The mechanism
A green-list mark lives in a large number of small choices between near-equivalent words. The statistic is an excess of “green” choices across the whole document, so the signal is spread thin and redundant rather than concentrated anywhere removable.
Editing replaces some of those choices with your own. Each replaced word is one fewer scored position, so the excess shrinks. Light copy-editing typically leaves a strong signal intact; heavy rewriting degrades it; translation into another language effectively destroys it, since the word choices are entirely different ones.
What OpenAI measured
OpenAI published one of the first vendor figures for this. Alongside its EU text watermark, it reported that replacing 10% of words with synonyms reduced detection from about 92% to 66%, and that short passages, math answers and translated text are harder to detect.
That is a statement about one vendor's method under one test. It is consistent with the mechanism above, and it is why this product reports the share of your original words that changed, so a reader can place an edit on the same axis.
Why this matters for reading a result
This is the technical reason an absent mark cannot be treated as proof of human authorship, and it is worth stating in an appeal. Substantially edited generated text can test clean. So can text from a system that never applied a mark. So can text a person wrote.
Those three histories are indistinguishable to the test, which is exactly why WatermarkRemoverPro reports what it measured rather than issuing a verdict.
What editing can and cannot promise
WatermarkRemoverPro does now offer an on-device rewrite feature (see /rewrite), and it is bound by the same physics this page describes: it can reduce detectable evidence, and it cannot fabricate certainty about a watermark it has no key for. No tool, including this one, can honestly guarantee defeating a specific vendor's undisclosed watermark.
If you edited generated text and want to know whether a signal remains, that is a diagnostic question and this tool answers it, directly and for free.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. WatermarkRemoverPro's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.