Sometimes, in some detectors, by amounts that vary enormously depending on the tool, the text, and which detector you check against. It is an accurate answer and a less satisfying one than either the marketing claim or the flat denial. There is no widely accepted independent benchmark comparing these products under reproducible conditions, and many prominent success-rate claims come from the vendors themselves. The tools have also changed underneath the label: some still swap synonyms and reorder clauses, while others hand the whole passage to a language model and return a full rewrite. What can be said with more confidence concerns the costs the pricing pages tend to omit: rewriting changes your wording, your emphasis, and sometimes your meaning, and Turnitin reports AI-paraphrased text as a distinct category. This article covers what these tools do to a draft, what the research actually found, and what to do instead.
What these tools do to a draft
The category has changed, and descriptions written two years ago no longer fit it.
Older tools worked mainly by substitution: swap words for synonyms, reorder clauses, occasionally split a sentence. Some still do. Newer ones pass the text to a language model and ask for a full rewrite, which can change sentence count, paragraph structure, register, emphasis, and occasionally the content itself.
That difference matters when you are deciding whether to use one. A substitution tool leaves your structure and alters your word choices. A model-based rewriter can return something that argues slightly differently from what you wrote. Product pages do not always explain which approach a tool uses, so the only way to know is to compare the output against your original line by line.
Run that comparison before you accept anything. It takes a few minutes and it is an easy step to skip.
Why scores move
Detection scores can fall after rewriting, sometimes substantially, although the result varies by text, by rewriting method, and by which detector you check against. The reason is less mysterious than either side of the argument suggests.
These systems estimate a likelihood from patterns in the text. Change enough of the text and the estimate changes. Nothing more complicated than that, and it says nothing about whether the writing got better, or whether a different detector would agree, or whether the result holds next month after the detector is updated. What a detection score can and cannot support covers why a moved number is weaker evidence than it looks.
Research supports the general direction while being narrower than it is usually quoted. A study from Maryland and Harvard found that recursive paraphrasing can substantially reduce detection performance across several detection approaches, with the largest reported effects coming from repeated rounds of rewriting rather than a single pass. The paper reports that this was achieved with limited degradation to text quality by the measures it used. We should state that accurately rather than bend it toward our own argument. The study does not establish that rewriting ruins prose. It establishes that detection does not hold up against a determined rewriting attack.
The costs that are not on the pricing page
The research measured detection performance. What follows is what we see in practice when comparing rewritten drafts against originals, offered as observation rather than as a finding.
Specific words get replaced with general ones. If you wrote that a policy was clumsy and the output says inefficient, a judgment has quietly been removed. Precision tends to be the first thing lost.
Register drifts upward. Output from these tools often reads more formal than the input, in a way instructors and editors encounter frequently and recognize.
Connotation errors appear. Substitution does not track how a word feels. Replacements come back subtly wrong in ways a reader notices without always being able to name.
Evidence gets smoothed away. A rewriter has no way to know that a particular example was the point of a paragraph, and full-rewrite tools sometimes drop or generalize it.
None of these show up in a detection score, which is the awkward part. The number may improve while wording, emphasis, or evidence shifts in ways the score does not capture.
What Turnitin already reports
AI-paraphrased text is not an unwatched gap. Turnitin's AI writing report includes its own reported category for text it judges likely AI-generated and then likely revised with a paraphrasing tool, highlighted separately, with QuillBot named as an example in the documentation. Turnitin also states that its AI-generated category can include text modified by tools intended to defeat detection.
Two qualifications keep this honest. The category describes a two-step pattern, AI text that was then paraphrased, rather than any text a paraphraser touched. And Turnitin does not publish how well that category performs against any specific product, so nobody outside the company can say how often it succeeds. Turnitin reports AI-paraphrased text as a separate category covers the reporting structure, along with Turnitin's published file and language requirements.
The more reliable point has nothing to do with detection. A person reads the work afterward, and that person assigns the grade. Prose rewritten repeatedly to satisfy a scoring system can develop noticeable shifts in register, emphasis, or internal consistency, and those are things a human reader picks up on. Clearing the machine and losing the reader is not an improvement on the situation you started in.
Improving a draft without optimizing for a detector
The framing worth adopting is not how to move a number. It is whether the draft says what you meant, which is a question with an answer that does not expire when a model is updated.
Check that every sentence still carries your meaning. Read the rewritten version against your original. Anywhere the claim shifted, revert it.
Put your specific examples back. Named cases, figures, and concrete details are what make writing yours, and they are what generalizing rewriters remove first.
Correct substitutions that changed the connotation. If a replacement word is more formal but less true, it is the wrong word.
Check that transitions match the actual logic. Rewriters insert connectives freely, and a "therefore" that does not follow is worse than no connective at all.
Read a paragraph aloud. Sentences you would never say out loud are usually the ones a reader will stumble over, whatever any tool reports.
Read your institution's AI policy before your assignment, not after. Policies differ substantially on what assistance is permitted and what must be disclosed, and no tool changes what yours says.
Keep your drafts and revision history. If a question ever arises, that is the material that speaks to it.
Where a detector genuinely helps here is narrow. It points at passages a reader might misread, and then you decide what to do with them, in your own words. Unfox is an AI detector that reports at the sentence level for exactly that reason, and if it becomes something you use regularly, plans and word limits are on the pricing page.
When a rewriting tool is the wrong instrument
After an accusation. Rewriting the document changes the artifact under discussion and is very hard to explain afterward. What to do instead if you have been accused sets out the alternative, which is process evidence rather than a new file.
When the writing is already yours. If you wrote it and a detector disagreed, you have a measurement problem rather than a text problem, and degrading the text to satisfy a measurement is a poor trade.
When policy is the actual question. If your institution prohibits AI assistance, the question is about the work rather than the software, and no rewriting pass changes that.
The narrow case where these tools earn their place: a specific passage you have already identified, that you want tightened while keeping your meaning intact, with you reading every change. Used that way it is an editing tool. Used as a blanket pass over a finished document, it is a way to hand over decisions about your own writing to something that does not know what you meant.
FAQ
Do AI humanizers actually work?
They can lower scores in some detectors, and results vary widely by tool, by detector, by text length, and by how aggressively the rewriting is set. There is no widely accepted independent benchmark comparing these products under reproducible conditions, so many prominent success-rate claims come from the vendors. Rewriting also changes wording and emphasis in ways a score does not capture.
What is the best AI humanizer for Turnitin?
We are not in a position to name one. There is no widely accepted independent comparison that establishes a best product for Turnitin, and Turnitin does not publish performance against individual products. Turnitin does report AI-paraphrased text as a separate category, which makes any product claiming reliable results against it hard to verify.
Can Turnitin detect humanized text?
Turnitin's report includes a category for text judged likely AI-generated and then likely revised with a paraphrasing tool, and states that its AI-generated category can include text modified by tools intended to defeat detection. How well that performs against any particular product is not published.
What actually lowers an AI detection score?
Changing the text changes the estimate, which is why rewriting moves the number. Whether that is worth doing is a separate question. If the writing is yours, a lower score buys very little, and the trade can cost precision or voice in ways the score does not show.
Does rewriting help after I have been accused?
No, and it usually makes things harder. Modifying the document after a concern is raised changes what is being discussed. Version history, dated drafts, and being able to explain your reasoning are what address the question actually asked.
Is using an AI rewriting tool against the rules?
That is set by your institution or publisher, not by the tool. Policies vary on what assistance is permitted and what must be disclosed, and tightening a passage you wrote is often treated differently from generating text you did not. Read the policy that applies to you.
