AI for Researchers

Session 4: Writing, Publishing & Integrity

What publishers actually require, why detectors cannot enforce it, and how to disclose correctly

{{Presenter Name}}

Landscape as of August 2026

2-Minute Recap

The Series So Far

Session 1 — The Capability/Failure Map for Research Tasks: every task sits in Safe, Unreliable or Dangerous.
Session 1 — The 2026 Tool Landscape Taxonomy: Tier 1 — General chatbots · Tier 2 — Research-specific tools · Tier 3 — Deep-research agents.
CRIT — Context / Role and register / Instructions and constraints / Task, then iterate — is our prompting template.
Session 2 — The Literature-Discovery Tool Matrix: six citation-grounded tools, compared from their own documentation.
Session 3 — The Summary-Trust Triage: when an AI summary obliges you to reopen the paper.
You have found the papers and read them. Today you write — and someone else sets the rules.

Today (Session 4): the first session where the answer is not "it depends on your judgement" but "it depends on your target journal's own page".

Learning Objectives

By the End of This Session You Will Be Able To…

Explain the distinction between using AI as an editor versus a ghostwriter, and where each falls within discipline-appropriate norms.
Compare the publisher policy spectrum — from prohibition through structured disclosure to permissive-with-transparency — and locate a given journal on it.
Apply ICMJE-aligned rules to determine when and how AI use must be disclosed, given that AI can never be listed as an author.

Objectives 4–6 continue on the next slide.

Learning Objectives

…And You Will Also Be Able To

Evaluate why AI-generated-text detectors are unreliable and should not be relied on for policy compliance.
Run a policy-compliant AI-assisted editing pass on a paragraph of academic writing.
Draft a correct, publication-ready AI-use disclosure statement.

Section 01

Editor or Ghostwriter?

Every policy in this session turns on one question: did the tool improve how your ideas are expressed, or did it supply the ideas?

Literacy Foundation

One Spectrum, Not Two Boxes

Editing — you supply the content

Grammar, spelling, punctuation
Tightening a sentence you wrote
Translating your own draft
Reordering a paragraph you wrote

Ghostwriting — the tool supplies it

Drafting a section from a prompt
Writing the discussion "in the style of"
Generating a literature summary as text
Producing claims you did not verify

The spectrum is continuous; the policies are not. Most publishers draw their line somewhere in the middle — and they do not all draw it in the same place.

The Rules, In Their Words

One Publisher Already Names the Spectrum

"We distinguish various uses for AI and related technologies as: assistive (and no longer requiring disclosure), generative (requiring disclosure), and prohibitive." Sage, Artificial intelligence policy ^[21]

Assistive: "tools to improve language, grammar, or structure… do not require disclosure". ^[21]
Generative: content that "directly impacts the research methodology, analysis, results and/or conclusions" — disclose in methods or acknowledgements. ^[21]
Sage lists translation and literature-review assistance as disclosable — many researchers assume they are not. ^[21]

The Evidence — 1 of 2

Why "Just Polishing" Is Not a Small Thing

+50.6%more time writing a paper in English (moderate-proficiency countries) ^[12]

2.5×more likely to have had a paper rejected because of its English ^[12]

12.5×more likely to be asked to improve the English at revision ^[12]

Survey of 908 researchers across eight nationalities. ^[12]
38.1% of non-native speakers had a paper rejected over English, against 14.4% of native speakers. ^[12]
Around 30% of early-career non-native speakers often skip English-language conferences. ^[12]

The Evidence — 2 of 2

The Equity Argument — and Its Limits

Over one million abstracts, 2012–2024: writing complexity improved globally, "with non-native English-speaking countries showing notable advances". ^[13]
The authors credit better connectivity and "the adoption of AI-assisted writing tools" for the convergence. ^[13]
Publishers agree in principle: Taylor & Francis welcomes AI for "supporting authors to express content in a non-native language". ^[20]
Two honest limits: the convergence evidence is observational, not causal ^[13] — and polishing carries its own risk, which Section 04 quantifies. ^[10]

Section 02

The 2026 Policy Landscape

Eight rulebooks, one universal rule, and a table you should not trust for longer than a year.

