Whether AI humanizers work is four different questions, and they have different answers. Lumping them together is why the discussion on this topic is so unsatisfying: someone says yes, someone says no, and both are right about different questions.
Here they are separately. Two are yes, one is no, and one cannot be answered by us or by anyone else selling you one of these tools. We sell one, which you should weigh when reading this.
The least interesting question and the only one with a clean answer.
Yes, reliably, because the properties a humanizer changes are the properties detectors measure. Sentence-length variation, word-level predictability, filler density. Change them and the reading changes. There is no mystery in this part.
You can verify it in one pass rather than taking our word for it: run a paragraph, read the before and after. We score with our own detector and label the reading as ours. Three rewrites a day at up to 260 words, no account.
What should make you suspicious is a specific number attached to this claim. "Lowers scores by 94%" is a figure someone chose, because it depends entirely on the input text, the setting, and which detector you ask. Our own house prose, written by humans, reads in the low 80s on our detector, which tells you how little a single number means in this category.
Yes on one axis, and no on the one people care about most.
Yes, on filler. Cutting "it is important to note that" and "in today's rapidly evolving landscape" is a real improvement, and it is most of what a rewrite does that a reader would thank you for. A 1,400-word AI draft usually says its content in 1,100 words.
No, on substance. A rewrite cannot add information that was not in the draft. The characteristic weakness of AI-written text is that it covers a topic without knowing anything specific about it: no numbers you gathered, no example from your own work, no claim anyone could disagree with. That reads as empty whatever the sentence rhythm. No tool fixes this and ours does not either.
And it costs you something. Where your writing had a distinctive turn of phrase, a heavier rewrite will replace it with a more ordinary one, because unusual is what it is removing. On fiction or anything voice-driven that is a bad trade. For authors specifically, this is the main point.
The question people ask least and should ask first.
No, and this is the honest answer rather than a disclaimer. A rewrite operates on wording with no model of what your text is claiming, so three things can go wrong: a hedge gets upgraded into a definite statement, a figure shifts, or a detail appears that was never there.
That last one is not theoretical. While testing for these pages, a Balanced pass returned the sentence "Roughly a year into it" when nothing in the source mentioned any timescale at all. An invented specific that reads as authoritative is the real failure mode of this entire category, well ahead of style, and it is invisible if you read the output for flow.
Which is the practical argument for one feature over all others: a tool that shows you which words changed. Most do not. Without it you cannot audit twenty passes across a long document, and reading the prose will not catch it because the prose is optimised to read well. The check, in full.
And a vendor that claims to is telling you something useful about itself.
Turnitin does not expose its verdict to third-party products. Neither do GPTZero or Originality in the way that would be required. So no external tool can measure its own output against them, which means every published "99% bypass rate" against a named checker is a number the vendor produced by some method it does not describe.
This is the single most useful test for judging anyone in this category, us included: if a tool claims a pass rate against a checker it cannot query, treat its other claims as unreliable too. It is not a borderline exaggeration, it is a measurement that cannot exist.
Two further reasons the question is the wrong one. Detectors disagree substantially with each other on identical text, so "will it pass" has no single answer. And they false-positive on human writing, especially for writers using English as a second language and on formal registers like abstracts and literature reviews, so passing is not the same as being believed. The false-positive evidence is here.
If you have already been accused, no tool helps and a rewrite makes it worse: your draft history is the evidence that defends you, and rewriting replaces an accumulating record with one block of changed text. Gather evidence instead. Whether using one of these tools is permitted at all is set by your institution or client, not by us.
It depends which of four things you mean. They reliably lower a detector score, because they change the exact properties detectors measure. They partly improve the writing, by cutting filler, but cannot add substance that was never in the draft. They do not reliably preserve your meaning, which is the risk most people never ask about. And nobody can tell you whether your document will pass a specific checker, because those services do not expose their verdicts to third-party tools.
No, and a tool claiming to is telling you something important about its other claims. Turnitin does not expose its verdict to third-party products, so no external tool can measure its own output against it. Any published bypass rate against a named checker is a number the vendor produced by a method it does not describe. Detectors also disagree substantially with each other on the same text, so the question has no single answer even in principle.
Sometimes, and it is the most under-discussed risk in the category. A rewrite works on wording with no model of what you are claiming, so a hedged finding can be upgraded into a definite one, a figure can shift, or a detail can appear that was never there. In our own testing a pass returned "Roughly a year into it" when the source mentioned no timescale at all. This is why a word-level diff matters more than a better score.
On one axis yes and on the axis that matters, no. Cutting connective filler is a genuine improvement, and it is most of what a reader would thank you for. But a rewrite cannot add information the draft never had, and the characteristic weakness of AI writing is covering a topic without knowing anything specific about it. A heavier rewrite also flattens distinctive phrasing, because unusual wording is exactly what it removes.
Because nothing stops them. The claim cannot be verified by the vendor making it, since the checkers in question do not expose their verdicts to outside products, and it cannot easily be falsified by a customer either. Treat it as the clearest available signal to distrust the rest of the page. Our own position is that we score with our own detector and label the reading as ours, which is the only measurement we can honestly stand behind.
No, and it can actively hurt. Detectors false-positive on genuinely human writing, particularly for writers using English as a second language and on formal academic registers, so a high score on your own work is a fact about the instrument rather than evidence about you. If you have been formally accused, rewriting destroys the draft history that would have defended you. Gather dated versions, notes and supervisor correspondence instead.
The mechanism behind all four answers above.
Read the explainer →The audit for question three, worked through.
See the method →Why passing is not the same as being believed.
Read the evidence →Three rewrites a day, no account. Answer question one in one pass.
Run a paragraph →Three rewrites a day, no account. Read the before and after, and read what changed.