Search for a prompt to humanize AI text and you will find lists of fifty, each promising output that reads as though a person wrote it. Most of them are variations on one instruction, and that instruction is the one that changes least.
Understanding why is more useful than another list, because it tells you which prompts are worth typing.
Why "write like a human" barely moves anything
A language model produces text by repeatedly choosing a next word from a probability distribution. What makes machine output recognisable is that those choices sit near the top of the distribution, consistently, across a whole document. That is what low perplexity means, and it is a property of the sampling procedure rather than of the persona the model was asked to adopt.
Tell it to write like a human and it adopts a more casual register. It still picks predictable words within that register. The tone changed and the statistics did not, which is why people report a prompt "working" when they read the output and then find the score barely moved.
There is a related trap worth naming. Ask a model to avoid a word and it will usually reach for the nearest synonym, so "delve" becomes "explore" across every instance and the repetition simply relocates. The instruction has to say restructure, not substitute, or you get the same shape with fresh vocabulary.
Ask for the edit, not the impression
The prompts that do something name a constraint you could verify yourself. Compare "make this sound natural" with "rewrite so sentence lengths vary between 6 and 30 words, with at least two sentences under 10 words per paragraph". The second is checkable, so the model can actually satisfy it.
Five that earn their place:
1. Constrain the rhythm
"Rewrite this so sentence length varies substantially. Include at least two sentences under ten words in every paragraph, and do not let three consecutive sentences have similar length."
2. Ban the vocabulary
"Rewrite without using: delve, leverage, comprehensive, robust, nuanced, furthermore, moreover, it is important to note, in today's. Do not substitute near-synonyms for these, restructure the sentence instead."
3. Kill the shape
"Remove the opening paragraph that restates the question and the closing paragraph that summarises. Begin with the actual claim."
4. Force commitment
"Remove every hedge. If a claim is qualified, either state the qualification concretely or delete it. No 'generally', 'often', 'can be seen as'."
5. Cap the paragraphs
"No paragraph over 80 words. Where an argument turns, use a two-sentence paragraph."
Order matters more than most lists admit. Run the structural instructions first, since prompts 1, 3 and 5 change where sentences begin and end, and then apply the vocabulary pass. Reverse that and the restructuring reintroduces the phrases you just banned.
The three models need different instructions
This is where generic lists fall down. ChatGPT over-hedges and over-formalises, so prompts 2 and 4 do the most work. Claude produces well-balanced paragraphs of uniform weight, so prompts 1 and 5 matter more and prompt 4 barely applies. Gemini fragments into lists and headings, so the useful instruction is the opposite of brevity: ask it to connect its bullet points into argued prose and supply the reasoning it left implicit.
Applying ChatGPT advice to Gemini output makes it worse, which is the single most common mistake in this area.
What a prompt to humanize AI text cannot do
Two things. The first is that generation-time instructions only shape what the model produces next; they do nothing about evenness that survives into the draft, which is why most people still end up editing afterwards. Restructuring existing sentences is a different operation from generating better ones, and it is what an AI to human rewriter does to text you already have.
Measure rather than assume, either way. Read a prompted draft aloud and count how many sentences run to a similar length, because that is the property being scored and it is one you can check without any tool. If four consecutive sentences land within a few words of each other, the prompt did not do what you asked, whatever the prose feels like.
The second is more important. No prompt makes an argument yours. If the reasoning came from the model, the prose reading naturally changes nothing about authorship, and that is the question an academic integrity process actually asks. Universities are consistent on this even where they permit AI assistance.
Worth knowing what a realistic result looks like. A well-prompted draft usually reads better to a person and moves a detector score by less than people expect, because the two are measuring different things. If a prompt takes a paragraph from 95% to 90%, that is roughly what prompting alone achieves; the larger movement comes from restructuring afterwards. Anyone promising a specific score from a prompt is selling something.
For the editing side of the job, the manual restructuring method covers what to change by hand, and the phrase list feeds directly into prompt 2 above.