Yes, you can run a manuscript through an AI checker before you submit, and the way that works in practice is checking it section by section rather than as one document. Paste a section into the box, press check, and you get a 0 to 100 score with one of three verdicts: reads as AI-generated, borderline, or reads as human-written. The check needs at least 65 words and costs one credit whatever the length. Checking the abstract costs exactly what checking the whole methods section costs. The free allowance is 10 credits a day per network, with no account, no login, no card and no email address, which is enough to cover an entire paper in one sitting. Two things are worth settling before you paste anything. First, a flagged methods section is not evidence of anything. Methods are written to be reproducible, and reproducible writing is uniform and predictable, which is exactly what these scores respond to. Second, the journal question about AI use is answered with a written statement from you, not with a number from a classifier. This page covers how to run the check properly, which sections to trust, where a rewrite actually helps, and what to produce if anyone ever asks.
Many journals now ask about generative AI somewhere in the submission portal or the author guidelines. The wording differs by publisher, by journal and by year, so the only safe move is to read the guidelines for the journal you are submitting to, on the day you submit. Those policies tend to converge on a short set of principles rather than a numeric threshold.
A disclosure statement is short. Name the tool, say which part of the work it touched, and say what you did to verify the output. One or two sentences in the methods or the acknowledgements, plus whatever field the portal gives you. It should not contain a detector score, because a score from one company says nothing about the policy you are disclosing against.
On the underlying question, briefly. Rewriting a draft you were permitted to generate is ordinary editing. Where the policy that binds you does not permit generated text, that rule covers the draft however it is edited afterwards. Read the policy that applies to you rather than the one that applies to your colleague down the corridor. The AI detector gives you the reading; the disclosure gives the journal what it actually needs.
Procedure written to be reproducible uses fixed terminology and no elegant variation. You call the same instrument the same name in the same construction every time. Sentences are short, declarative and structurally alike, and reporting checklists push whole paragraphs into a fixed order. That is highly predictable prose, and predictability is what the score measures.
There is a second reason specific to your field. Standard passages, such as sample preparation, consent language and statistical software descriptions, appear in near-identical form throughout the published literature, and a model trained on that literature reproduces them easily. Your correct conventional version and a generated one look alike. The resemblance is a property of the genre, not a finding about you.
The practical consequence is to read the sections of a manuscript against each other rather than against a cut-off. If methods is the only part reading as machine-written while the discussion reads as human, that pattern is the explanation. The false positives page goes into why this happens in detail.
Split at section headings: abstract, introduction, methods, results, discussion, conclusion. Where a section runs long, split again at subheadings rather than mid-argument, because the rhythm measurements need whole paragraphs to mean anything. Every real section clears the 65-word minimum, and because a check costs one credit at any length, ten sections is one free day.
Strip everything that is not your own prose before you paste, so the reading describes your writing rather than your bibliography. Then record the result for each section in a small table. The spread across sections is the finding, not any single number, because the section that reads differently from the others is the one that tells you something.
The registry of 44 named AI writing tells runs in your browser with no daily limit, and it is the better first stop on a manuscript. It highlights the exact phrase instead of scoring the section, and a phrase is something you can fix without touching your terminology.
On a multi-author paper the person signing the disclosure is usually the corresponding author, signing for text they did not write. Ask each co-author what they used and where, before submission rather than after a query arrives. It is a two-line email, and it is far cheaper than a correction notice.
Then check the sections you did not draft, and every citation inside any passage a tool helped produce. A fabricated or mismatched reference is the one item a reviewer can settle without you, because a reference either exists and says what you claim or it does not. Resolve the DOI, open the paper, and confirm the sentence you attributed to it.
Do the same for any number that reached the discussion from memory rather than from your results table. Rewriting changes how a sentence sounds, not whether it is true. A rewriter works only from the text you give it, so a wrong figure, a mismatched citation or a claim your data does not support comes back out of the pass reading better and still wrong.