The Rules, In Their Words

The One Rule Nobody Disagrees About

"Chatbots (such as ChatGPT) should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work." ICMJE ^[14]

"AI tools cannot meet the requirements for authorship as they cannot take responsibility for the submitted work." COPE position statement ^[16]

Science: "artificial intelligence tools cannot be authors" — nor may cited sources be AI-authored. ^[22]
Of publishers and journals that had guidance in 2023, AI authorship was prohibited in 96% and 98%. ^[6]
The corollary matters more: a human is accountable for every sentence, including the ones a tool wrote. ^[14]

Series Artefact

The Publisher AI-Policy Comparison Table (1 of 3)

Landscape as of August 2026. Every row is quoted from that publisher's own current policy page, fetched 2026-07-28 — see [17]–[19]. Row order is not a ranking.
Publisher	Policy category	What must be disclosed, and where	Exempt / prohibited
Elsevierjournals policy, updated June 2026	Structured disclosurenamed section, fixed template	New section before the references, headed "Declaration of generative AI and AI-assisted technologies in the manuscript preparation process", using their template sentence. ^[17]	Exempt: "Basic checks of grammar, spelling and punctuation". Prohibited: AI authorship. ^[17]
Springer NatureNature Portfolio editorial policy	Structured disclosureMethods section	"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]	Exempt: "AI assisted copy editing" (readability, style, grammar). Prohibited: generative AI images. ^[18]
Wileypublishing ethics guidelines	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]	Exempt: tools used "solely for spelling, grammar, and general editing". ^[19]

Series Artefact

The Publisher AI-Policy Comparison Table (2 of 3)

Landscape as of August 2026, fetched 2026-07-28 from each publisher's own page — see [20]–[23]. Category labels are ours, not the publishers' — except Sage's three tiers.
Publisher	Policy category	What must be disclosed, and where	Exempt / prohibited
Taylor & FrancisAI policy	Permissive with transparencyjournal may override	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]	Prohibited: "Text or code generation without rigorous revision"; synthetic data standing in for missing data. ^[20]
Sageartificial intelligence policy	Structured disclosurethree named tiers	"Generative" use — anything affecting methodology, analysis, results or conclusions — declared at submission, "within the methods or acknowledgements". ^[21]	Exempt: "assistive" language tools. Prohibited: AI in peer review or editorial work. ^[21]
Science / AAASeditorial policies	Restrictive, disclosure-gatedrelaxed Nov 2023	"In the cover letter and in the methods section or acknowledgments section"; prompt and tool version for research use. ^[22]^[23]	Prohibited without editor permission: AI-generated images and multimedia. ^[22]

Series Artefact

The Publisher AI-Policy Comparison Table (3 of 3)

Landscape as of August 2026, fetched 2026-07-28 — see [14], [24], [25]. The handout carries all eight rows in one table, each with a link to the body's own page.
Publisher / body	Policy category	What must be disclosed, and where	Exempt / prohibited
JAMA Networkinstructions for authors	Permissive with transparencymost prescriptive element list	Acknowledgment (or Methods, if part of the design): content created, tool name, "version and extension numbers, and manufacturer", dates. ^[24]^[25]	Exempt: "basic tools for checking grammar, spelling, references". Prohibited: AI-drafted Letters to the Editor. ^[24]
ICMJEthe standard most medical journals adopt	Structured disclosuretwo places, by use case	"In both the cover letter and the submitted work": writing assistance in the acknowledgments, data, analysis or figures in the methods. ^[14]	No AI authorship; authors must be able to "assert that there is no plagiarism", including in AI-produced text and images. ^[14]

