← series index
# Session 4: Writing, Publishing & Integrity
*AI for Researchers — attendee handout · Landscape as of August 2026*
---
## At a glance
| | |
|---|---|
| **Session** | 4 of 7 — Writing, Publishing & Integrity |
| **You'll learn** | Where the line between AI-as-editor and AI-as-ghostwriter actually falls, what each major publisher requires and where, why AI-text detectors cannot be used to enforce any of it, and what the rules become when you are the reviewer. |
| **You'll practise** | Running a policy-compliant editing pass on a rough methods paragraph — suggestion prompting versus rewrite prompting — and drafting the disclosure statement it requires, in a real publisher's format. |
| **Prerequisite** | None. Session 1's CRIT template is recapped in two minutes. |
| **Tools referenced** | None endorsed. Every prompt here works in any general assistant or writing tool; the obligations attach to *what you did*, not to which tool you used. |
---
## The one-paragraph version
Almost every journal you will submit to now has an AI policy, and almost nobody complies with it. Across 5,114 journals and 5.2 million papers, **70% of journals had adopted an AI policy** [1]. Within the same study's 164,579-paper full-text sub-corpus, only **76 of the 75,172 papers published since 2023 (~0.1%, cumulative) disclosed AI use**; the *quarterly* disclosure rate rose from 0.01% in early 2023 to just **0.43% in 2025 Q1** — the quarter in which the study put the ratio of detected AI-content prevalence to disclosure at about **40:1** (three figures, three windows, one study — they are consistent, not contradictory) [1]. An independent lexical analysis of 7.3 million articles estimates **57% showed evidence of LLM influence by 2025**, up from 12% in 2023 [29]; set against the 0.43% disclosure rate, that is roughly two orders of magnitude of gap (57 ÷ 0.43 ≈ 130), across two different corpora and methods. Corroboration from a mandatory declaration field: **5.7%** of 25,114 manuscripts submitted to 49 BMJ Group journals disclosed it, 87% of those for improving their writing [3]. An independent excess-vocabulary estimate puts at least **13.5% of 2024 PubMed abstracts** through an LLM [4]. All of these usage estimates are group-level statistics, and none "can distinguish whether AI is used merely for linguistic polishing… or for substantive text generation" [1]. The gap is mostly not deception: more than 90% of 5,229 surveyed researchers think editing your own paper with AI is fine, but the rules for whether that needs declaring differ between publishers, and sometimes between two journals at the same publisher [7][26]. Detectors cannot close the gap and should not be trusted to try: the broadest independent test (2023) found them "neither accurate nor reliable" [8], seven 2023-era detectors false-flagged **61.3%** of human-written essays by non-native English speakers [9], merely polishing your own published prose (2024-era pipeline: chatgpt-4o-latest, scored by GPTZero) took it from **97–100% classified human** to **75–85% flagged AI** — while altering the meaning of 2–3 sentences per letter [10] — and an August-2026 re-examination found guideline-compliant light editing flagged **38–80%** while humanizer-evaded AI text was flagged **under 4%** [30]. The defensible position is the boring one: keep the content yours, diff every polished paragraph, and disclose in the format your target journal asks for.
---
## The policy quick-reference table
> **Check your target journal — this table is not the last word.** Publisher pages give you the *shape* of the requirement; the journal's own guide for authors gives you the *wording*, and they diverge. On one day in July 2026, two Elsevier journals asked for two different headings for the same declaration [26]. Taylor & Francis states plainly that "some journals may not allow use of Generative AI tools beyond language improvement" [20]. **Read your target journal's guide for authors the week you submit.** Every cell below is quoted from that body's own page, fetched 2026-07-28.
