Existing reviews first, then the seminal work, then the work that cites it; strands 9 to 16 are practice. Grades are defined in README.md; the evidence against is gathered in against.md. Sources from outside journals are scored in sources.md; a source below 10 of 20 is marked here as an example of practice only.
0. Existing reviews
Searched first, as Garousi 2019 asks (protocol.md 3). Each covers one part of the article; none covers the whole.
| Hartley 2004 |
narrative survey, 31 studies |
structured abstracts |
adds practice and the other front matter (strand 4) |
measured |
| Turner 2012 |
Cochrane review, 50 evaluations |
endorsement of CONSORT and completeness |
read with Stevens 2014 as the limit of that result |
measured |
| Stevens 2014 |
systematic review |
101 guidelines other than CONSORT; 15,249 records retrieved, 26 evaluations of nine of them; only the PRISMA summed score significant (SMD 0.53, 99% CI 0.02 to 1.03, endorsing against non-endorsing journals); insufficient evidence |
adds venue checklists in machine learning, which it does not cover |
measured |
| Blanco 2019 |
scoping review |
109 references, 31 interventions to raise adherence to reporting guidelines, 11 evaluated |
studies what authors and journals do with the article, not training or tools |
measured, descriptive |
| Galipeau 2015 |
systematic review |
training in writing, editing and peer review; the title reports a knowledge gap; results not read |
teaches no one; describes the article |
supported with caveat |
| Stoll 2022 |
systematic review |
plain language summaries: 7,714 records, 90 included, 33 empirical; evidence on criteria scarce |
adds journals' own rules (Taylor 2024 (Taylor & Francis), PLOS 2026) and the summary experiments (strand 12) |
measured |
| Gainey 2022 |
scoping review |
author instructions for plain language summaries in 534 health journals; 27 gave them |
adds what the instructions require in two publishers |
measured, descriptive |
| Krukowski 2023 |
narrative review |
graphical abstracts: purpose and design advice; benefits called emerging |
keeps measured outcomes apart: attention, citations, understanding |
supported with caveat |
| Bhavsar 2025 |
cross-sectional audit |
AI-chatbot policies of 162 STM publishers, 2023-09 to 2023-12: 56 had one |
reads it with Ganjavi 2024 and venue rules (strand 15) |
measured, descriptive |
This search did not find a systematic review of IMRaD, of highlights, or of machine-learning checklists beyond the corpus studies in strand 9. It also did not find one that spans front matter, checklists, availability statements and AI rules together.
1. History and prevalence of IMRaD
| Gross 2002 |
survey (history) |
the scientific article in English, French and German from the 17th century to the present |
supported with caveat |
| Atkinson 1999 |
survey (corpus history) |
Philosophical Transactions 1675-1975: rhetoric moves from author-centred to object-centred |
supported with caveat |
| Bazerman 1988 |
survey (rhetoric) |
the experimental report took its form in answer to rhetorical and social situations |
supported with caveat |
| Sollaci 2004 |
seminal (counts) |
1,297 articles in BMJ, JAMA, Lancet and NEJM, 1935-1985: IMRaD above 10% by 1950, above 80% in the 1970s, the only pattern in the 1980s |
measured |
| Gastel 2016 |
handbook (Day's) |
teaches IMRaD as the default shape of a primary research paper |
convention, untested |
| ICMJE 2026 |
editorial standard |
original research usually divided into IMRAD; reviews, case reports and editorials may differ |
convention, untested |
| Lin 2012 |
forward (corpus) |
433 empirical articles, 39 disciplines: plain IMRD is not the default; ILM[RD]C ranks first |
measured |
| Moskovitz 2024 |
forward (corpus) |
biology mostly IMRaD; under half in engineering and social science; most computational papers not IMRaD |
measured |
| GMU 2026 |
university writing page (example of practice only, 7.5/20) |
teaches IMRaD as the shape of lab and research reports and cites no study; the page names no fields where it fails |
convention, untested |
2. Structure by checklist: reporting guidelines
| Simera 2010 |
survey |
a catalogue of reporting guidelines for health research, the base of the EQUATOR library |
convention, untested |
| Schulz 2010; Hopewell 2025 |
seminal guideline and update |
CONSORT 2010 and CONSORT 2025: checklist and flow diagram for randomised trials |
convention, untested |
| Page 2021 |
guideline |
PRISMA 2020: checklist for systematic reviews |
