← series index

# Sources — Session 4: Writing, Publishing & Integrity

*AI for Researchers · dated, annotated reference list · compiled 2026-07-28, revised 2026-09-02 · Landscape as of August 2026*

Entry format (one line per source, numbering matches the `[n]` footnote markers used in `slides.html` and `handout.md`):

`- [n] Author/Org (Year). *Title*. URL — accessed YYYY-MM-DD. [type: peer-reviewed | policy | primary-doc | news] — one-line note on what claim(s) it grounds.`

Allowed `type` values:
- `peer-reviewed` — peer-reviewed paper or arXiv/preprint literature
- `policy` — official publisher/funder/institutional policy page
- `primary-doc` — primary tool/vendor documentation (not a blog or listicle)
- `news` — recent news/benchmark article; informs tool discovery only, never the sole ground for a factual claim

**Policy-source rule for this session.** Session 4 makes claims about publisher and funder rules that attendees will act on. Every policy row in the deck's comparison table and in the handout links **the publisher's or funder's own current page**, fetched directly on 2026-07-28 — never secondary reporting, a library guide, or a vendor summary. Where a policy has changed recently, the change and its date are stated on the slide.

---

## References

### A. The disclosure gap — the load-bearing evidence

- [1] He, Y., & Bu, Y. (2026). *Academic journals' AI policies fail to curb the surge in AI-assisted academic writing*. Proceedings of the National Academy of Sciences 123, e2526734123. https://www.pnas.org/doi/10.1073/pnas.2526734123 (open preprint: https://arxiv.org/abs/2512.06705) — accessed 2026-07-28. [type: peer-reviewed] — **The single primary study behind both halves of the disclosure-gap statistic, and the reason the stat may appear on a slide at all.** Analysis of 5,114 journals and over 5.2 million papers. Quoted verbatim on the slide: "despite 70% of journals adopting AI policies (primarily requiring disclosure), researchers' use of AI writing tools has increased dramatically across disciplines, with no significant difference between journals with or without policies"; and "full-text analysis on 164 k scientific publications reveals a striking transparency gap: Of the 75 k papers published since 2023, only 76 (~0.1%) explicitly disclosed AI use." Also grounds the observation that growth is highest in "Non-English-speaking countries, physical sciences, and high-OA journals", the policy-category counts on the deck's policy-spectrum slide ("3,556 journals require disclosure, 1,529 journals do not mention policies, 27 journals strictly prohibit, and 2 journals have open policies"; "96.8% of journals allow authors to use AI for writing and editing, and 62.9% permit language and grammar checking"), and the underreporting ratio quoted on the deck: the disclosure rate rose "from 0.01% in early 2023 to 0.43% in the first quarter of 2025", at which point "the ratio was approximately 40:1. In other words, for every 40 papers showing statistical evidence of AI usage, only one formally disclosed it." **Scope note (verified against the full text, 2026-09-02): the paper's three headline figures have three different windows and denominators and are complementary, not contradictory — the deck and handout state each with its scope.** ~0.1% (76/75,172) is the *cumulative* disclosure rate for papers published since 2023, within the 164,579-paper full-text sub-corpus; 0.01% → 0.43% is the *quarterly* disclosure-rate trend inside that same sub-corpus (Fig. S6); 40:1 is the *2025-Q1* ratio of detected AI-content proportion to disclosure rate. Do not average or substitute them. **Method, stated correctly (corrected 2026-07-28 after re-reading the full text).** The disclosure count is a direct full-text search of 164,579 publications. The AI-usage estimate is *not* a commercial detector: the authors "employed maximum likelihood estimations (MLE) combined with multiple robustness checks to quantify the actual prevalence of AI-generated content" — a **group-level distributional estimate**, which the authors are explicit "provides probabilistic estimates at the group level rather than definitive judgments on individual manuscripts". ZeroGPT is one of *three* secondary robustness checks, alongside a dictionary-based keyword analysis and Kobak et al.'s excess-word analysis [4] — i.e. [4] is not an alternative to this paper's method, it is one of its corroborations. The limitation that matters for this deck is the authors' own, and it is the one the slides state: "we are unable to distinguish whether AI is used merely for linguistic polishing which is typically permissible under current policies, or for substantive text generation that may violate regulations."
- [2] Wang, Y. (2026). *Policy snapshots cannot establish journal AI-policy failure*. Proceedings of the National Academy of Sciences 123(27), e2616276123. https://www.pnas.org/doi/10.1073/pnas.2616276123 — accessed 2026-07-28. [type: peer-reviewed] — The published critique of [1], cited on the same slide so the deck does not overstate a contested causal claim. Quoted: the data "document diffusion and public disclosure but do not establish the headline inference that journal AI policies have 'failed to promote transparency or restrain AI adoption'", because the analysis uses 2025 policy snapshots rather than journal-specific adoption dates. Independently valuable for the policy-spectrum slide: in the January 2025 classification there were "only 27 strict-prohibition journals and 2 open-policy journals, compared with 3,556 disclosure-required journals, and 96.8% of disclosure-policy journals allow writing and editing support" — i.e. the modal 2026 policy is *permit-with-disclosure*, not prohibition (the same counts appear in [1] itself). He and Bu's reply (https://doi.org/10.1073/pnas.2617762123) is noted but not quoted. Fair to both sides, and stated on the slide: **He and Bu concede the central point in their own limitations section** — "there is a certain time lag from the introduction of a policy to the manifestation of its effects… Consequently, our results do not necessarily imply that the policies are ineffective". The disagreement is about how far the title's claim outruns the data, not about the measurements.
- [3] AlFayyad, I., Zeegers, M. P., Bouter, L., Macdonald, H., & Schroter, S. (2025). *Authors self-disclosed use of artificial intelligence in research submissions to 49 biomedical journals: A cross-sectional study*. medRxiv 2025.10.24.25338574. https://www.medrxiv.org/content/10.1101/2025.10.24.25338574v1 — accessed 2026-07-28. [type: peer-reviewed] — Corroborates [1] from the publisher's own submission system rather than from published full text, which removes the "they disclosed in the cover letter, not the paper" objection raised by [2]. 25,114 empirical research manuscripts submitted to 49 BMJ Group journals between 8 April and 6 November 2024, all passing through a *mandatory* AI-declaration question: "A total of 1,431 submissions (5.7%) disclosed the use of AI." Of those, 87% (1,248/1,431) reported using it "to improve the quality of their writing"; generative AI chatbots were the most common tool (812/1,431; 56.7%); translation was 7.5%. Authors' conclusion quoted on the slide: 5.7% "is substantially lower than proportions reported in surveys" (28%–76%), and "Authors may be uncertain about what AI use requires disclosure or may be hesitant to declare it." Preprint at time of access; flagged as such on the slide.
