Search this and you get two separate product categories: rewriters that never show you a number, and detectors that hand you one with no way to act on it. So people run both by hand, in two browser tabs, comparing a score from one vendor against a rewrite from another and hoping the difference means something. It usually does not, for a reason worth two minutes.
We built ours the other way round. The same detector we sell standalone scores your text before and after every rewrite, and shows both numbers with sentence-level highlights, so you can see which specific lines still carry the machine signal. A rewrite that moves nothing shows up as a rewrite that moves nothing, which is the one thing a single-tool humanizer can never tell you.
Most people already do this manually, with two browser tabs. Three things go wrong every time.
Rewriting always makes prose feel different, which is exactly why unverified humanizing is so easy to sell. Without a before number and an after number from the same model, you have no way to separate a rewrite that changed the statistical signal from one that just moved words around. The felt improvement is not evidence.
Humanize in one tool, check in another, and any difference could be the rewrite, or it could be that the two models use different signals and different thresholds. That is not a hypothetical: on the same DeepSeek text in a University of Missouri study, one detector returned 100% and another returned 3.24%. Comparing across models tells you about the models, not about your edit.
Source: Alshammari and Rao (arXiv:2507.17944), six detectors across 294 samples.
A single document percentage tells you almost nothing actionable. What helps is knowing which lines are carrying the signal, because on most drafts it is the same three places: the opening restatement, the transitional scaffolding, and the closing summary. Those are usually the weakest writing in the piece. A separate detector will not map its verdict back onto the sentences a rewriter is about to touch.
Three steps, in this order, and the middle one is the least interesting part.
1. It scores what you pasted. The detector reads your draft and returns a Humanization Score with sentence-level highlights. Before anything is rewritten, you can see where the signal sits.
2. It rewrites the lines that scored, not the whole document. The highlights from step 1 are what the rewriter works on, which is the point of doing this in one loop: a standalone humanizer has no idea which sentences mattered, so it rewrites everything and risks your good paragraphs to fix your weak ones.
3. It scores the result. Same model, same scale, side by side with the original. That is the feature: you are reading a measurement, not trusting a claim.
This matters more than it sounds. The score you see after a rewrite comes from the identical detector we sell as a standalone product, not a lenient in-house scorer tuned to flatter the rewriter. That is also the honest limit of it: a better score is evidence about our detector, and another tool with different thresholds may read the same paragraph differently.
On what that detector is: long-form English of roughly 300 words or more, against 15 current commercial model families, sits at roughly 88 to 92% accuracy in our internal benchmark, dropping to roughly 70 to 78% under about 100 words. Those are our own results, not an independent evaluation, and we label them that way. Our published false-positive measurement is 1,180 academic papers at 5.85%, with the per-document dataset downloadable.
Sources: our accuracy methodology and published benchmark, both of which link the dataset.
The number is useful in a narrow way. Over-reading it is the mistake this whole page exists to prevent.
Detectors read perplexity, how predictable each next word is, and burstiness, how much that varies across the document. A rewrite that varies sentence length and prunes template phrasing changes both. Nothing about where the text came from changes. The published research is consistent on how far that goes: field-wide accuracy fell to about 26% on paraphrased AI text in the 14-tool Weber-Wulff study, and the strongest detectors dropped to between 52% and 71% on humanised reasoning output in the Missouri study.
Substantially reduced is not eliminated. We will not print the word "undetectable", because detectors retrain and a guarantee about a moving target is marketing.
Sources: Weber-Wulff et al. (2023); Alshammari and Rao (2026).
It usually means the draft was already in human register and the detector was reacting to something structural instead: formal academic phrasing, technical vocabulary, or the low burstiness that second-language English and heavily edited prose both produce. That is worth knowing, because the fix in that case is not more rewriting. It is understanding that the signal was never about provenance.
Common, and not your fault. Our own benchmark puts the false-positive rate at 5.85% on academic papers, and a Stanford study measured a 61.3% average false positive rate across seven detectors on non-native TOEFL essays. Use the highlights to see what the tool is reacting to, then keep your drafts and version history, which is the only evidence in this area that no tool can manufacture.
Source: Liang et al. (2023), Patterns (Cell Press). The seven detectors did not include ours.
The loop makes the number honest. It does not make the number mean more than it does.
It will not make text undetectable, and we will not use the word. See the numbers above for why.
The score is ours. A better Humanization Score is evidence about our detector, not a universal clearance. Another tool may read the same paragraph differently, which is the entire reason the two-tool comparison above is unreliable.
Both halves are English-only. The scoring model is English-optimised, so a Humanization Score on other languages is not a reading you should act on.
Long documents get chunked. A 10,000-character ceiling applies to each request, and a new free account caps each one at 1,500 words, so a dissertation goes through in passes rather than one paste.
A moved score is not a finished essay. The loop measures cadence, not whether your argument holds or your sources are real. A document can score perfectly human and still be thin, wrong, or unsourced, and no part of this tool is looking at that.
The loop runs 3 times a day on 300 words without an account. A new free account raises that to 260 words per rewrite plus 3 scans a day of detection, and paid tiers start at $9.99/month for 20,000 words ($7.49 billed annually). Full ladder on the pricing page.
It is two tools in one workflow: a detector that scores how machine-written text reads, and a rewriter that changes the cadence, with the detector running before and after so you can see what the rewrite actually did. Most products give you one or the other, which means you either rewrite blind or get a number with no way to act on it. Running both in one loop is what makes the number useful.
You can, and many people do with two browser tabs, but you lose the comparison. Different detectors use different signals and thresholds, so a difference between tools tells you about the models rather than about your edit. In one published study, the same DeepSeek text scored 100% on one detector and 3.24% on another. A before and after from the same model is the only reading that isolates what the rewrite changed.
No. The score shown after a rewrite comes from the identical model we sell as a standalone detector, on the same scale. That is also why we are explicit that a better score is evidence about our detector specifically, not a universal clearance. If we used a lenient in-house scorer the number would be meaningless, which is the problem this page is about.
Not necessarily, and treat anyone who promises that with suspicion. The published research shows rewriting cuts detection substantially across the board, with field-wide accuracy falling to about 26% on paraphrased AI text and the strongest detectors dropping to between 52% and 71% on humanised output. Substantial reduction is not elimination, detectors retrain, and we will not print the word "undetectable".
That is useful information rather than a failure. It usually means the draft was already in human register and the detector was reacting to something structural: formal academic phrasing, technical vocabulary, or the low burstiness that second-language English and heavily edited prose both produce. More rewriting will not fix a signal that was never about provenance.
Meaning is tuned for ahead of score movement, and the loop is part of why: because the rewriter only touches the sentences that actually scored, it leaves the rest of your argument alone instead of reworking paragraphs that were never the problem. Read the output anyway on cited or numerical material, and check the figures survived.
Yes. 3 rewrites a day at up to 300 words each with no account at all, both scores shown. On a new free account you get 260 words per rewrite, plus 3 scans a day of detection. Paid starts at $9.99/month for 40,000 words, or $7.49/month billed annually.
The rewriter itself, with the score shown before and after every pass.
Open the humanizer →The same model on its own, with sentence-level highlights and no rewriting.
Open the detector →What the before-and-after number measures, and how to read a partial improvement.
Understand the score →1,180 academic papers, 5.85% measured rate, per-document dataset downloadable.
Check our numbers →The no-account version: 3 rewrites a day, up to 300 words each.
Try it free →If you wrote it yourself and it still scores high, start here.
Read the guide →Score the draft, rewrite only the lines that scored, score it again on the same model. That comparison is the whole product. 3 rewrites a day free, no signup and no card.