Before using an AI humanizer for researchers, understand the constraint: in academic writing it is not style, it is precision. A hedge is a claim. "The data suggest" and "the data show" are different findings, and only one of them may be true of your results. A rewriter sees two phrasings of a sentence; a reviewer sees two different assertions about what you demonstrated.
So the recommendation here is narrow and it runs against the usual sales pitch. Use the lightest setting, keep it away from your methods and results, and check your terminology afterwards. A rewrite is a reasonable aid for an introduction or a cover letter. It is a poor idea for the sentences your contribution rests on.
Four places a humanizer changes the science rather than the prose.
Hedging, in both directions. Hedges read as a detector tell, so rewrites strip them. In an introduction that tightens the writing. In a results section it can turn a cautious observation into a claim you did not make, and the reverse is just as bad: a qualifier added to a finding you stated firmly has also changed your result.
Terms with an ordinary-English twin. Significant, power, bias, validity, control, model, theory, hypothesis. Each has a precise sense in some field and a loose sense everywhere else. A rewriter has no way to know which you meant, and a substitution reads perfectly well while meaning something else.
Numbers and their units. Digits survive; the words around them are exposed. Roughly 40% returning as 40% deletes a hedge you placed deliberately, and percentage points traded for percent is a different measurement rather than a rewording.
Citations. The reference marker stays and the proposition drifts. A softened statement of a finding, still carrying the citation, now attributes to the source something the source did not report. In peer review that is worse than clumsy prose.
The general treatment is in humanizing without changing meaning. Research writing is the acute case because the cost is a misreported finding rather than a weaker sentence.
A manuscript is not one kind of writing, and the answer differs across it.
Reasonable to rewrite: the introduction and background, the discussion’s broader framing, a cover letter to an editor, a plain-language summary, a grant lay abstract, a conference bio. Here the wording is not load-bearing and AI-assisted drafts that read mechanical are a genuine problem a humanizer solves well.
Keep the humanizer away from: methods, results, statistical statements, the abstract’s claim sentences, limitations, and anything carrying a figure or a citation. These are the sentences a reviewer reads as assertions about what you did and found.
Work a paragraph at a time rather than a section at a time. It is the only scale at which you will genuinely read each change, and it is also what the free humanizer allowance suits. Run the humanizer on one paragraph, read the pair, keep what survives, move on.
Three modes: Light, Balanced, Maximum. Use Light. It works on sentence rhythm and connective filler and leaves your vocabulary alone, which is exactly the property a manuscript needs. Balanced edits phrasing more broadly, so keep it to non-technical prose. Maximum replaces most of the wording, is a Pro mode, and has no place in a manuscript section. The free tier covers Light and Balanced.
Most major publishers now have a stated position on generative AI in manuscripts. The common shape is that AI cannot be an author, and that AI assistance in drafting should be declared, often in the methods or acknowledgements. Policies differ and they change, so read the one for your target journal rather than trusting a summary, including this one.
The point worth making is that a rewrite does not answer a disclosure question. If your journal asks you to declare AI assistance, running the text through a rewriter does not remove the obligation; it only makes the assistance harder to see, which is a worse position to be in if it emerges later. Co-authors have a claim on knowing too.
We sell a rewriter, so weigh that framing accordingly. We would rather say it than let the page imply the tool settles something it cannot.
One: search your technical terms. Keep a short list of the five or ten terms that are terms of art in your field and search the rewritten passage for each. Missing or altered ones are defects. Under a minute, and it catches the hardest category to spot by reading.
Two: check every hedge. Compare the modal and qualifying language against the original. Suggest, show, demonstrate, may, appears to are findings, not style.
Three: read every sentence with a citation or a figure. Confirm the proposition still matches what the source reported and what your data show.
Four: re-run the humanizer score, and accept the honest answer. We score the passage before and after with our own detector and label the reading as ours. If you reverted enough that the reading is back where it started, that passage needs editing by hand. That is a legitimate result and a better one than a manuscript you cannot defend.
On unpublished data, read our privacy policy and make your own judgement before pasting anything under embargo or covered by a data agreement into any third-party tool, ours included.
On parts of it, and with care. The introduction, background, discussion framing, cover letter and lay summary are reasonable candidates because the wording is not load-bearing. Methods, results, statistical statements, limitations and the abstract’s claim sentences are not: those are read as assertions about what you did and found, and a substituted term or a stripped hedge changes the claim. Use Light, work a paragraph at a time, and check your terminology after every pass.
Four things. Hedges, where suggest and show are different findings and a rewrite strips or adds them freely. Technical terms with an ordinary-English twin, such as significant, power, bias and validity, where a synonym reads well and means something else. Numbers, where units and comparatives are exposed even though digits survive. And citations, where the marker stays while the proposition around it drifts away from what the source reported.
Usually yes if you used it in drafting, and you should read your target journal’s policy rather than any summary, including ours. The common position among major publishers is that AI cannot be an author and that assistance in drafting should be declared, often in the methods or acknowledgements. A rewrite does not answer this question. It changes the prose, not the obligation, and it makes the assistance harder to see rather than absent.
Light, in essentially every case. Light works on sentence rhythm and connective filler and leaves your vocabulary alone, which is the property a manuscript needs. Balanced edits phrasing more broadly, so keep it to non-technical prose such as a cover letter or a lay summary. Maximum replaces most of the wording, is a Pro mode, and should not be used on a manuscript section at all.
That is your judgement to make against your own obligations, and we will not make it for you. Read our privacy policy, and consider any embargo, data-sharing agreement or ethics approval that applies to the material. For anything genuinely sensitive the conservative answer is not to paste it into a third-party tool. Rewriting an introduction does not usually require including your data at all.
We are not going to claim that. We cannot see another system’s verdict, screening tools retrain against rewriting techniques, and a promise about a system we do not control would be marketing rather than a measurement. What we provide is a before and after reading from our own detector, labelled as ours, which tells you whether the statistical properties this family of tools reads have moved.
This page is about you as a researcher: which sections are safe, what your journal’s disclosure policy means, and the checking habit before a draft reaches a co-author. The section-by-section walkthrough of working through a paper is a separate guide, linked below. Read this one to decide whether and where to rewrite, and that one for the mechanics once you have decided.
The mechanics once you have decided which sections to touch.
See the guide →The general method: what drifts first and how to check it.
Read the method →The same problem where a drifted word changes an obligation.
Read the guide →Why detectors over-flag non-native English, with the research to cite.
Read the evidence →What our score cannot tell you, published rather than buried.
See the limits →The full tool, scored before and after by our own detector.
See the tool →Score a passage and read what changed before it reaches a co-author or a reviewer. 3 scans a day at up to 1,000 words, no account needed.