convention, untested |
| Hopewell 2008 |
guideline |
CONSORT for abstracts: the items a trial abstract must report |
convention, untested |
| Liu 2020 |
guideline extension |
CONSORT-AI adds 14 items (11 extensions, 3 elaborations) to the 25-item CONSORT 2010 checklist, from a Delphi survey of 103 then 91 stakeholders and a 31-person meeting |
convention, untested |
| COS 2019 |
working-group pilot (owner's account) |
MDAR minimum-standards checklist for life-science journals; pilot at 13 journals, 211 authors, 289 manuscripts; about 80% found it useful; framework paper is Macleod 2021, not read |
convention, untested |
| Turner 2012 |
forward (Cochrane review) |
50 evaluations, 16,604 trials: 25 of 27 outcomes (22 checklist items plus blinding and a summary score) favour endorsing journals, 5 significantly; allocation concealment 45% vs 22% (RR 1.81) |
measured |
3. Genre analysis: what each part does
| Swales 1990 |
seminal |
CARS: Move 1 establish a territory, Move 2 establish a niche, Move 3 occupy the niche |
supported with caveat |
| Swales 2004 |
survey |
revises CARS (Move 3 becomes presenting the present work) and reviews move studies |
supported with caveat |
| Swales 2012 |
textbook (with Feak) |
teaches the moves to graduate writers, section by section |
convention, untested |
| Hill 1982 |
seminal |
hourglass: the introduction narrows from field to study, the discussion widens back |
supported with caveat |
| Hyland 2004 |
model (with Tse) |
interactive and interactional metadiscourse, from 240 dissertations, 4 million words |
supported with caveat |
| Yang 2016 |
forward (corpus) |
highlights behave as a genre of their own, dense in evaluative language |
supported with caveat |
4. Abstracts and front matter
| Ad Hoc Working Group 1987 |
seminal proposal |
structured abstracts for clinical articles, with fixed labelled parts |
convention, untested |
| Haynes 1990 |
forward |
revisits the proposal after journals adopted it, and refines the headings |
supported with caveat |
| Hartley 2004 |
survey (31 studies) |
structured abstracts are longer (+21%; +35% when independently published), usually more informative, sometimes easier to read and search |
measured |
| Hartley 2014 |
survey update |
spread beyond medicine; early studies biased by rewritten abstracts and student judges |
supported with caveat |
| Taddio 1994 |
before and after |
300 abstracts, 33 criteria: quality 0.57 unstructured vs 0.74 structured |
measured |
| Sharma 2006 |
controlled comparison |
600 dental abstracts, 29-item checklist: scores rose only in journals that switched |
measured |
| Hartley 2003 |
rewriting study |
24 psychology abstracts: longer, more informative, more readable, judged clearer by 48 authors |
measured |
| Budgen 2008 |
randomised paired judgement |
software engineering, 64 judges, 25 papers: completeness +6.65 on an 18-item questionnaire, clarity +2.98 on 1-10; 40 of 57 preferred structured |
measured |
| Pferschy-Wenzig 2016 |
observational |
one journal, 2014-2015: papers without a graphical abstract had more downloads, views and citations |
supported with caveat |
| Kim 2022 |
observational |
4,205 gastroenterology articles: with a graphical abstract, median citations 14 vs 12 |
supported with caveat |
| Ibrahim 2017 |
crossover |
44 articles tweeted both ways: visual abstract 7.7 times the impressions, 2.7 times the visits |
supported with caveat |
| Oska 2020 |
crossover, random order |
40 articles: visual abstract tweets drew about twice the views of citation or key-figure tweets |
measured |
| Chapman 2019 |
randomised trial |
41 manuscripts: visual beat plain-English abstracts on engagement; public engagement low throughout |
measured |
| Elsevier highlights |
publisher guideline (example of practice only, 9.0/20) |
3 to 5 bullets of at most 85 characters, meant for discoverability |
convention, untested |
| Elsevier GA 2026 |
publisher guideline (example of practice only, 9.0/20) |
graphical abstract at least 1328 x 531 pixels, 300 dpi, labels inside the image, no separate caption; the stated aim is a quick takeaway and discovery |
convention, untested |
5. How researchers read
| Tenopir 2008 |
survey series since 1977 |
readings per scientist nearly doubled 1977-2005; minutes per reading fell from 48 to 31 |
supported with caveat |
| Tenopir 2015 |
same series, 2012 wave |
about 21 articles a month, about 32 minutes each |
supported with caveat |
| Nicholas 2007 |
log analysis |
