Two readers, asking two different questions
An instructor marking a coursework paper and an editor handling a submission want to know different things. The instructor may be shown a percentage and has to decide what it means for one assignment under one course policy. The editor works to a publisher’s written rules. Those rules ask whether the authors said what they used and whether they stand behind every sentence. Who, or what, typed the sentences is a smaller question.
That changes how you prepare. For a coursework paper, the risks are what a detector is known to get wrong and whether you can show your drafts. For a manuscript, the risk named in nearly every policy below is failing to disclose. ICMJE says nondisclosure “may be construed as misconduct in some circumstances”.
In many schools a “research paper” is a course assignment, not a journal article. If that is your paper, skip to the Turnitin section. If you are submitting to a journal, the policy and screening tables are the parts to keep.
What 7 publisher and editorial policies say
We read each policy on 26 September 2026, on its owner’s site. COPE’s site blocks automated requests, so its text was read through a text-only reader proxy; the link goes to COPE’s own page. Words in quotation marks are the policy’s own wording, not ours.
| Body | Can AI be an author? | Disclosure, and where | Source |
|---|---|---|---|
| ICMJE Recommendations (medical journals that follow them) | No. Chatbots “should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work”. | Yes. Describe the use “in both the cover letter and the submitted work”. Separately, section II.A says “Use of AI for writing assistance should be reported in the acknowledgment section”. | ICMJE, section V.A, read 26 September 2026 |
| COPE position statement, Authorship and AI tools | No. AI tools “cannot meet the requirements for authorship as they cannot take responsibility for the submitted work”. | Yes, in “the Materials and Methods (or similar section)”, saying which tool was used and how. | COPE position, read 26 September 2026 |
| Springer Nature (Nature Portfolio editorial policy) | “Assigning authorship or accountability to AI systems or tools” sits in the band marked “Not permitted”. | Authors should “Disclose AI use for trust and transparency”. The live page names no section for it. Language polishing is in the permitted band. | Nature Portfolio AI policy, read 26 September 2026 |
| Elsevier (generative AI policy for journals) | No. Authors “should not list AI tools as an author or co-author, nor cite AI tools as an author”. | Yes, in “a separate AI declaration statement” at submission, which “will appear in the published article”. Basic grammar, spelling and punctuation checks are exempt. | Elsevier policy, read 26 September 2026 |
| Wiley (research integrity guidelines) | No. These tools “cannot fulfil the role of an author and must not be listed as one”. | Yes, when submitting. If it is missing, it “may be requested during peer review or after publication”. | Wiley guidelines, read 26 September 2026 |
| Taylor & Francis (AI policy) | No. Generative AI tools “must not be listed as an author”. | Yes: the tool’s full name and version number, how it was used and why. Articles need “a specific statement of AI usage disclosure”. The live page does not name a section. | Taylor & Francis policy, read 26 September 2026 |
| IEEE (submission and peer review policies) | Not stated on the submission-policy page we read. | AI-generated content goes “in the acknowledgments section”, naming the system and the sections it touched. For editing and grammar, disclosure is “not required, but recommended”. | IEEE Author Center, read 26 September 2026 |
Six of the seven rows bar AI from authorship, and ICMJE and COPE give the reason: a tool cannot take responsibility for the work. Every policy here that addresses accountability, which is all of them except IEEE’s submission-policy page, puts it on the humans. COPE says authors “are fully responsible for the content of their manuscript, even those parts produced by an AI tool”. Taylor & Francis: “Authors are accountable for the originality, validity, and integrity of the content of their submissions”. Wiley: “Authors remain fully accountable for their submissions and any tools or sources they use”.
They disagree about where the statement goes. ICMJE wants it in the cover letter and the manuscript, with writing help in the acknowledgments. COPE points to the methods section. Elsevier wants a separate declaration that is printed with the article. IEEE uses the acknowledgments. Follow your venue’s own instructions for authors, because a correct statement in the wrong place still has to be moved.
On language help, ICMJE’s guidance to editors is unusually direct. “Many forms of use can be very helpful, especially for those authoring content not in their primary language.” It treats spell-checking and grammar correction as “generally viewed as acceptable”. Content generation, by contrast, “must be considered in the context of the work”. The ICMJE Recommendations index carries the note “Updated January 2026”.
Nature Portfolio’s policy is now a green/amber/red risk framework, which is why its row reports the band each use falls in.
Science (AAAS) is missing on purpose. On 26 September 2026 its editorial-policies page returned a 403 error and then a CAPTCHA, and we do not fill a row from second-hand summaries.
