 tells. Search for dashes first, then read the last few words of every sentence, where the -ing endings hide.
Next, delete the last sentence of each paragraph and see whether you miss it. AI loves to close on a tiny moral. Most paragraphs read better without one.
Try the same with the whole first paragraph. An AI draft often starts for real in paragraph 2.
Check drafts in batches whenever you can. A tell that hides in a single piece jumps out across 10. A favorite sentence shape shows up in all 10 once you line up 10 hooks from the same model.
You can give a model the first pass. Pick a different one from the model that wrote the draft. A 2024 study (Panickssery et al.) found AI models score their own writing higher than anyone else’s, so a draft from Claude likely gets a tougher read from ChatGPT.
Paste the 9 tells in and ask it to flag each one. The rewriting stays with you.
6. Read it out loud
Read the whole piece out loud once. Yes, out loud, even if your only audience is a cup of coffee.
You’ll trip over every sentence you’d never say to a friend.
Mark the spots where you ran out of breath, since those sentences are too long. The spots where your own attention drifted are where your reader will leave.
Rewrite those the way you’d say them, then mix up the lengths while you’re in there.
For a tougher test, have your phone read the piece to you. A flat robot voice won’t act the lines out the way your own voice does, so a dull sentence has nowhere to hide.
7. Check the numbers, then give it a human read
AI models state wrong numbers in a very confident voice.
Google’s guidance on generative AI content tells site owners to manually fact-check and review anything AI wrote before it goes live. On my drafts, a figure with no named, dated source behind it gets cut.
Click every link the model gave you. A model can cite a real-sounding report at a URL that goes nowhere, or one that says something else entirely.
Check the people as well. In 2026, researchers in Warsaw found AI models inventing the same experts again and again. One of them, an “Elena Vasquez” who has never existed, is listed as an author on 77 papers, each claiming a journal that doesn’t exist either.
Claude is the model that kept reaching for that name. It’s also the model I draft in, so this check stays in my routine.
Search your draft for any expert you never interviewed.
A quick shortcut for all of this is to ask the model for a list of every factual claim in the draft, with the source it used for each. The claims with a blank beside them are your to-do list.
Every video and every post gets a human read before it publishes. Reading what my AI staff produce takes about 6 hours of my day.
Borrowed wording is a separate risk for that read to catch, which I cover in my post on AI content plagiarism.
A before and after with my own channel numbers
A one-line prompt packs 7 of the 9 tells into 2 sentences.
| Before, from a one-line prompt | After, humanized |
|---|---|
| “In today’s fast-paced digital landscape, training AI on your brand voice isn’t just a nice-to-have, it’s a game-changer. Moreover, businesses that leverage voice training can unlock consistency, credibility, and connection, ensuring their content truly resonates.” | “From March to late May 2026, my YouTube channel ran on AI scripts with none of my own material behind them. Monthly organic views slid from about 2 million toward 1 million. The training data landed in May 2026. A video published without it pulled 3.2K views in 3 weeks, while one published with it in the same week pulled 25.2K views in 3 days.” |
The one-line-prompt paragraph has negative parallelism, a group of 3, AI vocabulary, a throat-clearing opener, a tacked-on -ing ending, same-size sentences and no person anywhere on the page.
The humanized version has dates and numbers from someone who was there. You could argue with it, which is a good sign a human is talking.
About 2 months of videos from the untrained stretch, which I ran on purpose, are still up on the channel.
AI humanizer tools change words and leave the voice empty
Let me ask you this. Would you let a stranger rewrite your sales page with a thesaurus and publish it unread?
That’s close to the job a humanizer tool does. You paste in a draft. A tool such as QuillBot or Humbot rewrites the sentences for you.
In SlashGear’s January 2025 test of 10 humanizers, QuillBot’s rewrite scored 100% AI on QuillBot’s own detector. Humbot turned “Tap your Apple ID” into “Faucet your Apple ID.”
The SlashGear reporter’s own advice was to rewrite the copy yourself.
On August 27, 2025, Turnitin added detection for humanizer-altered text. You’d be paying for a rewrite that can break your meaning while the detector companies keep closing the gap.
I’d skip the score chasing altogether.
In a test I published in 2025, I ran 10 articles, a mix of human and AI writing, through AI detectors. All 10 came back flagged as AI. My guide to AI detection explains why that happens.
Keep the human part real
You can also overcorrect, usually when you try too hard to sound casual.
| Overcorrection | Why it backfires | Do this |
|---|---|---|
| Planted typos | A typo reads as careless to a customer and does nothing for the voice | Fix the errors and vary your sentence lengths |
| Invented stories | A made-up client or a fake “last week” moment breaks trust the day someone asks about it | Use a real number or a real customer line, even a small one |
| Slang you’d never say | Readers who know you can hear the costume | Write in the words you use on a call |
| Jokes on a timer | A quip in every paragraph sounds as scripted as the draft you started with | Keep the one line that made you laugh |
If a customer asked you about any line on the page tomorrow, you’d want to be able to answer.