- [4] Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). *Delving into LLM-assisted writing in biomedical publications through excess vocabulary*. Science Advances 11, eadt3813. https://www.science.org/doi/10.1126/sciadv.adt3813 (open preprint: https://arxiv.org/abs/2406.07016) — accessed 2026-07-28. [type: peer-reviewed] — An independent, **corpus-level** estimate of how much AI-assisted writing is actually happening, arrived at by a different route from [1] and — as noted in [1]'s methods — reused there as one of that paper's three robustness checks. It is on the slide as convergent evidence, not as a substitute for [1]. Method: excess-word frequency analysis across more than 15 million PubMed abstracts, 2010–2024, modelled the way excess-mortality is estimated. Quoted: "at least 13.5% of 2024 abstracts were processed with LLMs", a lower bound that "differed across disciplines, countries, and journals, reaching 40% for some subcorpora". The named excess words ("delve, intricate, meticulously, realm, pivotal, showcasing") are used on the slide only as illustration, never as a detection method.
- [5] Wolfrath, N., Patel, S., Flitcroft, M., Banerjee, A., Somai, M., Crotty, B. H., & Kothari, A. N. (2026). *Rising Prevalence of Detected AI-Generated Text in Medical Literature: Longitudinal Analysis in Open Access Articles*. arXiv:2603.19316. https://arxiv.org/abs/2603.19316 — accessed 2026-07-28. [type: peer-reviewed] — Second corroboration of the disclosure side of the gap, in a single journal with a clear disclosure policy: 7,251 open-access JAMA Network Open articles, January 2022 – March 2025, with acknowledgment sections manually reviewed for named LLMs. "Of all articles, fifteen articles (0.2%) were found to disclose utilizing an LLM in the writing process." The companion figure — articles flagged as >10% AI-generated rising "from 0.0% in the first month analyzed (January 2022) to 11.3% in the final month (March 2025)" — is detector-derived (Originality.AI) and is therefore **not** quoted as a measurement on any slide; only the manually verified 0.2% disclosure figure is used. Preprint.
- [6] Ganjavi, C., Eppler, M. B., Pekcan, A., Biedermann, B., Abreu, A., Collins, G. S., Gill, I. S., & Cacciamani, G. E. (2024). *Publishers' and journals' instructions to authors on use of generative artificial intelligence in academic and scientific publishing: bibliometric analysis*. BMJ 384, e077192. https://doi.org/10.1136/bmj-2023-077192 (open access: https://pmc.ncbi.nlm.nih.gov/articles/PMC10828852/) — accessed 2026-07-28. [type: peer-reviewed] — Grounds the "the rules genuinely differ, and the difference is not random" slide and the handout's *check your target journal* caveat. Top 100 largest publishers and top 100 highest-ranked journals, screened May and October 2023: "Among the top 100 largest publishers, 24% provided guidance on the use of GAI… Among the top 100 highly ranked journals, 87% provided guidance on GAI. Of the publishers and journals with guidelines, the inclusion of GAI as an author was prohibited in 96% and 98%, respectively… When disclosing the use of GAI, 75% of publishers and 43% of journals included specific disclosure criteria." Conclusion quoted: heterogeneity persists "in some instances among affiliated publishers and journals", and "Lack of standardization places a burden on authors". Cited with its 2023 screening dates visible on the slide, because the policy landscape has moved since (see [1], [2]).
- [7] Kwon, D. (2025). *Is it OK for AI to write science papers? Nature survey shows researchers are split*. Nature 641, 574–576. https://doi.org/10.1038/d41586-025-01463-8 — accessed 2026-07-28. [type: news] — **First-party survey report from a major publisher; labelled on the slide as self-reported survey data, per the series source standard.** 5,229 respondents contacted in March 2025 via e-mails to recently published authors, Springer Nature's market-research panel, and the Nature Briefing; the article itself states the sample is subject to response bias and under-represents authors in China. Used for the *why* of the disclosure gap, not for prevalence: more than 90% of respondents think it acceptable to use generative AI to edit or translate one's own paper; 65% think generating text is ethically acceptable; but only around 28% said they had actually used AI to edit a paper — and of those who had, the reported split was 18% who did not disclose it versus 10% who did. The slide quotes the survey's own framing: researchers "generally feel it's acceptable to use AI chatbots to help to prepare manuscripts, relatively few report actually using AI for this purpose — and those who did often say they didn't disclose it."
- [29] Siler, K. (2026). *The diffusion of large language models in published academic articles*. Proceedings of the National Academy of Sciences 123(22), e2605754123. https://doi.org/10.1073/pnas.2605754123 — accessed 2026-09-02 (added 2026-09-02, hence the out-of-section number; PNAS blocks direct fetching — abstract verified verbatim via the Crossref API). [type: peer-reviewed] — The upper-bound usage estimate that makes the disclosure gap two orders of magnitude wide. Quoted from the abstract: full texts of "7.3 million journal articles" from four major publishers (Elsevier, Frontiers, MDPI, PLoS), 2020–2025, analysed via "a corpus of 228 focal words exhibiting sharp post-2022 frequency increases consistent with LLM output"; "By 2025, an estimated 57% of published articles exhibited evidence of LLM influence, up from 12% in 2023"; "LLM adoption in academic writing is pervasive but socially stratified." **Pairing rule used on the slide and in the handout:** the gap is presented as *our* juxtaposition of Siler's 57% (2025) with He & Bu's 0.43% 2025-Q1 disclosure rate [1] — 57 ÷ 0.43 ≈ 130, i.e. roughly two orders of magnitude — with the caveat that the two studies use different corpora and methods; no gap figure is attributed to Siler himself. Cited alongside the published methodological challenge: Topaz & Bahl's PNAS letter "Lexical change is not a calibrated measure of LLM prevalence or its determinants" (doi:10.1073/pnas.2620928123) contests converting word-frequency shifts into prevalence rates; Siler's reply is doi:10.1073/pnas.2621834123. The deck's "what none of them proves" bullet carries that contestation.