many users of journal sites view a few items and do not come back (bouncing) |
supported with caveat |
| Renear 2009 |
review |
scientists read strategically: search, filter, scan, link and extract fragments |
supported with caveat |
| Hubbard 2017 |
survey, 260 undergraduates and researchers in biological sciences at one UK university |
experienced readers rank figures and results first, undergraduates the abstract; all rank the introduction low |
supported with caveat |
| Keshav 2007 |
advice |
three passes: 5-10 minutes on title, abstract, introduction, headings, conclusions; up to an hour with the figures; 4-5 hours for a novice to re-derive it |
convention, untested |
| Mensh 2017 |
advice |
context, content, conclusion at the scale of paper, section and paragraph |
convention, untested |
6. Prose inside the structure
| Gopen 1990 |
rhetoric |
readers expect linking information in the topic position and new information in the stress position at the end |
supported with caveat |
| Pinker 2014 |
book |
the curse of knowledge and a self-conscious style (hedges, metadiscourse, nominalisation) explain academic prose |
supported with caveat |
| Sword 2012 |
corpus, about 1,000 articles surveyed, 500 scored, ten fields |
under 10% met her three criteria of engagement, accessibility and craft (secondary) |
supported with caveat |
| Plavén-Sigray 2017 |
corpus |
709,577 abstracts, 123 journals, 1881-2015: FRE falls with year (r = -0.93), general jargon rises (r = 0.96) |
measured |
| Weinberger 2015 |
corpus |
over one million abstracts, eight fields: about half the common style advice goes with fewer citations |
supported with caveat |
| Hillier 2016 |
corpus |
732 climate abstracts: more narrative abstracts are cited more, largely through journal identity |
supported with caveat |
7. Alternative shapes and bundles
| Chambers 2013; Nosek 2014 |
seminal format |
Registered Reports: Stage 1 review of question and method, in-principle acceptance, then results |
convention, untested |
| Chambers 2022 |
survey |
history, early evidence and limits of Registered Reports; the count of over 300 journals offering them is from the COS page, not from this paper |
supported with caveat |
| Findley 2016 |
pilot |
results-free review at one political science journal: 19 submissions, 3 published |
supported with caveat |
| Buckheit 1995 |
seminal (reproducible research) |
the article advertises the work; the code and data are the work |
convention, untested |
| Olah 2017 |
essay |
research debt: explanation is undervalued work; interactive articles as a remedy (Distill) |
convention, untested |
| eLife 2020 |
platform (example of practice only, 9.5/20) |
Executable Research Article: readers re-run and edit the code behind a figure in the browser |
convention, untested |
| Groth 2010 |
seminal |
nanopublication: one assertion with its provenance, as linked data |
convention, untested |
| Octopus |
platform (example of practice only, 7.5/20) |
eight linked units: problem, hypothesis, method, results, analysis, interpretation, applications, review |
convention, untested |
| Hartley 1999 |
proposal |
structured articles: headings from writing research, tried on 20 published papers |
supported with caveat |
8. Templates and standards in use
| elsarticle |
Elsevier, on CTAN |
layout options (preprint, review, 1p, 3p, 5p), front matter (author, affiliation, email, abstract, graphical abstract, highlights, keywords), reference style |
| llncs |
Springer, on CTAN |
page layout, institute and keywords macros, theorem-like environments, the splncs04 reference style |
| IEEEtran |
IEEE, on CTAN |
journal and conference modes, IEEE keywords block, drop-cap first paragraph, IEEE reference style |
| acmart |
ACM, on CTAN |
formats (sigconf, acmsmall and others), CCS concepts, keywords, rights block, DOI, teaser figure, figure descriptions for accessibility |
| arXiv |
preprint server |
takes TeX source with no house template; HTML versions since 2023, because 2.4% of research outputs met accessibility guidelines (Frankston 2024) |
| JATS |
NISO standard |
machine-readable article metadata, including an article-type attribute |
| ICMJE 2026 and journals |
article types |
research article, review, case report, letter, each with its own word limits and abstract format |
All of them fix layout, metadata and reference style; none fixes the argument, which stays with the writer. The IEEEtran journal sample holds an Introduction, a Conclusion and appendices; the elsarticle sample holds one example section.
