How to Humanize AI Text in ChatGPT and Claude: A Free, Research-Based Tool

Eli Josiah

10 min read

·

Sep 22, 2026

--

Listen

Share

Press enter or click to view image in full size

If you want to humanize AI text from ChatGPT or Claude without rewriting every line yourself, this guide is for you. In this article, I will give you my free Claude skill and custom GPT, built on 11 research sources, adapted to research, policy, speeches, articles, grants and fiction.

A note before you read: this article was written with Claude, using the tool I will give you. The research questions, the design decisions and the final review are mine

Let me start with something we can all relate to. You go to ChatGPT or Claude and ask it for a draft. It could be an assignment. It could even be work. You desperately don’t want anybody to think ChatGPT wrote it, whether or not you did your own research.

Then the draft comes back and suddenly you see the familiar A.I markers everywhere; “let’s delve,” “a pivotal moment” here, a “rich tapestry” there, a sentence that ends with “highlighting the importance of”, and a last paragraph assuring you the future is bright. Every paragraph is the same length.

It can be frustrating. And before you know it, you have spent more time editing than it would have cost you to write the whole thing from scratch.

I built a free tool that will help you solve this problem, that forces GPT and Claude to write like human, without all of these familiar AI markers. It is a Claude skill, with a custom GPT version for ChatGPT, and it is now in its second version. This time it is built on 11 research sources, and it changes its approach depending on what you are writing, because a research paper, a speech and a short story should never sound the same.

Why does ChatGPT and Claude writing sound like AI?

Here is the thing. AI models do not write badly. They write clearly. The problem is that they write everything the same way.

A 2025 study in PNAS found that these models keep a heavy, noun-packed style even when you ask them to write casually. The researchers called it genre misalignment. In plain English, your speech, your report and your blog post all come out in one voice. Readers feel it, even when they can’t put their finger on it.

Why “make it sound human” doesn’t work

I am sure you have tried this prompt before. You add “make it sound human” to the end of your prompt, and what comes back? The same gibberish, just wearing different adjectives.

That happens because there is no single human style. Nobody writes a grant proposal the way they write an op-ed, and neither sounds like a short story.

I learned this the hard way. Three weeks ago I released the first version of this skill. I built it using the findings from Wikipedia’s field guide on the signs of AI writing and turned it into rules: drop “delve” and “tapestry”, write “is” where the model wants “serves as”, that kind of thing.

People used it, and the feedback kept pointing to one big flaw: it treated every kind of writing the same. So I went back to the drawing board…

What the research says about AI writing patterns

This time I did not lean on one source. Instead, I consulted 11 established research on this subject; but two did most of the heavy lifting.

The first was Wikipedia’s “Signs of AI writing” page, a detailed catalogue of the words, phrases and formatting habits that give AI away. It is a good start, and that’s where many so-called AI-Claude Skills stop. But you should know what it is. It describes itself as an advice page the community has not reviewed. It was written to help Wikipedia editors catch AI edits, not to help writers like you and me.

So, I had to do some digging, and then I found a second, and this time, far more comprehensive paper that went beyond words and focused on narrative of styles between human and AI text. It is a 2026 study called StoryScope, by Jenna Russell and colleagues at the University of Maryland and Google DeepMind. They took 10,272 short stories written by people and asked five AI models (Claude, GPT, Gemini, DeepSeek and Kimi) to write their own version of each. Then they ignored the style completely and looked only at how each story was put together. Structure alone told human and AI stories apart with a macro-F1 score of 93.2%.

Now this is the part that stopped me in my tracks. When the researchers edited the AI stories to strip out the clichés and flowery language, detection barely moved. It went from 95.5% to 93.9%. So swapping out the obvious words does not really fix your draft. The AI shape underneath is still sitting there.

What is that shape? AI narrators spelled out the story’s theme in 77% of stories, while human writers did it in 52%.

AI showed emotion through the body, the tight chest and the cold sweat, 81% of the time, against 38% for people. And human writers named real books, places and people almost twice as often as the models did.

To be fair, StoryScope is still a preprint and it only looked at fiction, so I use it directly for stories and more carefully for everything else.

After that I read nine more studies, and together they cover the different problems with AI writing, from the way it builds sentences to what readers notice and what detectors get wrong.

One study found that models use the heavy “highlighting its importance” pattern two to five times more than people do, and pile up noun-heavy phrases up to twice as often. Others found that AI writing lacks a person behind it: students wrote “I think” and “in my view” far more than ChatGPT did, and ChatGPT asked fewer questions and made fewer personal asides.

Two studies found that AI flattens rhythm, dropping the very short and very long sentences we use without thinking, and that it lets in less fear and disgust and more joy than people do. Two more, both working with expert editors and professional writers, traced the padding: irrelevant content, thin information, clichés, over-explaining and no real detail.

And yes, readers notice. When five heavy ChatGPT users voted together, they got only one article wrong out of 300, and many of them pointed to a lack of originality, which tells you readers are also quietly asking, “Haven’t I read this somewhere before?”

How the free AI humanizer works for different types of writing

So what did I do with all of this? How does this new model I built differ from the first?

Well, this new model works in two layers. The first layer is a set of rules that apply to every single draft. The writing should fit what it is. The rhythm should move, with short sentences next to long ones. Someone should actually take a position: firmly when the evidence is strong, and honestly unsure when it is not. Every sentence should earn its place. And ideas should connect with real linking words, not the same “Additionally” every other line.