The humanizer rewrites every sentence in one click, up to 12,000 characters per run, roughly 2,000 words, with a 25-word minimum, at one credit per 50 words. On a manuscript that makes it a tool for prose, not for precision.
Introduction framing and discussion prose are where the stock openers, the stacked hedges and the uniform sentence lengths collect, and a rewrite there costs you nothing you need. Methods, results and any sentence carrying a number are different. The exact reagent name, the exact statistical test and the term you defined on page two are content, not packaging. Read every rewritten technical line against the original.
If a passage in your methods still reads flat after that, leave it flat. Flat is what a correct methods section reads like, and a paper smoothed to one even voice from abstract to conclusion has lost something a careful reader expects to find. For the arithmetic on long manuscripts, the humanize 5,000 words page works through the credit math.
The reading describes patterns in the text you pasted. It does not identify who wrote it, and no detector output can. Different systems are trained on different text and set their cut-offs differently, so a reading here is not a prediction of what any publisher's screening will say. The what an AI score means page explains the number itself.
The errors are also not spread evenly across authors. Liang and colleagues at Stanford HAI found in 2023 that 61% of TOEFL essays by non-native English speakers were falsely flagged as AI across seven detectors. What that study points to is careful textbook construction, and the word a writer is certain of rather than the one they half remember. That is how a great deal of research writing in a second language reads, and the page for second-language writers covers it directly.
If an editor or a reviewer raises the question, do not argue about the number. Ask which passages prompted it, then answer with what a classifier cannot supply: your data, your code, your drafts and your correspondence.
Most of it you already have. Version history in Overleaf, Word or Google Docs. Git history on the manuscript and the analysis. Notebooks and raw data files with timestamps. Dated drafts circulated to co-authors. A preprint posted before submission timestamps the text publicly. Put it in date order, because the order is what makes it read as a record rather than a pile.
The signed drafting record adds what finished files cannot show, which is the shape of the session: when you worked, how much you cut and rewrote, and whether large blocks arrived in a single paste. It signs that record with an ECDSA P-256 key and commits it to your exact text with a SHA-256 hash, so anyone can check the signature and the match on the verify page without contacting us. It stores lengths and timings, never your content, and it is free with no daily limit.
It records forward, not backward. Start it on the paper you are drafting now rather than the one already under review, because the record only exists from the moment you begin it.
You can, but running it section by section is the better use of the tool. One reading across an entire manuscript averages your flattest and most distinctive writing together, which hides the section you need to look at. A check needs at least 65 words and costs one credit whatever its length, so the abstract and the methods section cost the same, and ten free credits a day covers a full paper. Rewriting is capped at 12,000 characters per run in any case, roughly 2,000 words.
That is expected, and it is the ordinary reading for a methods section. Reproducible procedure uses fixed terms, no synonyms and repeated sentence shapes on purpose, which is the shape of predictable text. Check the introduction and discussion as well. If methods is the only section reading as machine-written, that pattern is the explanation. The same low variability is what Liang and colleagues at Stanford HAI pointed to in 2023, when 61% of TOEFL essays by non-native speakers were falsely flagged as AI across seven detectors.
That depends on the journal, and journal policies differ on where the line between language editing and text generation falls. Read the author guidelines for the specific journal on the day you submit rather than relying on what a sister journal said last year. When you do disclose, keep it to what you used, which part of the manuscript it touched, and how you verified the result.
No tool can promise a particular result from a particular system, and we do not claim one. What you can do is check your own sections, read the named tells, fix what you agree with, and answer the disclosure question honestly. The submission question is settled by your statement, not by any score.
Ask the editor which passages prompted the comment, then respond to those specifically. Supply version history, dated drafts, analysis code and raw data, and explain the writing conventions that make the flagged section read as uniform. Keep it factual and short. Detector output does not identify authorship, and a response built on your materials is stronger than one built on a counter-score.