Three things that won’t work
Clearing formatting, running the text through a character cleaner and retyping it yourself all keep the same words. The mark is in the words, so it stays.
How much needs to change
OpenAI’s test is the best guide we have. Swapping 10% of the words left about two-thirds of watermarked passages still detectable. Swapping 25% left only about one in six. So about a quarter of your words is the number to aim for, a bit more on long pieces.
How to do it in NeverPrompted
Paste your text, open Advanced, and under Word-change target choose 25% (textGrain). NeverPrompted keeps rewriting until about that share of your words has changed, and tells you what it reached.
It won’t trade your meaning for the number. Names, figures and dates are locked, and a rewrite that drifts from what you said is dropped. If it can’t reach the target safely, it says so and shows how far it got.
When the number can’t be reached
Some writing leaves little room to move. A recipe with exact quantities, a legal clause, a paragraph of code comments or a list of names has words that can’t change without changing the meaning. In those cases NeverPrompted stops short of the target and shows the percentage it actually reached, rather than forcing replacements that would make the text wrong.
Short pieces behave differently too. OpenAI says shorter text is harder to detect to begin with, so a two-sentence answer carries a weaker mark than a long essay. For long pieces, choose a target a little above 25% so you have a margin, then read the result through once, start to finish, before you use it.
Why we’re upfront about the limits
Plenty of sites will promise a clean result. We can’t, because the only test that counts belongs to OpenAI and isn’t available to us or to you. What we can do is give you a number you can check, the share of your words that changed, and the same published figures OpenAI used so you can see where that number sits.
Then make it sound like you
A watermark is only one signal. Read the result aloud, put back the phrases you’d really use, and add something only you know. That’s what makes it yours.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. NeverPrompted's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.