| Publisher / body | Category | Disclose what, and where | Exempt | Their own page |
|---|---|---|---|---|
| **Elsevier** (journals policy, updated June 2026) | Structured disclosure | New section before the references: "Declaration of generative AI and AI-assisted technologies in the manuscript preparation process", using their template sentence [17] | "Basic checks of grammar, spelling and punctuation" — but disclose "when an AI tool makes substantive changes to sentence structure or organization" [17] | [elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals](https://www.elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals) |
| **Springer Nature / Nature Portfolio** | Structured disclosure | "Use of an LLM should be properly documented in the Methods section (and if a Methods section is not available, in a suitable alternative part)" [18] | "AI assisted copy editing" — readability, style, grammar, spelling, punctuation, tone — **does not need to be declared** [18] | [nature.com/nature-portfolio/editorial-policies/ai](https://www.nature.com/nature-portfolio/editorial-policies/ai) |
| **Wiley** | Structured disclosure + at submission | Use to "substantially edit, develop, or translate" must be "described transparently and integrated into the manuscript", and disclosed at submission [19] | Tools used "solely for spelling, grammar, and general editing" — but the tool's terms must not let its provider "train" on your content [19] | [authorservices.wiley.com/ethics-guidelines](https://authorservices.wiley.com/ethics-guidelines/index.html) |
| **Taylor & Francis** | Permissive with transparency | A statement in the article giving "the full name of the tool used (with version number), how it was used and the reason for use" [20] | Not specified as an exemption; the journal may impose stricter limits [20] | [taylorandfrancis.com/our-policies/ai-policy](https://taylorandfrancis.com/our-policies/ai-policy/) |
| **Sage** | Structured disclosure (three named tiers) | "Generative" use — affecting methodology, analysis, results or conclusions — declared at submission, "within the methods or acknowledgements"; includes translation and literature-search help [21] | "Assistive" tools that "improve language, grammar, or structure" [21] | [us.sagepub.com/en-us/nam/chatgpt-and-generative-ai](https://us.sagepub.com/en-us/nam/chatgpt-and-generative-ai) |
| **Science / AAAS** | Restrictive, disclosure-gated | "In the cover letter and in the methods section or acknowledgments section"; for research use, the prompt and tool version [22] | None stated for writing use; AI images need explicit editor permission, and editors "may decline to move forward" on inappropriate use [22] | [science.org/content/page/science-journals-editorial-policies](https://www.science.org/content/page/science-journals-editorial-policies) |
| **JAMA Network** | Permissive with transparency | Acknowledgment (or Methods if part of the design): what was created, tool name, "version and extension numbers, and manufacturer", dates of use [24][25] | "Basic tools for checking grammar, spelling, references" [24] | [jamanetwork.com/journals/jama/pages/instructions-for-authors](https://jamanetwork.com/journals/jama/pages/instructions-for-authors) |
| **ICMJE** (adopted by most medical journals) | Structured disclosure, two places | "In both the cover letter and the submitted work": writing assistance in the acknowledgments, data/analysis/figures in the methods [14] | Not specified; authors must still be able to "assert that there is no plagiarism" in AI-produced text or images [14] | [icmje.org — Defining the Role of Authors and Contributors](https://www.icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html) |
**The one rule with no exceptions:** AI cannot be an author, anywhere. ICMJE: chatbots "cannot be responsible for the accuracy, integrity, and originality of the work" [14]. COPE: they "cannot take responsibility for the submitted work" [16]. Science: "artificial intelligence tools cannot be authors" [22]. In the first systematic audit (screened 2023), AI authorship was prohibited by 96% of publishers and 98% of journals that had any guidance [6].
**Enforcement is no longer hypothetical.** In May 2026, arXiv announced a one-year submission ban where there is "incontrovertible evidence that the authors did not check the results of LLM generation" — hallucinated references are named as exactly such evidence, and subsequent submissions must first be accepted by a reputable peer-reviewed venue [31]. Note the trigger: unverified LLM output, not LLM use. The same verification habits this handout drills are what keep you on the right side of it.
---
## Disclosure-statement templates
Fill the brackets truthfully and change nothing else. If your journal's guide for authors gives different wording, **their wording wins**.
**Template 1 — Elsevier format** (new section at the end of the manuscript, before the references) [17]
```
Declaration of generative AI and AI-assisted technologies in the manuscript
preparation process
During the preparation of this work, the author(s) used [NAME OF TOOL /
SERVICE] in order to [REASON]. After using this tool/service, the author(s)
reviewed and edited the content as needed and take(s) full responsibility
for the content of the published article.