### B. Why AI-text detectors cannot enforce any of this

- [8] Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). *Testing of detection tools for AI-generated text*. International Journal for Educational Integrity 19, article 26. https://doi.org/10.1007/s40979-023-00146-z — accessed 2026-07-28. [type: peer-reviewed] — The broadest independent test of detection tools, and the source of the deck's flat statement that detectors are not fit for enforcement. 12 publicly available tools plus two commercial systems (Turnitin, PlagiarismCheck), tested on an original document set with machine translation and obfuscation conditions. Quoted verbatim on the slide: "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." The asymmetry matters for the argument: the tools miss AI text *and* misfire on human text. **Labeled on the deck as what it is — a 2023 test of the 2023 detector cohort — and paired on the same slide with the August-2026 re-examination [30], which reaches the same verdict on current commercial detectors.**
- [9] Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). *GPT detectors are biased against non-native English writers*. Patterns 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779 (full text: https://www.cell.com/patterns/fulltext/S2666-3899(23)00130-7) — accessed 2026-07-28. [type: peer-reviewed] — **The documented-bias study.** Seven widely used GPT detectors evaluated on 91 TOEFL essays written by non-native English speakers and 88 US eighth-grade essays. Quoted verbatim: the detectors "incorrectly labeled more than half of the TOEFL essays as 'AI-generated' (average false-positive rate: 61.3%)"; "All detectors unanimously identified 19.8% of the human-written TOEFL essays as AI authored, and at least one detector flagged 97.8% of TOEFL essays as AI generated"; while "the detectors accurately classified the US student essays". Mechanism quoted on the slide: detectors key on low text perplexity, so they "inadvertently penalis[e] individuals with limited linguistic proficiency" — and using ChatGPT to enrich the vocabulary of the same TOEFL essays dropped the average false-positive rate from 61.3% to 11.6%. Also grounds the equity link: ICLR papers with first authors from non-English-speaking countries showed lower perplexity than those of native-English-speaking counterparts. **Historicized on the deck (2026-09-02): the rates and the perplexity mechanism are attributed to the seven 2023-era detectors this study tested — whether current commercial detectors still key on perplexity has not been independently verified, so the deck does not claim it. What has been re-verified is the outcome: the 2026 re-test [30] still finds human text over-flagged (9–15% of unmodified recent abstracts, non-STEM far above STEM), so the equity lesson is carried by the 2023 measurement plus the 2026 skew, not by asserting an unchanged mechanism.**
- [10] Wang, J., Jairath, N., Shue, E., Perrotta, P. L., & Hu, G. (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 (open access: https://pmc.ncbi.nlm.nih.gov/articles/PMC12276437/) — accessed 2026-07-28. [type: peer-reviewed] — The study that makes the detector problem personal for the attendee who only polishes their own prose. Research Letters published in JAAD (2020 and 2024; 58 letters from authors in China, Korea and Japan, matched with US-affiliated letters) were polished by ChatGPT ("chatgpt-4o-latest") with the instruction to "enhance the readability and flow" while preserving "original meaning and intent", then scored by GPTZero. Quoted verbatim on the slide: "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"; similar results with Originality.AI. Second finding used in the demo script: a focused review of 10 letters flagged as 100% AI-generated after polishing "revealed alterations in meaning for on average 2 to 3 sentences per letter, emphasizing the necessity of careful human proofreading postpolishing." **Historicized on the deck (2026-09-02): the polishing model (chatgpt-4o-latest) and the detector (GPTZero) are both named as a 2024-era pipeline wherever the numbers appear. The flagging-risk half is re-grounded in the current generation by [30]; the meaning-drift half is taught as a property of rewriting rather than of one model version, which is why the diff practice survives the historicization.**