9. Checklists at ML and NLP venues
| Gundersen 2018 |
survey (400 papers) |
IJCAI and AAAI papers document 20% to 30% of the variables needed to reproduce them; none documents all |
measured |
| Dodge 2019 |
seminal (NLP checklist) |
argues for reporting validation scores and budgets; checks 50 random 2018 EMNLP papers against a checklist (secondary: the count is read in Magnusson 2023), the baseline later reused |
supported with caveat |
| Pineau 2021 |
seminal (practice report) |
2019 NeurIPS: checklist mandatory at submission and at camera-ready, 6,743 submissions; model of the Nature life-science checklist |
convention, untested |
| Pineau 2021 |
same |
about 40% of authors said they had code at submission; reviewers confirmed 71.5% of those (so about 29% overall, derived: 40% x 71.5%); 74.4% had code at camera-ready |
measured, descriptive |
| Pineau 2021 |
same |
34% of reviewers found the checklist answers useful; 36% of submissions judged error bars not applicable; the authors state there is no evidence yet that quality improved |
supported with caveat |
| Magnusson 2023 |
forward (analysis of the NLP checklist) |
10,405 anonymous responses, EMNLP in 2020 and 2021, NAACL and ACL in 2021; 45.9% of submissions link code; those that do score 0.30 higher on reviewer-rated reproducibility |
measured, descriptive |
| Magnusson 2023 |
same |
acceptance rate rises with the number of yes answers; 44% of submissions collect new data and are 5% less likely to be accepted |
measured, descriptive |
| Magnusson 2023 |
same |
against 50 manually checked 2018 EMNLP papers, 8 of 10 items are reported more often after the checklist; rates were flat across the following year |
supported with caveat |
| Magnusson 2023 |
same |
reviewers of NAACL and ACL in 2021: 59.9% found the checklist somewhat useful, 17.0% very useful, 23.2% not useful |
measured, descriptive |
| Coakley 2026 |
corpus, 56,800 papers at five AI venues, 2014-2024 (preprint, abstract read) |
papers sharing code and data rose from 11% to 64%; the authors say the rise predates the checklists and reflects open-science norms; the 28% to 64% reproducibility estimate is inferred from documentation, not tested |
supported with caveat |
| Rogers 2021 |
proposal |
a checklist for responsible data use in NLP, argued from legal and ethical principles; no outcome measured |
convention, untested |
| Nanayakkara 2021 |
thematic analysis |
2020 NeurIPS required a broader impact statement; the statements vary in specificity, valence and uncertainty |
supported with caveat |
| Ashurst 2022 |
dataset and analysis |
all 2020 NeurIPS impact statements; the requirement was dropped in 2021 for a checklist; lessons on incentives, guidance, deliberation |
supported with caveat |
| Kapoor 2023 |
survey and case study |
data leakage found in 17 fields, 294 papers; proposes a model info sheet; a corrected civil-war prediction comparison shows no gain over logistic regression |
measured |
| Goldberg 2024 |
experiment |
234 2024 NeurIPS submissions voluntarily checked by a language-model assistant; over 70% of the 78 authors who answered the post-usage survey found it useful; causal effect on papers not established |
supported with caveat |
| NeurIPS 2021; NeurIPS 2026 |
venue policy |
2021: 5 questions, 18 items, yes/no/n.a., a no is not grounds for rejection. 2026: 16 questions including a language-model usage declaration; omitting the checklist means desk rejection |
convention, untested |
10. Artifacts and badges
| Collberg 2016 |
seminal audit |
402 papers whose results rest on code: the code was found and built within 30 minutes for 32.3% (secondary: read in a summary of the technical report, not opened; the CACM article uses 601 papers) |
measured |
| Saucez 2019 |
practice report |
two artifact committees at SIGCOMM venues awarded ACM badges to 12 and 28 papers |
measured, descriptive |
| Heumüller 2020 |
corpus (789 ICSE papers, 2007-2017) |
604 papers (76.6%) describe a software artifact; in 2017 only 58.5% of them made it public; a small positive correlation of linking an artifact with citations |