```
*When to use:* any Elsevier journal, and as a serviceable default elsewhere. *Watch out for:* journal-level variants — some Elsevier journals still head this section "…in the writing process", end it "…the content of the publication", and drop the comma after "work" [26]. Copy the wording from your journal's own guide for authors, not from here.
**Template 2 — JAMA Network format** (Acknowledgment section; Methods instead if the AI was part of the research design) [24][25]
```
Use of AI: The author(s) used [TOOL NAME], version [VERSION AND EXTENSION
NUMBERS], from [MANUFACTURER], on [DATE(S)], to [BRIEF DESCRIPTION OF WHAT
IT DID AND TO WHICH PORTIONS OF THE MANUSCRIPT]. The author(s) reviewed,
edited, and take responsibility for the integrity of the content generated.
```
*When to use:* JAMA Network journals, and any journal that asks for tool versions. *Watch out for:* this is the most demanding element list in the table — record the version and dates **while you work**, not from memory at submission.
**Template 3 — Taylor & Francis / general "name, version, use, reason" format** (a statement within the article) [20]
```
Declaration of AI use: [TOOL NAME] (version [VERSION]) was used to [HOW IT
WAS USED] because [REASON FOR USE]. The author(s) reviewed all output,
verified all content and references, and take full responsibility for the
final text.
```
*When to use:* Taylor & Francis, and journals that specify content but not a template. *Watch out for:* "how" and "why" are separate requirements in the T&F wording — answer both.
**For ICMJE-following journals,** put the writing-assistance sentence in the acknowledgments **and** describe the use in your cover letter; if AI touched data, analysis or figures, that description belongs in the methods instead [14].
---
## Two prompt templates
**A. Policy-compliant editing pass (suggestion mode)** — keeps the content, and the accountability, yours.
```
Context: This is a [section] I wrote myself. I am submitting to a journal
that permits AI-assisted language editing provided I disclose it.
Role: Act as a copy-editor for an academic [field] manuscript. You are not
a co-author and not a reviewer.
Instructions and constraints:
- Return a NUMBERED LIST OF SUGGESTIONS, not a rewritten text.
- For each: exact original wording, proposed wording, one-line reason.
- Change no number, statistic, symbol, reference or identifier.
- Change no word expressing certainty, hedging, or interpretation.
- Add no content, citation or claim not already present.
- If something is unclear or looks wrong, ask me a QUESTION. Do not fix it.
Task: List the changes you would make to the text below.
[paste your text]
```
*When to use:* any language pass on your own writing. *Watch out for:* accept the suggestions one at a time, by hand. Current-generation assistants usually respect the suggestion-only constraint — but "usually" is not a compliance strategy, so re-prompt if it returns a rewrite anyway.
**B. Disclosure statement** — fills a required template without embellishing it.
```
Context: I used an AI assistant to [exactly what you did]. I reviewed and
accepted the changes individually. No content, numbers, citations or
interpretations were generated by the tool.
Instructions and constraints:
- Use EXACTLY this template and change nothing outside the brackets:
[paste your target journal's required template]
- [REASON] must be factual and under 20 words. Do not overstate or
understate.
- Add no sentences, hedges, apologies or justifications.
Task: Produce the declaration.
```
*When to use:* immediately after any disclosable use, while you still remember the version and dates. *Watch out for:* tools like to "improve" the publisher's template sentence. Diff it against the original wording before you paste it in.
---
## Before you submit — checklist
- [ ] **1. Open the target journal's guide for authors and read the AI paragraph.** Not the publisher's page — the journal's [20][26]. Ninety seconds.
- [ ] **2. Classify what you actually did.** Basic grammar/spelling? Usually exempt. Restructured sentences, translated, summarised sources, generated any text, drafted code, or made a figure? Disclosable at most publishers [17][18][19][21].
- [ ] **3. Diff every AI-touched paragraph against your original.** Polishing altered the meaning of 2–3 sentences per letter in published, peer-reviewed text (measured on 2024-era polishing; the habit, not the number, is the point) [10].