- [11] Koga, S. (2025). *ChatGPT-polished scientific writing and AI detection: Cohorts, baselines, fairness*. JAAD International 24, 242–243. https://doi.org/10.1016/j.jdin.2025.10.017 (open access: https://pmc.ncbi.nlm.nih.gov/articles/PMC12721137/) — accessed 2026-07-28. [type: peer-reviewed] — Peer-reviewed correspondence on [10], cited for the "what to do instead" slide and for methodological honesty about [10]'s cohort definition (US affiliation is an imperfect proxy for native-English status). Quoted: detector scores "can conflate long-standing differences across language backgrounds with recent changes in tool use, and… using such scores as a proxy for authorship or quality can disadvantage non-native authors. The main question, therefore, is whether the content is correct and clear, not which tool was used to polish the language." Also the source of the deck's recommended replacement for detection: "A better safeguard is rigorous peer review that checks claims, methods, and data… the process should also include strict verification that references exist and match the text."
- [30] Karr, J. A., Jr., Khvatskii, G., Hua, T., & Chawla, N. V. (2026). *Why AI Detection Fails for Academic Integrity*. arXiv:2608.11256v2, 2026-08-20. https://arxiv.org/abs/2608.11256 — accessed 2026-09-02 (added 2026-09-02, hence the out-of-section number). [type: peer-reviewed] — **Preprint, labelled as such wherever cited; the August-2026 re-examination that lets the deck keep teaching the 2023 detector verdicts [8][9] as baselines rather than as current claims.** Controlled study of published English abstracts (four domains, 2013–2015 vs 2023–2025 cohorts) against commercial detectors at threshold 0.50. Quoted verbatim from the abstract: "Light refine (abstract only) edits, a proxy for guideline-compliant AI assistance, are flagged at 38 to 80%. Unmodified 2023 to 2025 originals are flagged at 9 to 15%, with non-STEM rates far above STEM (p < 0.001)"; "After Undetectable AI humanization, evasion is near-total: fewer than 4% of AI-labeled rewrites remain flagged (post-humanization detection rate < 4%; FNR > 96%). Honest AI-editing results in a higher sanction risk than humanizer-assisted evasion." Grounds the deck's current-generation detector bullets (slides 22–23), the handout's checklist item 10, and the demo script's decline-to-run-a-detector line. Note what it does *not* re-verify: whether current detectors still key on perplexity — see the historicization note at [9].

### C. The equity argument for language polishing

- [12] Amano, T., Ramírez-Castañeda, V., Berdejo-Espinola, V., Borokini, I., Chowdhury, S., Golivets, M., González-Trujillo, J. D., Montaño-Centellas, F., Paudel, K., White, R. L., & Veríssimo, D. (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 — accessed 2026-07-28. [type: peer-reviewed] — Quantifies the burden that language polishing is meant to relieve, so the equity argument on the slide rests on measurement rather than sentiment. Survey of 908 environmental-science researchers across eight nationalities: non-native English speakers from moderate- and low-proficiency countries spend a median 50.6% and 29.8% more time writing a paper in English, and 46.6% and 90.8% more time reading one; "38.1% and 35.9%… have experienced paper rejection due to English writing, while only 14.4% of native English speakers have" (≈2.5×); 42.5% and 42.6% report being asked to improve their English during revision, against 3.4% of native speakers (≈12.5×); around 30% of early-career non-native speakers report often or always deciding not to attend English-language conferences because of the language barrier.
- [13] Prakash, A., Aggarwal, S., Varghese, J. J., & Varghese, J. J. (2025). *Writing without borders: AI and cross-cultural convergence in academic writing quality*. Humanities and Social Sciences Communications 12, 1058. https://doi.org/10.1057/s41599-025-05484-6 (full text: https://www.nature.com/articles/s41599-025-05484-6) — accessed 2026-07-28. [type: peer-reviewed] — The counterpart to [12]: evidence that the gap has begun to close in the period AI writing tools became available. Over one million English-language social-science abstracts from Web of Science, 2012–2024, analysed for readability/complexity with a mixed generalised linear model plus lexical tracking of AI-associated adjectives and adverbs. Quoted: "The findings reveal a global improvement in writing complexity, with non-native English-speaking countries showing notable advances… Enhanced digital infrastructure and the adoption of AI-assisted writing tools appear to play a contributory role in this convergence." **Cited on the slide as association, not causation** — the design is observational and the AI signal is a lexical proxy; the slide says so.

