Prompts to Humanize AI Text: A 60-Text Test
We asked an AI model to rewrite 60 AI-written texts with the usual one-line humanize prompt and scored the results on ZeroGPT. The writing stayed accurate, but the AI score barely moved. Here is the test, why it happens, and prompts that give the model more to work with.
Short answer: a one-line prompt such as "rewrite this so it sounds human" gives you cleaner text, but it does little to change how an AI detector reads it. In our test the average ZeroGPT AI score fell only from 89% to 77%. What changes the result is giving the model, or yourself, something the draft does not already contain.
The test
In October 2026 we took 60 AI-written texts of about 250 words each: 30 written by Claude and 30 by DeepSeek, across academic essays, research abstracts, work emails, business reports, blog posts and personal statements. Before any rewrite, ZeroGPT scored them 89% AI on average.
We then gave each text to DeepSeek with the prompt most people try first:
Rewrite this so it sounds human-written. Return only the rewritten text.
Every rewrite was scored by ZeroGPT through the same detection API our AI Detector uses, and read by blind reviewers who checked whether it stayed faithful to the original and free of errors.
What happened
| Result | One-line prompt rewrite |
|---|---|
| Average ZeroGPT AI score | 77% (median 83%), against 89% before |
| Texts scored under 50% AI | 7 of 60 |
| Texts scored under 20% AI | 3 of 60 |
| Rewrites with no serious fault | 57 of 60 |
The prompt did what it promised in one sense: the rewrites were fluent and faithful, with almost no grammar errors or changed facts. It did not do what most people use it for. Most rewrites still read as AI to the detector.
This is one test, on short English texts, with one model and one detector. Other prompts, models and detectors will give other numbers.
Why a one-line prompt changes so little
Asked to "sound human", a model rewrites the text in its own default style: the same even pacing, the same tidy structure, the same kind of transitions. The words change; the habits that detectors and readers pick up on mostly stay. A model has no way to add what a human draft usually contains: your examples, your numbers, your opinion and the details you noticed yourself.
Prompts that give the model more to work with
These prompts are starting points, not a guarantee. We have not measured their detector scores; what they do is make the model work from your material instead of its defaults.
Add your own material first.
Rewrite the draft below for [audience]. Keep every fact, name and number exactly as written. Work in these points from my own notes: [two or three specific details, an example or an opinion]. Return only the rewritten text.
Ask for specific changes, not a vague tone.
Edit this draft. Vary sentence length, cut stock transitions such as "moreover" and "in conclusion", replace general claims with the specific ones already in the text, and do not add facts. Return only the edited text.
Edit in passes rather than all at once.
Go through this draft paragraph by paragraph. For each paragraph, tell me which sentence sounds most generic and suggest one concrete replacement. Do not rewrite the rest.
The last prompt keeps you in charge of the text, which is what most course and workplace policies expect anyway.
What works better than any prompt
- Write the parts only you can write. A specific example, a number from your own work, a reason you disagree.
- Edit by hand after any rewrite. Read it aloud; anything you would not say is worth changing.
- Check before you submit. Run the final version through an AI detector and read the result as a signal, not a verdict.
- Compare any automated rewrite with your draft. Whether it comes from a prompt or a dedicated humanizer, ours included, check that names, numbers and claims survived.
FAQ
Is there a prompt that makes AI text undetectable? No prompt can guarantee a detector result. In our test the common one-line prompt left 57 of 60 texts above a 20% AI score on ZeroGPT.
Can ChatGPT or Claude humanize their own text? They can rewrite it fluently. Whether the result reads as human depends far more on what you add than on how you phrase the request.
What is the best prompt for an essay? One that includes your own argument, sources and examples, and asks the model to keep facts exactly as written. Then follow your course's rules on AI use and disclosure.
Does asking it to "write like a human" hurt quality? Not in our test: 57 of 60 rewrites had no serious fault. The problem was the detector score, not the writing.