What publishers actually run on submissions
Less than the word “detector” suggests. Everything below comes from the operators’ own announcements or policy pages, not from independent testing. These are the named tools we found in primary sources:
| Tool | Run by, and when | Documented to check | What a flag leads to |
|---|---|---|---|
| Geppetto | Springer Nature, announced 12 June 2024. Springer Nature announced its donation to the STM Integrity Hub on 30 April 2025. | It splits a paper into sections and scores each one for “the probability that the text in them has been AI generated”. Its stated target is fake papers: the 2024 release calls AI-generated content “a classic indication of paper mill activity”, and the 2025 release says it detects “AI-generated nonsense text”. | High scores start “a human check by Springer Nature staff”. |
| SnappShot | Springer Nature, announced 12 June 2024. | Duplication in gel and blot images. It does not read text. | A signal “as to whether human assessment of the images is needed”. It does not decide. |
| Irrelevant-reference checker | Springer Nature, launched 7 April 2025 across almost all its journals and books. | References that do not relate to the manuscript. | The Research Integrity Group will “manually check the manuscript and decide whether the submission should be withdrawn”. |
| STM Integrity Hub | STM, the publishers’ trade association, for member publishers. | A “cloud-based environment for publishers to check submitted articles for research integrity issues”, including paper-mill patterns. Its releases list Springer Nature’s AI-text tool and Clear Skies’ Papermill Alarm. | Not described on the Hub page we read. |
| Elsevier's research integrity checks | Elsevier | AI tools that support “technical checks of submitted papers” and “research integrity checks”, plus duplicate-submission detection. It names no AI-text detector. | Not described on the policy page. |
| iThenticate v2 AI writing detection | Turnitin, through Crossref Similarity Check. Crossref post, 1 November 2023. | Turnitin’s AI writing detection, offered to iThenticate v2 users as a free preview “from Wednesday 1 November until the end of December 2023”. In 2023 it required English text of 300 to 15,000 words. | A percentage in the similarity report, not a decision. Crossref warned of “a higher incidence of false positives when the percentage is between 1 and 20”, marked with an asterisk. Turnitin’s own stated target (read 26 September 2026) is a false positive rate “under 1% for documents with over 20% of AI writing”. That is a vendor figure. |
The publishers’ own tools come with no error rate. We read three Springer Nature releases, the STM Integrity Hub page and Elsevier’s policy. None gives a false-positive or accuracy figure for Geppetto, SnappShot, the reference checker, the Hub or Elsevier’s checks. Accuracy comes up only in qualitative terms. Springer Nature says the reference checker went through testing “to ensure it provides a precise and reliable assessment”. In the donation release, Joris van Rossum of STM Solutions says wider use “will help to further train the tool and improve its accuracy”. Turnitin is the exception in the table: it states a target for its own detector, and that figure comes from the vendor, not from outside testing. How often a legitimate manuscript gets flagged by a publisher’s tool is not public, and we will not guess.
A person makes the decision. The three Springer Nature tools all send their flags to staff: Geppetto to a human check, SnappShot to image assessment, the reference checker to the Research Integrity Group. A flag starts a review. Nothing Springer Nature has published says a flag rejects a paper on its own.
The iThenticate row is the one to date carefully. Crossref’s post says Turnitin’s AI writing detection would appear as a free preview for iThenticate v2 users from 1 November “until the end of December 2023”, after which Turnitin planned a paid AI and paraphrase detection bundle charged per uploaded manuscript. It adds that the feature would be on by default and that administrators could switch it off. That is a 2023 statement. We could not confirm current availability or pricing, or whether journals have it switched on in 2026, so treat it as possible, not routine.
What research on scientific prose shows
Three peer-reviewed studies look at scientific writing specifically. They answer different questions, and mixing them up is how most summaries go wrong.
| Study | What was tested | Result as published | Limit |
|---|---|---|---|
| Gao et al., npj Digital Medicine, 2023 | 50 real and 50 ChatGPT-generated medical abstracts | GPT-2 Output Detector AUROC 0.94. Blinded reviewers spotted 68% of generated abstracts and marked 14% of real ones as generated. | Abstracts only, 100 samples |
| Desaire et al., Cell Reports Physical Science, 2023 | A 20-feature classifier built for Perspectives articles in Science | Over 99% accuracy. It misassigned one of 372 full documents; the GPT-2 Output Detector misassigned 20. | Built on one genre of one journal |
| Liang et al., Nature Human Behaviour, 2025 | 1,121,912 papers on arXiv, bioRxiv and Nature portfolio journals, Jan 2020 to Sep 2024 | Estimated LLM-modified content of up to 22% in computer science and up to 9% in mathematics and Nature portfolio papers. | An estimate across a population, not a verdict on any one paper |
Gao’s study used the kind of text a journal actually receives, which makes it the most relevant here. The detector separated real abstracts from generated ones well. The blinded human reviewers did worse: they “incorrectly identified 32% of generated abstracts as being real, and 14% of original abstracts as being generated”. Careful human readers were wrong about real work roughly one time in seven.