measured |
| Hermann 2020 |
survey of committee members |
no consensus on quality thresholds; terminology for the main purpose, replicability, is inconsistent |
supported with caveat |
| Timperley 2021 |
mixed methods |
153 researchers: expectations of creators, users and reviewers do not match; creators and reviewers lack reward |
supported with caveat |
| Winter 2022 |
forward (retrospective of a decade of AE) |
articles with artifacts do not get significantly more citations; articles with an "Available" badge have a working link among their references in 99.6% of cases against 90.3% without (pooled from Table 4); about 90% of eligible 2020 PLDI articles came with artifacts |
measured |
| Kidwell 2016 |
before and after with comparison journals |
Psychological Science open-data badges: under 3% of articles reported open data before, 23% after, 39% in the first half of 2015; no change in comparison journals; earned badges went with data that were available |
measured |
11. Documentation cards
| Mitchell 2019 |
seminal (model cards) |
short documents on intended use and measured performance across groups, shipped with a trained model |
convention, untested |
| Gebru 2021 |
seminal (datasheets) |
a question set for dataset creators, modelled on electronics datasheets |
convention, untested |
| Boyd 2021 |
think-aloud experiment |
23 engineers reviewed a flawed dataset, 11 with a datasheet; those with it raised ethical issues earlier and more often |
supported with caveat |
| Liang 2024b |
corpus (32,111 Hugging Face model cards) |
models with cards carry 90.5% of download traffic; filled-out rates: training 74.3%, limitations 17.4%, evaluation 15.4%, environmental impact 2.0% |
measured, descriptive |
| Liang 2024b |
intervention on 42 popular models |
adding a detailed card raised weekly downloads by 29.0% (95% CI 10.6% to 47.5%) in one batch; the other batch showed no clear effect |
supported with caveat |
12. Structured methods and plain-language summaries
| Marcus 2016 |
publisher editorial (owner's claim) |
Cell Press launched STAR Methods in 2016 as a methods format with a key resources table and persistent identifiers (editorial not read in full; format details from the publisher's author pages, secondary) |
convention, untested |
| Elsevier KRT 2026 |
publisher page (example of practice only, 9.5/20) |
key resources table lists resources with source and identifier; the page offers no study of benefit, only one anecdote of a caught typo |
convention, untested |
| Evanko 2025 |
before and after at one publisher (congress abstract, owner's data) |
a daily automated e-mail to authors of accepted-stage manuscripts lacking "RRID": published articles with RRIDs 33% (176 of 526) in April-July, 65% (330 of 505) in November-February; not controlled |
supported with caveat |
| Vasilevsky 2013 |
audit |
54% of research resources (antibodies, strains, cell lines and others) in sampled biomedical papers were not uniquely identifiable, whatever the journal's reporting rules |
measured |
| Babic 2019 |
text mining of about two million papers |
8.6% of cell lines named were on the problematic list; 3.3% in the 634 papers that used RRIDs; an association, not an effect |
supported with caveat |
| Jelicic Kadic 2016 |
audit |
of 1,738 Cochrane plain language summaries (2013-2015), none met every measured standard; the GRADE item 0.7%; overall adherence 57% |
measured |
| Karačić 2019 |
corpus |
readability (SMOG) of Cochrane texts: median 15.6 for scientific abstracts, 14.7 for plain language summaries, against a recommended sixth-grade level |
measured |
| Buljan 2018 |
three randomised trials |
students (n = 171), consumers (n = 99), doctors (n = 64): no difference in knowledge between an infographic and a text summary; the infographic was preferred (d = 0.48 reading experience) |
measured |
| Bredbenner 2019 |
randomised survey (n = 538) |
video and plain-language summaries gave higher perceived understanding than graphical abstracts and abstracts; all groups did well on the test; preference did not track comprehension |
measured |
| Taylor 2024 (Taylor & Francis) |
