To humanize AI text without changing meaning: use the lightest mode that clears your bar, work a paragraph at a time, and read the diff before accepting it. Drift happens when the rewrite is doing more work than it needs to. The four things that break first are citations, technical terms, hedges and numbers.
That is the whole method. The rest of this page is why each of those four breaks, how to tell quickly whether a rewrite kept your argument, and the case where a humanizer is simply the wrong tool to reach for. One position up front, since it shapes everything below: on work that carries citations, we would not run Maximum mode unattended. Light is the only setting we would trust without reading every changed sentence, and you should read them anyway.
Worth understanding, because it tells you exactly which parts of your draft are safe and which need checking.
Detectors mostly read two things. How predictable your word choices are, and how much that predictability wobbles from sentence to sentence. Flat and predictable reads as machine. Which means the edits that shift a score are edits to rhythm and phrasing. Not to content.
Which is good news. Sentence length carries almost no meaning, and neither does whether a paragraph opens with "Furthermore" or simply starts. Vary the one, delete the other, and your argument has not moved an inch. Those edits are free.
Drift starts later, when the rewrite runs out of rhythm to fix and reaches into your vocabulary instead. A synonym that fits the sentence may not fit the field. Significant is the obvious one: ordinary English in a blog post, a claim about a p-value in a results section. Theory and hypothesis sit beside each other in a thesaurus and name different objects in a methods chapter. Your tool cannot tell which you meant, and it will not ask.
Drift, then, is roughly proportional to how much vocabulary substitution you asked for. That is a setting. You control it.
Check these four and you have covered nearly every case where a humanized draft ends up saying something you did not. They are listed in the order they tend to go wrong.
The reference marker almost never moves. The claim it was attached to does.
Here is the shape of it. You wrote that one study measured an average false positive rate of 61.3% across seven detectors on TOEFL essays by non-native writers, their measurement rather than ours (Liang et al., 2023). The rewrite hands back "research suggests detectors may struggle with non-native writing." Fluent, and no longer supported by the figure you were citing. A reader who opens the source finds a claim you did not make. Nobody catches this on a read-through, because nothing reads wrong. It surfaces when someone checks, which is the one moment you least want it to.
Check every sentence carrying a reference. If the claim changed, put your original sentence back.
This is the one that catches people out. Terms of art look like normal words, so they're the first thing a rewrite reaches for. Significant, bias, power, validity, model, control. Each has a precise sense in some field and a loose sense everywhere else, and swapping the precise one for a casual paraphrase quietly changes a claim into a different claim.
The fix is to build a short do-not-touch list for the piece you're editing, then search for those words in the output and confirm they're still there.
Hedging reads as a detector tell, so rewrites strip it. In a blog post that is usually an improvement. In a paper it can amount to a retraction. "The data suggest a link" and "the data show a link" are two different claims, and only one of them may be true of your data. It runs the other way as well. A cautious qualifier bolted onto a finding you stated firmly has also changed your result, just in the direction nobody thinks to check.
Digits survive. It is the words wrapped around them that are exposed. Roughly 40% returning as 40% has quietly deleted a hedge you put there on purpose, and percentage points traded for percent is not a rewording, it is a different measurement. Anything reporting a result earns a slow pass. Digit by digit, qualifier by qualifier. No detector score is worth a misreported number.
The AI humanizer ships three modes. The gap between them is how much vocabulary each is willing to replace, which is also the gap in how much meaning each can move. If you want to humanize AI text without changing meaning, this single setting matters more than everything else on this page.
Works on rhythm, and on the connective tissue models overproduce. It varies sentence length, deletes "it is important to note that", and otherwise leaves your words where you put them. Use it wherever the wording is load-bearing. Papers, legal writing, documentation, a case study with real figures in it. Yes, it moves scores less than the other two. It also almost never surprises you, which on a thesis chapter is the property you actually want.
Edits more broadly, including phrasing. Right for general prose where the exact wording isn't carrying weight, which covers most blog posts, newsletters and internal writing. Read the diff, but you won't usually be reverting much.
Replaces most of the phrasing, and it is a paid mode: the free tier covers Light and Balanced. Genuinely useful for marketing copy you own outright and can rewrite freely. It is the wrong instrument for a dissertation chapter, and using it there is how people end up defending a sentence they did not write and do not agree with.
Start at Light. Re-score. Move up only if you need to, and only on the paragraphs that actually need it, which is usually not the whole document.
Not a full re-read. A targeted pass over the parts that actually carry risk.