### D. The cross-industry rules

- [14] International Committee of Medical Journal Editors (2025). *Recommendations for the Conduct, Reporting, Editing, and Publication of Scholarly Work in Medical Journals — II.A.4 Artificial Intelligence (AI)-Assisted Technology*. https://www.icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html — accessed 2026-07-28. [type: policy] — The reference standard behind most medical-journal policies, and the anchor of the deck's ICMJE segment (also carried forward for Session 7's medical deep-dive). Quoted verbatim: "At submission, the journal should require authors to disclose whether they used artificial intelligence (AI)-assisted technologies (such as Large Language Models [LLMs], chatbots, or image creators) in the production of submitted work… Authors who use such technology should describe, in both the cover letter and the submitted work in the appropriate section if applicable, how they used it. For example, if AI was used for writing assistance, describe this in the acknowledgment section… If AI was used for data collection, analysis, or figure generation, authors should describe this use in the methods. Chatbots (such as ChatGPT) should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work… Therefore, humans are responsible for any submitted material that included the use of AI-assisted technologies. Authors should carefully review and edit the result because AI can generate authoritative-sounding output that can be incorrect, incomplete, or biased… Authors should be able to assert that there is no plagiarism in their paper, including in text and images produced by the AI." Section II.A.3 adds the rule the handout uses: "Use of AI for writing assistance should be reported in the acknowledgment section."
- [15] International Committee of Medical Journal Editors (2025). *Recommendations — II.C Responsibilities in the Submission and Peer-Review Process (Confidentiality)*. https://www.icmje.org/recommendations/browse/roles-and-responsibilities/responsibilities-in-the-submission-and-peer-peview-process.html — accessed 2026-07-28. [type: policy] — Grounds the peer-review segment's framing that confidentiality is the *reason* for the reviewer rules, not an add-on. Quoted: "Manuscripts submitted to journals are privileged communications that are authors' private, confidential property, and authors may be harmed by premature disclosure of any or all of a manuscript's details… Editors should be aware that using AI technology in the processing of manuscripts may violate confidentiality… Instructions to reviewers should include guidance about AI use."
- [16] COPE Council (2023, last reviewed 13 February 2023). *COPE position statement: Authorship and AI tools*. https://publicationethics.org/guidance/cope-position/authorship-and-ai-tools — accessed 2026-07-28. [type: policy] — The cross-publisher statement that makes "no AI authorship" universal rather than publisher-specific; cited by Wiley's guidelines explicitly. Quoted: "AI tools cannot meet the requirements for authorship as they cannot take responsibility for the submitted work. As non-legal entities, they cannot assert the presence or absence of conflicts of interest nor manage copyright and license agreements. Authors who use AI tools in the writing of a manuscript, production of images or graphical elements of the paper, or in the collection and analysis of 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. Authors are fully responsible for the content of their manuscript, even those parts produced by an AI tool."
- [31] TechCrunch (2026, 16 May). *Research repository arXiv will ban authors for a year if they let AI do all the work*. https://techcrunch.com/2026/05/16/research-repository-arxiv-will-ban-authors-for-a-year-if-they-let-ai-do-all-the-work/ — accessed 2026-09-02 (added 2026-09-02, hence the out-of-section number; corroborated by 404 Media, https://www.404media.co/new-arxiv-rules-ai-generated-papers-ban/). [type: news] — **Grounds the arXiv one-year-ban bullet (slide 15) and the handout's enforcement paragraph. Explicit deviation from this session's policy-source rule, stated here because attendees may act on it:** the announcement was made via posts by Thomas Dietterich, chair of arXiv's Computer Science section, and no primary arXiv policy page could be located at revision time — so this claim rests on secondary press coverage, and the deck dates it only as "May 2026" (reported 2026-05-16; the exact announcement day is unverified). Per the coverage, quoting Dietterich: the ban applies on "incontrovertible evidence that the authors did not check the results of LLM generation" — examples given include "hallucinated references" and leftover "comments to or from the LLM"; the penalty is "a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted by a reputable peer-reviewed venue". It is not a prohibition on LLM use — authors retain full responsibility "irrespective of how the contents are generated"; per 404 Media it is a one-strike rule with moderator flag, section-chair confirmation, and appeal rights. Anyone relying on this in earnest should locate arXiv's own current moderation guidance first.