- [ ] **4. Re-verify every number, hedge, symbol and citation** in the touched text. Grounded or not, tools misdescribe sources.
- [ ] **5. Confirm no AI is listed as an author or as a cited author.** Universal, no exceptions [14][16][22].
- [ ] **6. Write the disclosure in the required format and the required place** — acknowledgments, methods, a named end-section, the cover letter, or the submission form. Several publishers want it in two places [14][19].
- [ ] **7. Record tool name, version and dates of use** in your project notes. JAMA already requires all three [25].
- [ ] **8. Check whether your funder has its own rule.** If a proposal is involved, NSF asks proposers to "indicate in the project description the extent to which, if any, generative AI technology was used" [28].
- [ ] **9. If you are reviewing rather than submitting: upload nothing.** Elsevier, Springer Nature, Wiley, Taylor & Francis, Sage and Science all prohibit putting a manuscript into an AI tool [17][18][19][20][21][22]; NIH **prohibits** reviewers from using generative AI for critiques and treats uploading application content as a confidentiality violation, certified in a signed agreement [27]; NSF prohibits uploading any proposal content [28].
- [ ] **10. Do not run a detector on your own text and panic.** The broadest independent test (2023) found detection tools "neither accurate nor reliable" [8]; seven 2023-era detectors false-flagged 61.3% of human-written non-native English writing [9]; and in August 2026, guideline-compliant light editing was still flagged 38–80% while humanizer-evaded AI text slipped through at under 4% flagged [30]. A correct disclosure is the protection; a detector score is not evidence about anyone.
---
## Further reading
- He, Y., & Bu, Y. (2026). *Academic journals' AI policies fail to curb the surge in AI-assisted academic writing*. PNAS 123, e2526734123. https://www.pnas.org/doi/10.1073/pnas.2526734123 — the disclosure gap, measured; read it with the published critique (Wang, PNAS 123, e2616276123) that argues the causal claim outruns the data.
- Siler, K. (2026). *The diffusion of large language models in published academic articles*. PNAS 123(22), e2605754123. https://doi.org/10.1073/pnas.2605754123 — the upper-bound estimate: 57% of 7.3M articles lexically LLM-influenced by 2025. Read with Topaz & Bahl's PNAS letter (doi:10.1073/pnas.2620928123), which contests turning word frequencies into prevalence, and Siler's reply (doi:10.1073/pnas.2621834123).
- Liang, W., et al. (2023). *GPT detectors are biased against non-native English writers*. Patterns 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779 — a 2023 study of 2023 detectors, and still the eight pages to send anyone who proposes detector-based enforcement.
- Weber-Wulff, D., et al. (2023). *Testing of detection tools for AI-generated text*. International Journal for Educational Integrity 19, 26. https://doi.org/10.1007/s40979-023-00146-z — the broadest independent test of the 2023 cohort; the phrase to quote is "neither accurate nor reliable".
- Karr, J. A., Jr., et al. (2026). *Why AI detection fails for academic integrity*. arXiv:2608.11256 (preprint). https://arxiv.org/abs/2608.11256 — the August-2026 update: compliant light editing flagged 38–80%, humanizer-evaded text under 4%; detectors punish the compliant and miss the evaders.
- Wang, J., et al. (2025). *ChatGPT-polished writing boosts the risk of human-authored manuscripts being miscredited as AI-generated*. JAAD International 22, 23–25. https://doi.org/10.1016/j.jdin.2025.05.013 — two pages, two numbers (from a 2024-era polishing pipeline), and the reason to diff.
- Amano, T., et al. (2023). *The manifold costs of being a non-native English speaker in science*. PLOS Biology 21(7), e3002184. https://doi.org/10.1371/journal.pbio.3002184 — what language polishing is actually for.
- ICMJE. *Recommendations — Defining the Role of Authors and Contributors*. https://www.icmje.org/recommendations/ — short, current, and the standard most of the table above is descended from.
*Full annotated source list for this session — 31 sources with the verbatim policy wording behind every cell: see `sources.md`.*
---
*AI for Researchers · Session 4: Writing, Publishing & Integrity · Landscape as of August 2026*