Reading the rewritten text on its own is close to useless for this. It reads well, which is the point, and a fluent sentence gives you no signal about whether it still says what you meant. The before and after side by side is where a changed claim becomes obvious. Our humanizer pairs the two sentence by sentence for that reason. Worth knowing what the score beside it is: the detector reading is for the whole passage, before and after, not a verdict on each line.
Take the five or six words that are technical in your field and search the output for each. Missing ones are your drift list. This takes under a minute and catches the failure that's hardest to spot by reading.
You already know which ones they are. That's typically a tenth of the document, and it's where nearly all of the real risk lives.
Paste the edited version back through the detector and look at what actually moved. If you reverted enough sentences that the score is back where it started, the honest conclusion is that this paragraph needs rewriting by hand rather than by tool. That's a legitimate outcome, and it's better than shipping a version you can't defend. We score before and after with our own detector and label it as ours, which tells you whether the statistical properties moved. It is not a prediction of what anyone else's detector will say.
If you wrote something yourself and a detector flagged it, preserving your meaning isn't really the problem you have. Rewriting replaces the very thing you're trying to defend, and it destroys the draft history that would have supported you. Gather process evidence instead: version history, notes, earlier drafts, a conversation about your own argument. How to prove you didn't use AI covers what that looks like in practice.
And the obvious one. A humanizer changes text; it doesn't change where the ideas came from. If a passage needs a citation, it still needs a citation afterwards. Rewriting copied material to make it look different is plagiarism with extra steps, and no detector score makes that not true.
It can, and how much depends almost entirely on how hard you push it. A light pass that varies sentence length and cuts transitional scaffolding rarely touches your argument, because it's editing rhythm rather than substance. A heavy pass that replaces most of the vocabulary is where drift starts, since a synonym that fits the sentence doesn't always fit the field. Use the lightest setting that gets you where you need to be, work in paragraphs rather than whole documents, and read the before and after side by side instead of accepting the whole rewrite at once.
Yes, and this is the failure that matters most in academic work. The break is usually not the citation marker but the sentence around it. If a rewrite softens a claim that a source supported directly, or merges two sentences that cited different sources, the citation is now attached to something the source didn't say. That's a citation error you introduced and can't defend in a viva or an integrity meeting. Check every sentence carrying a reference, and if the rewrite changed the claim, revert that sentence.
Figures normally survive, but the words around them aren't safe. Units, hedges and comparatives are the vulnerable part: roughly 40% can come back as about 40% or as 40%, and the difference between approximately and exactly is a real claim. Percentage point and percent are frequently swapped and they're not the same thing. Anything reporting a result should be checked digit by digit and qualifier by qualifier.
Paraphrasing restates a passage in different words, aimed at avoiding duplicate wording. Rewriting is the broad term for changing text at all, whether for tone, length or clarity. Humanizing is rewriting with a specific target: the statistical regularity detectors read, mainly uniform sentence length and highly predictable word choice. They overlap a lot in practice. The distinction matters here because a paraphraser is optimising for different words while a humanizer is optimising for different rhythm, and rhythm can usually be changed without touching what you said.
Light. It works mainly on sentence rhythm and on connective scaffolding, so it's the mode least likely to reach into your vocabulary. Balanced edits more broadly and suits general prose where exact wording isn't load-bearing. Maximum replaces most of the phrasing and is right for marketing copy you own outright, not for a dissertation chapter. On anything carrying citations or technical terms, start at Light.
Rewriting isn't what makes text original. If the ideas came from a source they still need a citation afterwards, and running a passage through a humanizer to make copied material look different is plagiarism with extra steps. A humanizer also doesn't run a similarity check, so it can't tell you whether your text matches a published source. That's a separate tool answering a separate question.
Without an account, 3 rewrites a day at up to 300 words each, which is enough to test the behaviour on a real paragraph before deciding anything. With a free account it's 260 words per rewrite. Since working in paragraphs is also what protects your meaning, the free allowance fits the method on this page reasonably well: a paragraph at a time, read the diff, keep what works.
The full tool: rewrite, then read the before and after with the two versions paired up.
See the tool →What our number measures, and what a before and after reading can and cannot tell you.
Understand the score →Rewrite and re-score in the same place, so you can see what each pass actually moved.
See the loop →The three modes and the sentence evidence behind each change.
Read the guide →The same engine framed as a converter, with a before and after on real output.
Convert some text →The right instrument when the problem is authorship rather than phrasing.
See the evidence list →Rewrite a paragraph and get a before and after reading from our own detector, labelled as ours, with the two versions paired sentence by sentence so you can see what moved. Read the pairs, keep your meaning. 3 rewrites a day, 300 words each, no account.