What "humanizing" text actually changes
A humanizer rewrites word choice, sentence rhythm and punctuation habits (the em dash used as a clause connector, stock phrases like "delve into" or "moreover") that occur more often in generated text than in ordinary prose, and that also happen to sit inside a statistical watermark's scored positions. Changing enough of them measurably reduces the evidence a detector or a watermark test finds.
That is a real, falsifiable effect: you can measure a before-and-after delta with the same arithmetic a detector uses. What it is not is a certainty. A vendor's own watermark key is not published anywhere, so no outside tool can confirm it has fully cleared a specific detector's specific test.
Why the "always passes" promise does not hold up
A keyed statistical mark is designed so that only the party holding the key can reliably test for it. That is the entire point of a keyed test, in the same way a locked door does not care how confidently a locksmith's advert reads. Anyone advertising a permanent, universal escape from every detector is describing something outside what the underlying method can support, not a feature they have built.
The honest version of this claim is a reduction, not an erasure: fewer of the statistical patterns a detector keys on, measured against the same detector arithmetic doing the checking. That is what "reduce detectable evidence" means on this site, and it is the strongest claim the underlying method supports.
What WatermarkRemoverPro's rewrite actually does
Two mechanisms, run together, entirely on your device: a deterministic pass over AI-tell punctuation and phrasing, and targeted candidate rewrites of the specific passages a real per-passage check flags, gated so a candidate that changes a number, a name or a negation is rejected outright rather than used. See /rewrite.
Checking is unlimited on every plan, because the computation runs on your device rather than metering server compute. Rewriting on the free plan has a weekly token budget on the Standard engine; a Pro subscription removes it and adds the Pro engine, a real local language model with more candidates per passage and the extended AI-tell library. Nothing about the document is ever sent anywhere, on any tier: see docs/REWRITE_PHILOSOPHY.md for exactly what is and is not claimed.
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.