Do journals check papers for AI writing?

Not necessarily. Wiley describes routine screening for duplicated text, which is a similarity check rather than an AI writing score. iThenticate sells AI writing detection as a paid add-on, so institutional licensing and enablement are one gate. A particular submission then has to meet Turnitin's file-type, language and qualifying-word requirements, so a missing AI section does not identify one cause by itself. In the seven named publisher and ethics-body policy pages we reviewed on 18 August 2026, none supplied a universal journal AI percentage to aim at. Those documents focused instead on authorship, responsibility and disclosure.

So "is there an AI rate limit for journals" is a question built on a step that may not exist for your submission. The answerable questions are narrower: what is actually run, where a number would come from if there is one, and what you are obliged to declare regardless.

Publishers do describe screening incoming manuscripts, and at least one of them writes out what the screen is for in a single sentence. From Wiley's publishing ethics guidelines, under the heading on plagiarism:

"Wiley routinely screens submitted manuscripts for duplicated text using tools such as Crossref Similarity Check. If significant and unexplained overlap is identified, we will follow COPE guidelines on handling plagiarism in a submitted manuscript or published article to resolve the issue."

Read the object of that sentence. Duplicated text. Crossref's page for the service then describes iThenticate as providing a Similarity Report "containing a Similarity Score and a highlighted set of matches to similar text". That description gives you the object being measured without asking a word-count search to stand in for the page.

That is a different measurement from an AI writing score, and it is different in kind rather than in degree. Who runs it, what your sentences are compared against, and why your own earlier paper can turn up as a match are all in who runs the similarity check on your manuscript, and what it compares against. One thing worth carrying into the next section: whether your manuscript goes through this particular check is a fact about your publisher, not about your field.

Same vendor, different purchase. Turnitin's iThenticate product page answers this in its own FAQ:

"iThenticate customers can access AI writing detection as a paid add-on. Once licensed and enabled, Turnitin's AI writing indicator, which is integrated within the Similarity Report, shows an overall percentage of the document that AI writing tools such as ChatGPT may have generated."

The feature list on the same page draws the same boundary: the Similarity Report appears under "Core features", while "AI writing detection" appears under "Add-ons".

Licensing and enablement are only the first gate. Turnitin's own report guide shows an "AI writing detection is unavailable for this submission" state whose possible reasons include an unsupported file type, an unsupported language, or qualifying text below 300 words or above 30,000 words. A separate state says detection was disabled at the time of submission. So "the journal checked my paper" can describe different events, and if someone quotes you a figure it is fair to ask which report it sat in and whether the file was eligible for that report.

A missing AI section therefore tells you only that this report does not display an AI result. It does not, by itself, distinguish licensing from submission eligibility or processing state. Ask the report owner which condition applied. The documented language boundary is covered separately in can Turnitin detect AI in non-English submissions?.

On 18 August 2026 we reviewed seven specific policy pages from six organisations. Naming them matters because this is a bounded document review, not a claim about every publisher or every journal.

None of those seven documents supplied a universal journal AI percentage for authors to aim at. They addressed questions such as who can be an author, who remains responsible for the manuscript, and when AI use should be disclosed. That finding is limited to those named pages as read on that date. It does not prove that no publisher or journal has an additional rule, and it says nothing about whether an institution licensed a detector.

Elsevier is not part of that seven-page finding because its policy page did not load reliably in our browser. We use one directly recoverable disclosure sentence from it below and make no absence claim from the failed retrieval.

The seven-page review gives no universal percentage to target. It gives a map of publisher-level disclosure questions, which still has to be narrowed to the journal and submission system handling your paper.

The recurring issue in the reviewed policies is telling people what was done, although the exact boundary varies. Elsevier's journals policy puts one version of that boundary in two sentences:

"Basic checks of grammar, spelling and punctuation do not need a declaration statement. However, when an AI tool makes substantive changes to sentence structure or organization of a part of the text, this should be disclosed."

That line is worth sitting with if any part of your workflow reshapes sentences, because it is drawn around what the tool did to the text rather than around what the tool is called.

COPE's position says AI cannot be an author because it cannot take responsibility, declare conflicts or manage licences. It then says authors who used AI in writing, producing images, or collecting and analysing data "must be transparent in disclosing in the Materials and Methods (or similar section) of the paper how the AI tool was used and which tool was used". That is guidance about responsibility and transparency, not a journal-specific score to target.

Where the statement goes, and what each publisher exempts, is the part that varies and the part people get wrong at submission. We put six of them side by side in publisher AI disclosure policies, compared.

Everything above is publisher-level, which is the floor rather than the answer. Your answer lives in narrower documents, and they are worth reading in this order.

If the answer comes back as "we do not have a number", that is an answer and not a brush-off. It tells you the trigger is not a threshold you can aim at, so the useful preparation is a record of how the manuscript was written, not a target figure.

Then keep the two questions apart, because they answer to different documents. A rewrite operates on wording. A disclosure rule asks what happened while the manuscript was being written, and nothing you do to the wording later travels back to that. We wrote our own position on it in does running your document through a rewriting tool change what you have to declare?, and Elsevier's "substantive changes to sentence structure or organization" line above is the version that matters at submission.

Where a report already exists and it marks particular paragraphs, that is the narrow situation HumanPen is built for. You upload the manuscript and the report together, it works only on the marked passages, and the rest of the file comes back as it went in. Eligible results can continue lowering AI for free.

What we are not going to give you is a target figure. None of the sources above sets one, inventing one is not ours to do, and a rewriting vendor announcing what counts as acceptable would be exactly the wrong party to ask.

Does every journal run an AI check on submissions? Do not assume it in either direction. Wiley publicly describes routine screening for duplicated text, while iThenticate sells AI writing detection as an add-on. Even where it is licensed and enabled, a submission still has to meet the report's file, language and qualifying-word requirements. Ask the journal or report owner which checks apply to this submission.

Is there an AI percentage limit for journal submissions? None of the seven named policy pages in our review supplied a universal limit. That does not rule out a title-specific instruction, so ask which document a quoted figure came from. The most-quoted number in this area is a display rule rather than a journal limit, which we unpacked in is 20% AI too high? why there is no absolute threshold.

My similarity report has no AI section. Does that mean the paper was not checked for AI? It means only that this report does not display an AI result. Possible reasons include licensing or enablement, an unsupported file type or language, qualifying text outside the documented word range, or the feature being disabled at submission. Ask the report owner which state applied rather than diagnosing it from the missing section alone.

Do I have to declare AI use if I rewrote everything myself afterwards? Read the exemption wording rather than reasoning from how much you changed. Elsevier's page exempts basic grammar, spelling and punctuation checks and says substantive changes to sentence structure or organisation should be disclosed. Other publishers word the same boundary differently, which is why the comparison page exists.

An editor asked about AI use in my paper. Is that a detection result? Not necessarily, and it is worth establishing before you answer. Ask what prompted the question and which report, if any, it came from. A question can come from a reviewer's impression, from a similarity finding, or from a portal declaration that did not match the manuscript, and those need different replies. A checklist for the pass before submission is in final manuscript check: AI disclosure, word count, citations and DOCX integrity.