### E. The publisher policies — one row of the comparison table each, each from the publisher's own page

- [17] Elsevier (2026, policy updated June 2026). *Generative AI policies for journals*. https://www.elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals — accessed 2026-07-28. [type: policy] — The Elsevier row of the comparison table and disclosure template 1 in the handout. Required declaration, quoted from the page: title of new section "Declaration of generative AI and AI-assisted technologies in the manuscript preparation process", placed "at the end of the manuscript… in a new section before the references list", with the statement, quoted byte-exactly from the page on 2026-07-28 (note the comma after "work", which several Elsevier journal-level guides omit — see [26]): "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." Exemption quoted: "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." Authorship: "Authorship implies responsibilities and tasks that can only be attributed to and performed by humans", so AI tools may not be listed as author or co-author. Reviewers/editors: "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." **Recently changed — say so on the slide:** the section heading and scope moved from "…in the writing process" to "…in the manuscript preparation process" with the June 2026 update, and many journal-level guides still carry the older heading (see [26]).
- [18] Springer Nature / Nature Portfolio (2026). *Artificial Intelligence (AI) — editorial policy*. https://www.nature.com/nature-portfolio/editorial-policies/ai and https://www.springernature.com/gp/policies/editorial-policies — accessed 2026-07-28. [type: policy] — The Springer Nature row. Quoted verbatim: "Large Language Models (LLMs), such as ChatGPT, do not currently satisfy our authorship criteria… 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) of the manuscript. The use of an LLM (or other AI-tool) for 'AI assisted copy editing' purposes does not need to be declared" — where AI-assisted copy editing means "AI-assisted improvements to human-generated texts for readability and style, and to ensure that the texts are free of errors in grammar, spelling, punctuation and tone… but do not include generative editorial work and autonomous content creation." Images: "Springer Nature journals are unable to permit its use for publication", with narrow listed exceptions. Reviewers: "we ask that… peer reviewers do not upload manuscripts into generative AI tools", and any AI-supported evaluation must be declared "transparently in the peer review report". The corporate policy page states the same three rules in summary form: "SN does not attribute authorship to AI. SN does not allow the inclusion of generative AI images in our publications. SN asks peer reviewers not to upload manuscripts into generative AI tools."
- [19] Wiley (2026). *Best Practice Guidelines on Research Integrity and Publishing Ethics — Artificial Intelligence*. https://authorservices.wiley.com/ethics-guidelines/index.html — accessed 2026-07-28. [type: policy] — The Wiley row and disclosure template 2 in the handout. Quoted: "If an author has used AI Technology to substantially edit, develop, or translate any part of a manuscript, its use must be described transparently and integrated into the manuscript… AI Tools used solely for spelling, grammar, and general editing are not included in these disclosure requirements." Authors "must also disclose their use of AI Technologies when submitting a manuscript to a Wiley journal. If disclosure is not provided at submission, it may be requested during peer review or after publication." Authorship, citing COPE: such tools "cannot fulfil the role of an author and must not be listed as one on an article." Distinctive to Wiley and worth a line on the slide: the **rights** clause — "Authors must ensure the AI Technology and the provider of that AI Technology does not gain any rights over the author's underlying content, including the right to 'train' their AI Technology on the content" — and the peer-review clause permitting AI only to "improve the clarity or quality of the written feedback", with disclosure to the handling editor, and never uploading the manuscript.
- [20] Taylor & Francis (2025, page dated 19 November 2025). *AI Policy*. https://taylorandfrancis.com/our-policies/ai-policy/ — accessed 2026-07-28. [type: policy] — The Taylor & Francis row, and the clearest publisher statement of the *legitimate* uses the deck's editor-vs-ghostwriter segment defends. Quoted: T&F "welcomes the new opportunities offered by Generative AI tools, particularly in: copyediting and language refinement, enhancing idea generation and exploration, supporting authors to express content in a non-native language, and accelerating the research and dissemination process." Disclosure requirement quoted: "Authors must clearly acknowledge within the article or book any use of Generative AI tools through a statement that includes: the full name of the tool used (with version number), how it was used and the reason for use." Prohibited: "Text or code generation without rigorous revision", synthetic data substituting for missing data, and inaccurate generated abstracts or supplements. Journal-level override quoted in the handout caveat: "some journals may not allow use of Generative AI tools beyond language improvement, therefore authors are advised to review the 'Instructions for authors' on the journal homepage prior to submission." Reviewers: "must not use artificial intelligence tools to generate manuscript and proposal review reports… However, Generative AI may be utilised to assist with improving review language."
- [21] Sage (2026). *Artificial intelligence policy* (publication ethics policies). https://us.sagepub.com/en-us/nam/chatgpt-and-generative-ai — accessed 2026-07-28. [type: policy] — The Sage row, and the policy the deck uses to ground the editor-vs-ghostwriter spectrum in a real publisher's own vocabulary. Quoted: "We distinguish various uses for AI and related technologies as: assistive (and no longer requiring disclosure), generative (requiring disclosure), and prohibitive." Assistive, quoted: "AI tools that make suggestions to improve or enhance your own work, such as tools to improve language, grammar, or structure, are considered assistive AI tools and do not require disclosure by authors or reviewers." Generative, quoted: "The primary or partial use of AI tools and/or LLMs that produce content such as references, text, images, or any other content that directly impacts the research methodology, analysis, results and/or conclusions must be disclosed upon submission", "within the methods or acknowledgements" — with translation of materials and literature-review assistance explicitly listed as disclosable. Prohibited uses include peer review and editorial work, fabricated references, and generated images presented as research images. Enforcement line quoted in the handout: "submissions will not be rejected solely because of the disclosed use of GenAI tools", but undisclosed inappropriate use may be rejected "at any time during the publishing process".