Critical Literacy

Three Categories — and a Moving Target

In a January 2025 classification of 5,114 journals: 3,556 required disclosure, only 27 prohibited AI outright, and 2 had an open policy. ^[1]^[2]
96.8% of them allow writing and editing support; 62.9% permit language and grammar checking. ^[1]
So the modal 2026 rule is not "don't" — it is "you may, and you must say so".
These pages change. Science moved from near-ban to disclosure in November 2023 ^[23], later relaxing the reviewer clause ^[22]; Elsevier renamed its declaration in June 2026 ^[17].
Enforcement has begun: in May 2026, arXiv announced a one-year submission ban on "incontrovertible evidence" of unchecked LLM output — hallucinated references named as exactly that evidence. ^[31]

Date every policy claim you make — including the ones on this slide, which are true as of 28 July 2026.

Critical Literacy

The Publisher Is Not the Last Word

On one day, two Elsevier journals asked for two different section headings and two different closing clauses for the same declaration. ^[26]
Taylor & Francis says it plainly: "some journals may not allow use of Generative AI tools beyond language improvement". ^[20]
Heterogeneity "persist[ed] in some instances among affiliated publishers and journals" in the first systematic audit (screened 2023). ^[6]
The rule for you: the publisher page tells you the shape of the requirement; the journal's own guide for authors tells you the wording. Check it the week you submit.

Section 03

The Disclosure Gap

Most journals now ask. Almost nobody answers. The size of that gap has finally been measured.

The Evidence

5,114 Journals, 5.2 Million Papers, 76 Disclosures

70%of journals had adopted an AI policy, "primarily requiring disclosure" ^[1]

~0.1%cumulative disclosure since 2023 — 76 of 75,172 post-2023 papers in the study's 164,579-paper full-text sub-corpus ^[1]

Use of AI writing tools rose "with no significant difference between journals with or without policies", fastest in non-English-speaking countries and the physical sciences. ^[1]
The quarterly disclosure rate, same sub-corpus, rose from 0.01% in early 2023 to 0.43% in 2025 Q1 — from almost nothing to nearly nothing. ^[1]
Also in 2025 Q1: the ratio of detected AI-content prevalence to disclosure was about 40:1 — "for every 40 papers showing statistical evidence of AI usage, only one formally disclosed it". ^[1]
Three figures, three windows, one study: ~0.1% is cumulative since 2023; 0.43% is the 2025-Q1 rate; 40:1 is the 2025-Q1 use-to-disclosure ratio. Consistent, not contradictory. ^[1]

Critical Literacy

How Far Can You Push That Number?

5.7% — of 25,114 manuscripts submitted to 49 BMJ Group journals through a mandatory declaration question; 87% of those reported using AI to improve their writing. ^[3]
0.2% — of 7,251 JAMA Network Open articles named an LLM in the acknowledgments. ^[5]
≥13.5% — of 2024 PubMed abstracts, by an independent excess-vocabulary estimate that the 40:1 study also ran as one of its own robustness checks. ^[4]^[1]
57% — of 7.3 million articles (Elsevier, Frontiers, MDPI, PLoS) showed lexical evidence of LLM influence by 2025, up from 12% in 2023. Set against the 0.43% 2025-Q1 disclosure rate: 57 ÷ 0.43 ≈ 130 — two different corpora and methods, but two orders of magnitude of gap either way. ^[29]^[1]
What none of them proves: the usage estimates are group-level and "cannot distinguish whether AI is used merely for linguistic polishing… or for substantive text generation" ^[1]; lexical prevalence estimates are contested in a published letter ^[29]; and the critique of the 40:1 study argues policy snapshots "cannot establish" that policies failed — a limit the authors concede. ^[2]

The Evidence — Self-Reported Survey Data

It Is Not Mostly Deception. It Is Confusion.