Google ranks helpful pages, however they were drafted
The other judge you might be worried about is Google. Google’s spam policies settle the AI question in 5 words. Scaled content abuse covers mass-produced content of little value “no matter how it’s created.”
Google grades the page on the value it gives a reader, whoever or whatever typed the first draft.
Put 2 pages next to each other on the same question. One carries a real person’s test results. The other reads like every page you’ve already closed.
I mean, wouldn’t you pick the page with a person on it?
In July 2025, Ahrefs checked 600,000 pages from the top 20 search results. It found 81.9% mixed AI with human writing, while pure AI pages made up 4.6%, fewer than 1 page in 20.
Graphite’s October 2025 study found the same lean in answer engines, where 82% of the articles ChatGPT and Perplexity cited were human-written. Both studies rely on AI detectors, so treat the exact figures as estimates.
My bet for answer engines is a page that carries your voice and your own data. My guides to generative engine optimization and answer engine optimization cover the full method for getting cited.
How my team gets a first draft that sounds like me
My early Claude projects had perfect grammar with none of my voice in them.
The fix was a document we call Voice DNA, a build I describe in Claude vs ChatGPT. It runs 15+ pages and came out of a recorded interview about what I believe and how I phrase it. One section lists words I’d never say.
A model that reads it stops reaching for them.
You can borrow the whole routine. Every piece I publish moves through 4 stages.
| Stage | Who or what | The job |
|---|---|---|
| 1. Research and outline | ChatGPT | Handles research passes and builds the outline |
| 2. Draft | The Claude project for that content type, holding Voice DNA | Every draft starts here |
| 3. Grammar check | Grammarly | Reads the last version ahead of the human read |
| 4. Human read | A trained human editor | Checks the draft against Voice DNA and signs it off |
The other tools I recommend are in my AI marketing tools list.
Voice DNA is layer 3 of the 6 layers in every First Movers AI clone, with trained human editors inside layer 6.
Your voice is the rare thing on the page now
Average writing is free in 2026 now that any chatbot will draft it on request. Pew’s August 2026 analysis found signs of AI authorship on over 1 in 3 pages published after ChatGPT’s release.
A voice people recognize is what’s left to compete on, which is good news since you already own one.
You’ve got strong opinions and stories from your own customers. Humanizing is the part of the work that was yours from the start.
FAQ on how to humanize AI content
What does it mean to humanize AI content?
Humanizing AI content means editing an AI-written draft until it carries a real person’s voice and firsthand experience. The work includes teaching the model your writing style, adding stories or data only you have, cutting common AI tells such as em dashes and throat-clearing openers, then giving the piece a final human read.
Does Google penalize AI content?
No. Google’s spam policy on scaled content abuse targets mass-produced pages of little value, however they’re made. Google’s guidance on generative AI content adds that the technology can be useful for research and structure. Ahrefs’ July 2025 study of 600,000 top-ranking pages concluded that Google neither rewards nor penalizes pages for using AI.
Do AI humanizer tools work?
Rarely. AI humanizer tools rewrite the wording of a draft, though the result can still score as AI and can break your meaning along the way. In SlashGear’s January 2025 test of 10 humanizers, QuillBot’s rewrite scored 100% AI on QuillBot’s own detector. Turnitin added detection for humanizer-altered text on August 27, 2025.
How do I make AI content sound human in ChatGPT?
AI content from ChatGPT sounds human once the model has your own writing to copy. Create a project and upload 20 of your best pieces. Then add custom instructions that set a sentence length range and ban em dashes. Sam Altman said on November 13, 2025, that ChatGPT now follows a no-em-dash rule placed in custom instructions.
Are em dashes a sign of AI writing?
Em dashes are a weak sign of AI writing on their own. Pew Research Center’s August 2026 analysis found em dashes appear about twice as often on the web as they did in 2023, though Pew notes that human writers use them too. A draft that stacks dashes with other tells, such as throat-clearing openers, looks far more machine-made.
Should I tell readers I used AI?
Telling readers you used AI is a choice Google encourages wherever a reader would reasonably expect it. Google’s guidance on creating helpful content says AI or automation disclosures are useful where someone might ask how a piece was created. At First Movers AI, a line on screen tells viewers when they’re watching Julia McCoy’s AI clone.
Build your voice setup live with my team
Say it’s Tuesday, 7 PM ET. You’re on a live build with my team while your best 20 pieces sit in a Claude project. The first draft that comes back sounds like you on a good day.
That’s a regular Tuesday inside AI Labs, where we’ve trained 3,000+ students. The newest flagship course, Build Your 2026 Claude Copywriter Setup, gives you the training files and system prompts for steps 1 and 2.
Join AI Labs and build your voice setup once. That setup is how you humanize AI content for every draft that follows.
Your next draft can come back in your voice on the first try.
Go get those 5 hours back.