- [22] American Association for the Advancement of Science (2026). *Science Journals: Editorial Policies — Image and Text Integrity: Artificial intelligence (AI)*. https://www.science.org/content/page/science-journals-editorial-policies — accessed 2026-07-28. [type: policy] — The Science/AAAS row, quoted verbatim from the current page: "AI-assisted technologies [such as large language models (LLMs), chatbots, and image creators] do not meet the Science journals' criteria for authorship and therefore may not be listed as authors or coauthors, nor may sources cited in Science journal content be authored or coauthored by AI tools. Authors who use AI-assisted technologies as components of their research study or as aids in the writing or presentation of the manuscript should note this in the cover letter and in the methods section or acknowledgments section of the manuscript (the location and details of the declaration will depend on use case; see our AI usage guidelines for specifics)… Editors may decline to move forward with manuscripts if AI is used inappropriately. Reviewers may not enter any part of the manuscript into an LLM or other AI tool because this could breach the confidentiality of the manuscript. However, reviewers may use AI tools to revise their own writing, provided that the system does not save or use inputs to train the model and that such use is transparently declared in the review." Images: "AI-generated images and other multimedia are not permitted in the Science journals without explicit permission from the editors." The authorship section adds: "artificial intelligence tools cannot be authors."
- [23] Thorp, H. H., & Vinson, V. (2023, 16 November 2023). *Change to policy on the use of generative AI and large language models*. Science Editor's Blog, AAAS. https://www.science.org/content/blog-post/change-policy-use-generative-ai-and-large-language-models — accessed 2026-07-28. [type: policy] — Dates the Science policy change on the slide, so attendees who remember the January 2023 blanket ban are corrected rather than confused. Quoted: "we initially took a very restrictive stance regarding the use of ChatGPT in preparing text and figures", replaced in November 2023 by disclosure in the cover letter and acknowledgments plus, in the methods, "The full prompt used in the production of the work, as well as the AI tool and its version". Comparing this text with [22] shows a **second, later revision**: the November 2023 wording said reviewers "may not use AI technology in generating or writing their reviews", whereas the current page permits reviewers to use AI to revise their own writing under stated conditions. The slide states both the change and the direction of travel.
- [24] JAMA Network (2026). *Instructions for Authors — Use of AI in Publication and Research* (JAMA and JAMA Network Open). https://jamanetwork.com/journals/jama/pages/instructions-for-authors and https://jamanetwork.com/journals/jamanetworkopen/pages/instructions-for-authors — accessed 2026-07-28. [type: policy] — The JAMA row. Quoted verbatim from the instructions: "Authors should report the use of artificial intelligence, language models, machine learning, or similar technologies to create content or assist with writing or editing of manuscripts in the Acknowledgment section or the Methods section if this is part of formal research design or methods. This should include a description of the content that was created or edited and the name of the language model or tool, version and extension numbers, and manufacturer. (Note: this does not include basic tools for checking grammar, spelling, references, etc.)" And: "The submission and publication of content created by artificial intelligence, language models, machine learning, or similar technologies is discouraged, unless part of formal research design or methods, and is not permitted without clear description of the content that was created and the name of the model or tool, version and extension numbers, and manufacturer. Authors must take responsibility for the integrity of the content generated by these models and tools." Also: "Use of AI, LLM, and chatbots to draft Letters to the Editor is not permitted."
- [25] Flanagin, A., Pirracchio, R., Khera, R., Berkwits, M., Hswen, Y., & Bibbins-Domingo, K. (2024). *Reporting Use of AI in Research and Scholarly Publication — JAMA Network Guidance*. JAMA 331(13), 1096–1098. https://doi.org/10.1001/jama.2024.3471 — accessed 2026-07-28. [type: policy] — First-party expansion of [24], used for disclosure template 3 in the handout because it is the most prescriptive element list any major publisher provides. Required elements quoted: "Name of the AI software platform, program, or tool", "Version and extension numbers", "Manufacturer", "Date(s) of use", "A brief description of how the AI was used and on what portions of the manuscript", and confirmation "that the author(s) takes responsibility for the integrity of the content generated". For AI used as part of the research itself, reporting moves to the Methods and adds: "Specify dates and prompt(s) used and their sequence and any revisions to prompts in response to initial outputs."
- [26] Elsevier journal guides for authors (2026), e.g. *Research Policy* (https://www.sciencedirect.com/journal/research-policy/publish/guide-for-authors) and *Sensors and Actuators Reports* (https://www.sciencedirect.com/journal/sensors-and-actuators-reports/publish/guide-for-authors) — accessed 2026-07-28. [type: primary-doc] — Evidence for the deck's "even inside one publisher, check the journal" slide, and the reason the handout's caveat is not boilerplate. On the same date, one Elsevier journal instructs authors to head the section "Declaration of generative AI and AI-assisted technologies in the writing process" with the statement ending "…full responsibility for the content of the publication", while another instructs "Declaration of generative AI and AI-assisted technologies in the manuscript preparation process" ending "…full responsibility for the content of the published article". Both journal-level guides also drop the comma after "work" that the publisher-level policy page carries ([17]), and one renders the placeholder as "[NAME TOOL / SERVICE]" rather than "[NAME OF TOOL / SERVICE]" — which is exactly why the handout tells authors that the journal's wording wins. Several also add a journal-level reviewer clause: "this journal does not currently allow the use of generative AI or AI-assisted technologies such as ChatGPT or similar services by reviewers or editors in the peer review and manuscript evaluation process."