Of 5,229 researchers surveyed, more than 90% said using AI to edit or translate your own paper is acceptable; 65% said generating text is. ^[7]
But only ~28% had actually used AI to edit a paper — and of those, 18% did not disclose it against 10% who did. ^[7]
The BMJ Group authors reach the same conclusion: authors "may be uncertain about what AI use requires disclosure or may be hesitant to declare it". ^[3]
Small wonder: "improve the language" is exempt at Springer Nature and Wiley, but disclosable at Sage once it changes structure. ^[18]^[19]^[21]

Survey figures are self-reported data from a publisher-run poll (Nature, March 2025); the article itself flags response bias and under-representation of authors in China. ^[7]

Section 04

Why Detectors Cannot Close the Gap

If you take one operational fact from this session, take this one: AI-text detectors are not evidence, and they are not neutral.

The Evidence — 1 of 3

"Neither Accurate Nor Reliable"

"the available detection tools are neither accurate nor reliable and have a main bias towards classifying the output as human-written rather than detecting AI-generated text. Furthermore, content obfuscation techniques significantly worsen the performance of tools." Weber-Wulff et al. (2023), 12 public tools plus Turnitin and PlagiarismCheck ^[8]

The failure is two-sided: they miss AI text, and they misfire on human text. ^[8]
That was 2023's verdict on 2023's tools. An August-2026 re-examination reads the same: guideline-compliant light editing flagged 38–80%, humanizer-evaded AI text flagged <4%. ^[30]
"Honest AI-editing results in a higher sanction risk than humanizer-assisted evasion." Evaders can opt out; honest writers cannot. ^[30]

The Evidence — 2 of 3

And the Errors Are Not Evenly Distributed

61.3%average false-positive rate on human-written TOEFL essays, across seven detectors (2023) ^[9]

19.8%of those human essays flagged as AI by all seven detectors ^[9]

11.6%false-positive rate after the same essays were given richer vocabulary ^[9]

The same detectors classified US eighth-grade essays with near-perfect accuracy. ^[9]
Mechanism, in that 2023 cohort: those detectors keyed on low text perplexity, so they penalised "individuals with limited linguistic proficiency". ^[9]
The skew outlived the cohort: in the 2026 re-test, unmodified human abstracts were still flagged at 9–15%, with non-STEM rates far above STEM. ^[30]

The Evidence — 3 of 3

Polishing Your Own Paper Flags Your Own Paper

"While 97% to 100% of the original letters were classified as human-generated, 75% to 85% of the polished versions were flagged as AI-generated, with 15% to 25% labeled at high confidence." Wang et al. (2025), published JAAD Research Letters polished for readability only — 2024-era pipeline: chatgpt-4o-latest, scored by GPTZero ^[10]

The instruction was simply to "enhance the readability and flow" while preserving "original meaning and intent". ^[10]
Second finding, and the reason Demo A ends with a diff: polishing "revealed alterations in meaning for on average 2 to 3 sentences per letter". ^[10]

Practice

What Protects You Instead

Disclose in the required format. A correct declaration converts a suspicion into a documented fact. ^[17]^[24]
Keep the record: tool, version, dates, what you asked it to do — the elements JAMA already requires. ^[25]
Diff every polished paragraph against your original before accepting it — measured on 2024-era polishing, but meaning drift is a property of rewriting, not of one model version. ^[10]
If you are accused: "The main question… is whether the content is correct and clear, not which tool was used to polish the language." ^[11]

Section 05

The Other Side of the Desk

When you review a manuscript or a grant, the rules stop being about disclosure and start being about confidentiality.

