ChatGPT humanizer prompt: three prompts and what they change

ChatGPT rewrites the text you paste according to the instruction you give it. A vague instruction ("make this sound human") leaves the model to decide what human means, and its default is the most likely wording for the widest audience, which is the pattern detectors learn. A specific instruction names the edits: shorter sentences of uneven length, no stacked qualifiers, no sentence that announces importance instead of adding a fact, no new facts, and a sample of the writer's own prose to match.

A custom GPT does not change this. OpenAI introduced GPTs on 6 November 2023 as custom versions of ChatGPT that combine instructions, extra knowledge and a chosen set of tools, built without code (OpenAI, Introducing GPTs, read 9 November 2026). A humanizer GPT in the store is a saved prompt with a name and, sometimes, a document of rules. Its results vary run to run for the same reason a pasted prompt's do.

This prompt names the tells and forbids invention. Paste it, then paste your text.

Rewrite the text below so it reads as if one person wrote it for one reader.

Rules:
1. Keep every fact, number, name, date and claim exactly as given. Add nothing. If a sentence needs a detail that is missing, leave a bracketed note instead of inventing one.
2. Vary sentence length. Some sentences under eight words, some over twenty, no fixed rhythm.
3. Remove sentences that announce importance without adding a fact. Remove "not X but Y" framing and state the point.
4. Remove stacked qualifiers (could potentially, may possibly). Keep one qualifier only where the source supports it.
5. Use plain verbs: is, has, does. No "serves as", "boasts", "features".
6. No groups of three for their own sake. Use the number of items the meaning needs.
7. No closing line that restates the paragraph. End on the last concrete fact.
8. Keep the same order of ideas unless a sentence is clearly out of place.

Return only the rewritten text, then a short list of what you changed and why.

Text:
[paste here]

The last line matters. Asking for a change list makes it possible to check the rewrite against the input, which is the only way to see whether rule 1 held.

This prompt uses a sample of your own writing, two or three paragraphs, so the model has a target other than its default.

Below are two samples. SAMPLE A is my own writing. SAMPLE B is a draft I want rewritten in my voice.

Study SAMPLE A for: average sentence length and how much it varies, how I open sentences, which connectors I use and which I never use, how I punctuate (commas, colons, parentheses, dashes or none), how formal I am, and any habits that repeat.

Then rewrite SAMPLE B to match those habits. Keep every fact, figure and claim in SAMPLE B exactly as it is. Do not borrow any fact or phrase from SAMPLE A. Do not add examples, opinions or transitions that are not in SAMPLE B.

Return the rewrite only, then three sentences naming which of my habits you applied.

SAMPLE A (my writing):
[paste 2 to 3 paragraphs you wrote without AI]

SAMPLE B (draft to rewrite):
[paste the AI draft]

A voice sample is the one input a hosted humanizer rarely takes. It is also the easiest way to lose facts: a model matching a style will sometimes carry a phrase across from the sample, which is why the prompt forbids it twice.

This prompt runs a rewrite, then a self-review against the original before the final version. It trades time for fewer dropped claims.

Work in three steps and show each one.

Step 1. Rewrite the text below so it reads as written by one person for one reader: uneven sentence lengths, plain verbs, no inflated significance, no sentence that only announces what comes next, no restating closer. Keep every fact, number, name, date and claim as given. Add nothing.

Step 2. Compare your rewrite with the original, claim by claim. List every claim in the original and mark it KEPT, CHANGED or MISSING in the rewrite. List any claim in the rewrite that is not in the original.

Step 3. Produce the final rewrite with every CHANGED or MISSING claim restored and every added claim removed.

Text:
[paste here]

Step 2 is the part most prompts skip. It makes the model do the comparison a careful editor would do, and it surfaces the dropped qualifier or swapped term before you use the text.

Three published sources answer this, and each answers a narrower question than the headline suggests.

Liang and colleagues found that several GPT detectors flagged about 61 percent of TOEFL essays by non-native English writers as AI, and that simple prompting strategies both reduced that bias and evaded the detectors (Liang et al. 2023, Patterns00130-7), read 9 November 2026). That is the direct evidence that a prompt changes a reading. Krishna and colleagues showed the same for a dedicated paraphraser, which dropped DetectGPT's detection rate from 70.3 percent to 4.6 percent at a 1 percent false positive rate (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026).

Both studies are from 2023, and detector makers have had three years to respond. GPTZero's technology page describes a Paraphraser Shield against paraphrasing and homoglyph attacks and states that its confidence categories are tuned so the average error rate is under 1 percent for high-confidence predictions on an internal evaluation set (GPTZero, how it works, read 9 November 2026). A prompt that moved a reading in 2023 may not move the same detector today, and no study this site could find has tested a named ChatGPT prompt against a 2026 detector.

Three things, and none of them is a detector setting.

• Meaning. A rewrite that reorders and reworks sentences can drop a qualifier, widen a claim or swap a technical term for a near-synonym. Krishna and colleagues judged their paraphraser's semantics "not appreciably modified" by automatic measures; no study has had editors rate prompted rewrites claim by claim. The published method on this site does exactly that, with two editors reading blind.
• Repeatability. ChatGPT samples its output, so the same prompt on the same text gives a different rewrite each run. A detector reading from one run is evidence about that run.
• The detector's date. Detector models change without notice. A reading is true for its detector on its date, and a prompt has no version number to cite.

No detector score should be the only evidence in a misconduct case. If you are a student, check your institution's AI policy before using any rewrite prompt; a humanizer, pasted or hosted, does not make undisclosed AI use acceptable.

A pasted prompt is free, open to inspection and changes nothing about where your text goes beyond the chat you are already in. Its weakness is drift: each run is different and the rules live in your clipboard. An open-source skill fixes the rules in a file that loads on demand; the most installed one is described in the Claude humanizer skill guide. A hosted humanizer adds request caps, strength modes and a monthly word pool, and charges for them; the manual edits those tools automate are set out in how to humanize AI text.

The order to try them is the order of cost. Start with Prompt 3 on a text you know well and read the Step 2 table. If claims go missing on a 300-word paragraph, a hosted tool running the same operation at scale will lose them too, and no detector reading changes that. The rest of this series is in the guides hub.