The same study found something anyone relying on a plagiarism checker should know. iThenticate gave the original, already-published abstracts a median similarity index of 100, and the generated ones a median of 27. Similarity software looks for text that already exists somewhere. Freshly generated text mostly does not, so a low similarity score is not evidence that AI was not used.
Desaire’s team took the opposite approach and built a narrow detector for a single genre. It worked very well on that genre. The stylistic differences it relies on, such as scientists’ heavier use of “however”, “but” and “although”, were observed only in Perspectives articles, and the authors write that “Additional studies will be needed to assess the generalizability of these observations to the full diversity of human scientific writing”. So the paper shows specialised detection can work on a narrow target. It says nothing about the general tool your university or publisher runs.
The Liang figures measure how word choice across whole fields has shifted since 2020, so they cannot be used against a single manuscript. The authors make no such claim. This is a different study from the 2023 Liang et al. paper on non-native English writers cited elsewhere on this site. The figures above are from the peer-reviewed 2025 version.
For general-purpose detectors, Weber-Wulff and colleagues tested a set of widely available tools and concluded that the tools “are neither accurate nor reliable”: they “all scored below 80% of accuracy”. Put together, the evidence says a detector tuned to one kind of scientific writing can do very well on that kind, and the general tools most people can reach have not been shown to. The wider evidence, with the studies that disagree, is on our accuracy page.
A student research paper checked by Turnitin
Turnitin’s file requirements for an AI writing report (read 26 September 2026) set a window of at least 300 words of long-form prose and no more than 30,000. The file must be .docx, .pdf, .txt or .rtf, under 100 MB, and in English, Spanish, Japanese or Arabic. The limits are stated per file, so check your paper against them before you submit. A thesis or final-year project handed in chapter by chapter has its own page: checking a thesis for AI.
Research papers carry more quoted material and longer reference lists than most essays. Whether those push a score up is covered in do AI detectors flag quotes and citations? Read it before any conversation with your instructor, not after. If you quoted a chatbot directly, cite it properly: how to cite ChatGPT gives APA’s template. A citation credits a quoted output; a disclosure statement describes how the tool shaped the work. A paper can need both.
The evidence that settles a dispute is the same for a course paper as for a thesis: dated drafts, your notes and sources, and the data you analysed. Keep them for as long as the grade can be questioned. If your course allows AI for language help, ICMJE’s line about writers working outside their first language is a fair point to raise with your instructor. Your course policy is still the one that applies.
If a journal questions your manuscript
A query from an editor about AI use is a question. The policies above tell editors to ask it: ICMJE says “editors should query authors and authors should disclose if and how LLM and AI were used”. Answer it the way you would answer a reviewer.
- Answer the question you were asked, in writing. Name the tools, what each was used for and which sections it touched. Taylor & Francis asks for the tool, its version, how it was used and why; IEEE asks for the system and the sections it touched.
- If your submission lacked a statement it needed, supply one now. Wiley’s policy expects this and says disclosure may be requested during peer review. ICMJE warns that nondisclosure “may require corrective action”. That is a reason to correct the record rather than defend the gap.
- Offer the record of the work. Dated drafts, version history, lab notebooks, analysis code and raw data. An integrity team is weighing the whole manuscript, not the score for one sentence.
- Ask which sections were flagged. Springer Nature says Geppetto scores a paper section by section. A concern about one section is easier to answer with that section’s drafts.
- Keep responsibility in view. Saying a tool wrote the sentence is no defence under the policies in the table that address accountability, which is all of them except IEEE’s page. Under those, the authors answer for the whole text.
COPE has published advice on exactly this situation. In case 26-11, “Suspected AI-generated manuscripts” (2026, read 26 September 2026 through the same text-only proxy), a reviewer had run a paper through an AI detector. COPE’s Council said the editor’s first step should be “to contact the authors in a neutral tone, quoting from their author guidelines or the related COPE guidance”. It also said that “AI detection tools are not infallible” and that “The raised concerns must be validated via other means than the AI-detection tool”. Two further points cut against authors. The journal “may also consider rejection solely based on lack of disclosure as required by journal policy”, so a missing statement can matter on its own. And COPE noted that the reviewer who uploaded the paper to a detector may have breached confidentiality, which is worth knowing if a query cites a reviewer’s detector result.
COPE’s separate topic discussion on editors who suspect AI was blocked on every route we tried on 26 September 2026, so it is not cited here.
To write the statement itself, the AI disclosure generator produces a first draft from what you tell it about your use. Check the result against your venue’s instructions, which override any template, ours and Elsevier’s included.