The Rules, In Their Words

Never Paste a Manuscript You Are Reviewing

Elsevier: reviewers "should not upload a submitted manuscript or any part of it into an AI tool as this may violate the authors' confidentiality and proprietary rights". ^[17]
Springer Nature: "peer reviewers do not upload manuscripts into generative AI tools"; AI-supported evaluation must be declared in the report. ^[18]
Science: "Reviewers may not enter any part of the manuscript into an LLM" — but may use AI to revise their own writing if inputs are not retained, declared in the review. ^[22]
Taylor & Francis and Wiley: same prohibition; language help on your own text only, disclosed to the editor. ^[20]^[19]

ICMJE gives the reason: manuscripts are "privileged communications that are authors' private, confidential property". ^[15]

The Rules, In Their Words

Funders Go Further Than Publishers

"NIH prohibits NIH scientific peer reviewers from using natural language processors, large language models, or other generative Artificial Intelligence (AI) technologies for analyzing and formulating peer review critiques… uploading or sharing content or original concepts from an NIH grant application, contract proposal, or critique to online generative AI tools violates the NIH peer review confidentiality and integrity requirements." NIH Notice NOT-OD-23-149, released 23 June 2023 ^[27]

Reviewers now certify compliance in a modified confidentiality agreement; accessibility technologies may be excepted with prior notice. ^[27]
NSF: reviewers "prohibited from uploading any content from proposals"; proposers "encouraged to indicate in the project description" any generative-AI use. ^[28]

Demo Preview

CRIT, Applied to a Compliant Editing Pass

C — Context: "This is my own methods paragraph for a journal that permits language editing with disclosure."
R — Role and register: "Act as a copy-editor for an academic methods section. Do not act as a co-author."
I — Instructions and constraints: "Return numbered suggestions with reasons. Change no numbers, hedges or claims. Do not add content."
T — Task: "List the changes you would make and why." Then iterate — and accept the edits yourself, one at a time.

Demo A runs this against a rewrite prompt on the same rough paragraph and diffs both. Demo B writes the disclosure statement Elsevier requires for it. ^[17]

References

References (1–15)

He, Y., & Bu, Y. (2026). Academic journals' AI policies fail to curb the surge in AI-assisted academic writing. PNAS 123, e2526734123. pnas.org/doi/10.1073/pnas.2526734123 — accessed 2026-07-28
Wang, Y. (2026). Policy snapshots cannot establish journal AI-policy failure. PNAS 123(27), e2616276123. pnas.org/doi/10.1073/pnas.2616276123 — accessed 2026-07-28
AlFayyad, I., et al. (2025). Authors self-disclosed use of artificial intelligence in research submissions to 49 biomedical journals. medRxiv 2025.10.24.25338574. medrxiv.org/content/10.1101/2025.10.24.25338574v1 — accessed 2026-07-28
Kobak, D., et al. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances 11, eadt3813. doi.org/10.1126/sciadv.adt3813 — accessed 2026-07-28
Wolfrath, N., et al. (2026). Rising prevalence of detected AI-generated text in medical literature. arXiv:2603.19316. arxiv.org/abs/2603.19316 — accessed 2026-07-28
Ganjavi, C., et al. (2024). Publishers' and journals' instructions to authors on use of generative AI. BMJ 384, e077192. doi.org/10.1136/bmj-2023-077192 — accessed 2026-07-28
Kwon, D. (2025). Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–576. Self-reported survey data. doi.org/10.1038/d41586-025-01463-8 — accessed 2026-07-28
Weber-Wulff, D., et al. (2023). Testing of detection tools for AI-generated text. Int. J. Educational Integrity 19, 26. doi.org/10.1007/s40979-023-00146-z — accessed 2026-07-28
Liang, W., et al. (2023). GPT detectors are biased against non-native English writers. Patterns 4(7), 100779. doi.org/10.1016/j.patter.2023.100779 — accessed 2026-07-28
Wang, J., et al. (2025). ChatGPT-polished writing boosts the risk of human-authored manuscripts being miscredited as AI-generated. JAAD Int. 22, 23–25. doi.org/10.1016/j.jdin.2025.05.013 — accessed 2026-07-28
Koga, S. (2025). ChatGPT-polished scientific writing and AI detection: cohorts, baselines, fairness. JAAD Int. 24, 242–243. doi.org/10.1016/j.jdin.2025.10.017 — accessed 2026-07-28
Amano, T., et al. (2023). The manifold costs of being a non-native English speaker in science. PLOS Biology 21(7), e3002184. doi.org/10.1371/journal.pbio.3002184 — accessed 2026-07-28
Prakash, A., et al. (2025). Writing without borders: AI and cross-cultural convergence in academic writing quality. Humanit. Soc. Sci. Commun. 12, 1058. doi.org/10.1057/s41599-025-05484-6 — accessed 2026-07-28
ICMJE (2025). Recommendations — II.A.4 Artificial Intelligence (AI)-Assisted Technology. icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html — accessed 2026-07-28
ICMJE (2025). Recommendations — II.C Responsibilities in the Submission and Peer-Review Process. icmje.org/recommendations/browse/roles-and-responsibilities/responsibilities-in-the-submission-and-peer-peview-process.html — accessed 2026-07-28