This first layer still carries the full Wikipedia checklist, plus some integrity rules: name your sources, give page numbers, and just say so when you don’t know something.

The second layer is where the real fix happens. There is a separate profile for each of six kinds of writing: research, policy briefs, speeches and talking points, articles and essays, grant proposals, and fiction. Ask it for a policy brief, and it puts the recommendation up front and says exactly who should act on it. Ask it for talking points, and it talks straight to the audience, repetition and all. If your request is vague, like “write an article on AI in schools”, it will stop and ask you one question first: research article or opinion piece? And if you upload a funder’s template, the template wins.

There are a few other things you will notice.

It spells in British English, the way we write in Nigeria, unless you tell it otherwise. Most models default to American spelling, and that mismatch is itself a giveaway.

It knows its own model’s bad habits. The Claude version watches for Claude’s love of em dashes and soft summary endings. The ChatGPT version watches for ChatGPT’s leftover citation code and favourite words.

It comes with a small scanner. The scanner checks your draft’s sentence variety, noun-heavy phrasing and hedging, how often you speak to the reader, and whether spellings are mixed. I set those thresholds myself, because the studies give comparisons, not fixed targets. So when it flags something, take it as a nudge, not a verdict.

The studies didn’t always agree, so I had to make some calls. Wikipedia treats “Additionally” as an AI tell, but Herbold found that people actually use more linking words overall. So the skill keeps your transitions and just asks you to vary them. And while most style guides tell you to avoid the passive voice, Reinhart’s numbers suggest a normal amount of it reads more human than none. So the skill lets it stay wherever your kind of writing uses it.

Does it work? I tested it on this article

Fair question. So I ran it on the article you are reading.

On the first draft, the word check found nothing to fix, because every flagged word was one I was quoting as an example. The real problems were deeper. The draft never asked you a single question. It barely hedged, about once every 300 words. Every section was roughly the same size. And it ended on a neat three-part line summing everything up, which is one of the exact Claude habits the skill is built to catch. The rewrite fixed all of that.

But let me be straight with you. That is one article. No one has run a proper blind test yet, where readers compare drafts with and without the skill. So the honest claim is that it removes the documented AI patterns. I am not promising it will make a draft sound exactly like you.

Can it bypass AI detectors?

No. And I didn’t build it to. In the StoryScope study, classifiers reading the raw text still scored above 99.7%.

There is also something about detectors you should know. Weixin Liang and colleagues (Patterns, 2023) found that detectors flagged more than half of a set of TOEFL essays written by non-native English speakers as AI-generated. And when the words in those essays were made fancier, the false alarms dropped.

Let that sink in for a moment. Detectors reward exactly the kind of inflated language this skill removes. If you write in Nigerian, Indian or Kenyan English, you have every reason to take those scores with a pinch of salt.

There are other limits too. The skill can’t fully judge its own work. When Shaib’s team asked AI models to spot the problems the expert editors found, they barely agreed with the experts at all. So your final read still matters most.

It will also need updates, because newer models pick up new habits, and I plan to review it every three months.

Two of the profiles, speeches and grant proposals, lean partly on standard writing practice, because nobody has studied AI against human writing in those forms yet. I have labelled those rules clearly.

And it does not change your responsibilities. If your school, client, funder or publisher wants you to disclose AI help, please do.

How to install the humanize skill in Claude

1. Download elijoskie-humanize-fieldguide.skill from the GitHub repository.
2. In Claude, click Customize, open Skills and choose Upload skill. Select the file.
3. If you installed version 1, delete it or switch it off so the two don’t clash.
4. Make sure code execution is on in your settings, because the scanner needs it.
5. Type / in the message box and pick the skill, or start with "use elijoskie-humanize-fieldguide". Then tell it what you are writing, for example "draft a policy brief".

How to set up the custom GPT in ChatGPT

1. Download elijoskie-humanize-fieldguide-GPT.zip from the repository and unzip it.
2. In ChatGPT, open Explore GPTs, click Create and switch to the Configure tab.
3. Paste everything in INSTRUCTIONS.txt into the Instructions box.
4. Under Knowledge, upload the 13 files in the knowledge folder, not the zip itself.
5. Under Capabilities, turn on Code Interpreter & Data Analysis and Web Search. You won’t need image generation.
6. Save it as private while you test it, and use SETUP-GUIDE.md for the rest.

Frequently asked questions

• Is this AI humanizer free? Yes, completely. The Claude skill and the custom GPT files are free and open source on GitHub.
• Does it work with ChatGPT? Yes. The repository includes a package for setting it up as a custom GPT, tuned to ChatGPT’s own habits.
• Will it get my work past Turnitin, GPTZero or other AI detectors? That isn’t what it’s for, and the StoryScope results suggest no editing tool can honestly promise that. It is built to remove the patterns that make readers stop trusting your writing.
• What kinds of writing does it handle? Research papers, policy briefs, speeches and talking points, articles and essays, grant proposals, and fiction.
• Do I need to know how to code? Not at all. In Claude it’s a file upload, and in ChatGPT it’s copy and paste.

Try it on your next draft

So here is my ask. Take one paragraph ChatGPT or Claude wrote for you this week. Run it through the skill, tell it what you are writing, and put the two versions side by side. Then ask yourself which one sounds like a person actually sat down and wrote it.

If a rule gets it wrong for your kind of writing, tell me by opening an issue on GitHub. Version 1 got rebuilt because people told me exactly where it failed them. Version 3 will be built the same way.