publisher guideline |
three formats (text, graphical, standalone article with its own DOI); a standalone summary with a recommended length of 3000 words or less (a guide, not a limit), peer reviewed including by lay experts; no outcome data |
convention, untested |
| PLOS 2026 |
journal instruction |
authors of research articles revised after review write an author summary, 2-3 one-sentence bullets under each of three questions, after the abstract |
convention, untested |
| eLife 2017 |
publisher survey (owner's claim) |
313 self-selected readers of digests; over 90% found most digests informative |
convention, untested |
| Markowitz 2024 |
corpus and experiments |
34,584 PNAS papers: significance statements are simpler than abstracts, with small effect sizes (d near 0.3 on word simplicity); readers understood GPT-written statements better than the published ones |
measured |
| Wang 2024 |
corpus |
over 3,000 pairs of abstract and significance statement across disciplines; ARI 18.9 for abstracts, 18.1 for statements |
measured |
13. Data and code availability statements
| Nosek 2015; COS 2024 |
standard and its update |
TOP guidelines of 2015, revised in 2025: seven research practices each at three levels (disclosed, shared and cited, certified), plus verification practices; Grant 2024 is the preprint, not read |
convention, untested |
| Hardwicke 2018 |
interrupted time series, one journal |
after a mandatory open-data policy at Cognition, data-available statements rose from 25% to 78%, reusable data from 22% to 62% of articles with a data-available statement; of 1,324 values in 35 articles, 64 did not reproduce within 10% |
measured |
| Stodden 2018 |
random sample, 204 papers in Science |
after the policy of sharing on request, artifacts were obtained for 44% and findings reproduced for 26% |
measured |
| Hamilton 2023 |
meta-analysis of 105 meta-research studies, 2,121,580 articles |
2016-2021: declared public data 8%, actual 2%; declared and actual public code both under 0.5% |
measured |
| Gabelica 2022 |
audit with author survey |
3,556 BioMed Central articles: 42% of statements say "on reasonable request"; of 1,792 authors willing to share, 93% did not reply or refused; 6.8% supplied data |
measured |
| Danchev 2021 |
cross-sectional, 487 trials |
JAMA, Lancet and NEJM after the ICMJE statement: 68.6% declared sharing, 0.6% of declared datasets were public |
measured |
14. Industry release practice
| OpenAI 2023 |
company report (owner's claim) |
the GPT-4 technical report says it gives no details on architecture, hardware, training compute, dataset construction or training method |
convention, untested |
| Bommasani 2023 |
external index |
10 developers scored on 100 disclosure indicators from public information: mean 37, highest 54, lowest 12 |
measured, descriptive |
| Bommasani 2024 |
follow-up |
after developers submitted their own reports, 14 developers averaged 58; indicators newly disclosed averaged 16.6 per developer; the authors say the index probably contributed |
supported with caveat |
| Loisel 2026 |
bibliometric preprint (abstract read) |
317 AI start-ups valued over a billion, 1998-2025: 1,389 peer-reviewed papers and 688 preprints with start-up authors; more than half (52.4%) have no qualifying output; about 0.1% of the 2025 AI literature |
supported with caveat |
| Rodriguez 2021 |
data brief (think-tank) |
universities produce more ML papers than most leading company labs; company papers draw more citations; the most cited are often joint |
supported with caveat |
| Gnewuch 2025 |
preprint, 49.8K papers at 10 AI venues (abstract read) |
industry-funded papers rose from under 2% (2015) to over 11% (2020); 12% of them reached high citation rates against 4% and 2% for the other two groups |
measured |
| Ahmed 2023 |
perspective |
industry's growing weight in AI research (one-line abstract; figures not read) |
convention, untested |
Not found by this search: a study of how paper-plus-blog-post releases are read, or a controlled comparison of how technical reports and peer-reviewed papers are cited.