References

References (16–31)

COPE Council (2023). Position statement: Authorship and AI tools. publicationethics.org/guidance/cope-position/authorship-and-ai-tools — accessed 2026-07-28
Elsevier (2026, updated June 2026). Generative AI policies for journals. elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals — accessed 2026-07-28
Springer Nature / Nature Portfolio (2026). Artificial Intelligence (AI) editorial policy. nature.com/nature-portfolio/editorial-policies/ai — accessed 2026-07-28
Wiley (2026). Best practice guidelines on research integrity and publishing ethics — Artificial Intelligence. authorservices.wiley.com/ethics-guidelines/index.html — accessed 2026-07-28
Taylor & Francis (2025). AI policy. taylorandfrancis.com/our-policies/ai-policy/ — accessed 2026-07-28
Sage (2026). Artificial intelligence policy. us.sagepub.com/en-us/nam/chatgpt-and-generative-ai — accessed 2026-07-28
AAAS (2026). Science journals: editorial policies — Artificial intelligence. science.org/content/page/science-journals-editorial-policies — accessed 2026-07-28
Thorp, H. H., & Vinson, V. (2023, 16 Nov). Change to policy on the use of generative AI and large language models. Science Editor's Blog. science.org/content/blog-post/change-policy-use-generative-ai-and-large-language-models — accessed 2026-07-28
JAMA Network (2026). Instructions for authors — Use of AI in publication and research. jamanetwork.com/journals/jama/pages/instructions-for-authors — accessed 2026-07-28
Flanagin, A., et al. (2024). Reporting use of AI in research and scholarly publication — JAMA Network guidance. JAMA 331(13), 1096–1098. doi.org/10.1001/jama.2024.3471 — accessed 2026-07-28
Elsevier journal guides for authors (2026), e.g. Research Policy; Sensors and Actuators Reports. sciencedirect.com/journal/research-policy/publish/guide-for-authors — accessed 2026-07-28
NIH (2023). NOT-OD-23-149: The use of generative AI technologies is prohibited for the NIH peer review process. grants.nih.gov/grants/guide/notice-files/NOT-OD-23-149.html — accessed 2026-07-28
NSF (2023). Notice to research community: use of generative AI technology in the NSF merit review process. nsf.gov/news/notice-to-the-research-community-on-ai — accessed 2026-07-28
Siler, K. (2026). The diffusion of large language models in published academic articles. PNAS 123(22), e2605754123. doi.org/10.1073/pnas.2605754123 — accessed 2026-09-02
Karr, J. A., Jr., et al. (2026). Why AI detection fails for academic integrity. arXiv:2608.11256. Preprint. arxiv.org/abs/2608.11256 — accessed 2026-09-02
TechCrunch (2026, 16 May). Research repository arXiv will ban authors for a year if they let AI do all the work. Secondary coverage of the arXiv CS announcement. techcrunch.com/2026/05/16/… — accessed 2026-09-02