### F. Funder and peer-review rules

- [27] National Institutes of Health (2023, released 23 June 2023). *NOT-OD-23-149: The Use of Generative Artificial Intelligence Technologies is Prohibited for the NIH Peer Review Process*. https://grants.nih.gov/grants/guide/notice-files/NOT-OD-23-149.html — accessed 2026-07-28. [type: policy] — The strictest rule an attendee is likely to encounter, quoted verbatim on the funder slide: "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 for grant applications and R&D contract proposals… Reviewers should be aware that 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." Enforcement quoted: "all NIH Peer Reviewers will be required to sign and submit a modified Security, Confidentiality and Nondisclosure Agreement certifying that they fully understand and will comply with… the prohibition on uploading or sharing content." Accessibility exception quoted in the handout: "Computer technologies that are used for accessibility needs may be granted an exception to this policy", with prior notice to the Designated Federal Officer. **Re-fetched and re-verified 2026-07-28:** all four quotations above are byte-exact against the live notice; the page shows Release Date "June 23, 2023", no later update and no superseding notice — its only related announcement remains NOT-OD-22-044 (30 December 2021). The notice also links a "Use of Generative AI in Peer Review FAQs" resource, which the deck does not quote.
- [28] U.S. National Science Foundation (2023, 14 December 2023). *Notice to research community: Use of generative artificial intelligence technology in the NSF merit review process*. https://www.nsf.gov/news/notice-to-the-research-community-on-ai — accessed 2026-07-28. [type: policy] — The second funder rule on the slide, and the one that also speaks to *proposers*. Quoted: "NSF reviewers are prohibited from uploading any content from proposals, review information and related records to non-approved generative AI tools" and "Proposers are encouraged to indicate in the project description the extent to which, if any, generative AI technology was used and how it was used to develop their proposal." Rationale quoted: "sharing proposal information with generative AI technology via the open internet violates the confidentiality and integrity principles of NSF's merit review process. Any information uploaded into generative AI tools not behind NSF's firewall is considered to be entering the public domain." **Re-fetched and re-verified 2026-07-28:** every quotation above is byte-exact against the live notice, still dated 14 December 2023, with no superseding notice on the page. One forward-looking sentence has since come due and is worth flagging to anyone using this in 2026 — "NSF will update the 2025 PAPPG to align with the requirements stipulated in this memorandum" — so proposers should also check the current Proposal and Award Policies and Procedures Guide; the deck makes no claim about PAPPG text. The notice additionally states that reviewers "may share publicly available information with current generation generative AI tools", which the deck does not quote.

---

**Counts for this session:** 31 annotated sources — 15 peer-reviewed/preprint ([1]–[6], [8]–[13], [29], [30]), 13 policy ([14]–[25], [27]–[28]), 1 primary-doc ([26]), 2 news ([7], first-party survey labelled as self-reported survey data wherever used; [31], secondary coverage of the arXiv announcement, deviation from the policy-source rule stated in its entry). Entries [29]–[31] were added in the 2026-09-02 revision and sit in their thematic sections with out-of-sequence numbers so that the pre-existing [n] markers in `slides.html` and `handout.md` did not need renumbering.

**Dropped during research and why.** (a) Vendor and editing-service summaries of publisher policies (Enago, thesify, university LibGuides) — useful for finding pages, never citable here; every policy claim was re-fetched from the publisher's own site. (b) A widely repeated "70% of journals have AI policies vs 0.1% of papers disclose" framing circulating without attribution — kept **only** after locating the single primary study [1] that reports both figures, and only with its method caveat and its published critique [2] on the same slide. (c) Per-document commercial-detector output as a measurement — specifically the "11.3% of articles flagged >10% AI-generated" figure in [5] (Originality.AI), which is cited as context but never quoted as a measurement, because the deck's own Section 04 shows per-document detectors are unreliable. **Correction of record (2026-07-28):** an earlier version of this note also placed [1] in this category, describing its usage estimate as ZeroGPT-derived. That was wrong — [1]'s primary method is a group-level maximum-likelihood estimate, with ZeroGPT only one of three robustness checks. [1]'s usage figures are therefore used on the deck, with the authors' own stated limitation (the method cannot separate permissible polishing from substantive generation) rather than a detector-reliability caveat. (d) A 1.7%-of-radiology-papers disclosure figure and a "24% to 76% of researchers use AI" range that appeared only in secondary summaries without a traceable primary — dropped.

---

*AI for Researchers · Session 4: Writing, Publishing & Integrity · Landscape as of August 2026*