15. AI in writing: policies
| Nature 2023 |
editorial |
no language-model tool is accepted as an author; use is documented in methods, acknowledgements or introduction |
convention, untested |
| ICMJE 2023 |
editorial standard |
chatbots are not authors; authors disclose AI-assisted technology at submission and describe it in the cover letter and the work; humans stay responsible. Added in 2023, in May by secondary accounts |
convention, untested |
| ACL 2023 |
venue policy (updated 2023-01-10) |
six cases for text, plus a note on code; language polishing needs no disclosure; low-novelty text must be flagged and checked; new text from a model is discouraged; disclosure goes through the Responsible NLP Research checklist |
convention, untested |
| ICML 2023 |
venue policy |
text generated by a language model is prohibited unless it is part of the experimental analysis; light editing is allowed; no automated checks, cases handled as plagiarism on complaint |
convention, untested |
| NeurIPS 2025 |
venue policy |
writing-only use needs no declaration; use that is important, original or non-standard goes in the methods; authors answer for everything; checks can come after publication. The text for 2023 was not found |
convention, untested |
| Bhavsar 2025 |
cross-sectional audit, 2023-09 to 2023-12 |
56 of 162 STM publishers (34.6%) had a public author policy; none allowed an AI author; 4 banned use outright; the abstract says 49 of 56 require disclosure and the text 46 of 56 |
measured, descriptive |
| ICLR 2025 |
venue blog (policy for the 2026 ICLR) |
use allowed with disclosure; substantial undisclosed use is desk rejected; fake references count as an ethics violation; detectors flag, a chair must find concrete evidence; the post gives no counts |
convention, untested |
| ICML 2026a |
venue policy |
two reviewer policies (A: no LLM, B: limited help); not about authors' prose |
convention, untested |
| arXiv 2025 |
platform blog (2025-10-31) |
computer science review and position papers must have passed peer review before posting; the reason given is a flood of AI-assisted surveys; the post gives no count |
convention, untested |
| Ganjavi 2024 |
cross-sectional audit, 2023-05 and 2023-10 |
guidance existed for 24% of the top 100 publishers and 87% of the top 100 journals; AI as author was prohibited in 96% and 98% of those; one journal banned its use for generating a manuscript; 12 journals conflicted with their publisher |
measured, descriptive |
16. AI in writing: measured use
| Liang 2025 |
corpus, 950,965 papers, 2020-01 to 2024-02 (numbers from the preprint) |
estimated share of LLM-modified abstracts reached 17.5% in computer science; mathematics and the Nature portfolio stayed at or under 6.3%; higher with frequent preprinting, crowded areas and shorter papers |
measured |
| Kobak 2025 |
corpus, over 15 million PubMed abstracts, 2010-2024 |
words that rose abruptly after 2022 imply at least 13.5% of 2024 abstracts were processed with a language model; up to 40% in some subcorpora; a lower bound |
measured |
| Markowitz 2024 |
experiments |
participants comprehended GPT-written significance statements better than the published ones in one experiment |
supported with caveat |
| ICML 2026b |
venue enforcement report (owner's account) |
hidden prompts in PDFs caught 795 reviews by 506 reviewers assigned the no-LLM policy; 497 papers (about 2%) desk rejected; the method catches careless use only |
measured, descriptive |
| Kim 2026 |
randomised experiment and survey (preprint, abstract read), reviewers of the 2026 ICML |
the policy each reviewer was given had near-zero effect on decisions; of 1,486 survey answers from reviewers and chairs, 22.5% under the ban said they used an LLM and 36.5% under the permissive policy used a disallowed form; the organisers are co-authors |
supported with caveat |
This search did not find a study that sets authors' declared language-model use against estimated use. The nearest, Kim 2026 and ICML 2026b, concern reviewers.