Against

This file lists the evidence against the standard paper shape, by theme, each with its source.

1. IMRaD hides the real process

Medawar 1963 argued that the paper misrepresents discovery: it narrates an inductive march from facts to conclusion, while the work ran from a guess to its test (revisited by Howitt 2014). Grade: supported with caveat (an argument, not a measurement).
ICMJE asserts the opposite, calling IMRAD "a reflection of the process of scientific discovery" (ICMJE 2026). Grade: convention, untested.
Historians of the genre find that its form grew out of rhetorical and social pressures over three centuries (Bazerman 1988, Atkinson 1999). Grade: supported with caveat.
Kerr 1998 names HARKing, a post hoc hypothesis written up as a priori, and reports survey data that some forms are widely practised. Grade: supported with caveat.
The standard literature is filtered: 96% of first hypotheses supported, against 44% in Registered Reports (Scheel 2021). Grade: measured.
Structured sections still carry spin: in 72 trials with a null primary outcome, 58% of abstract conclusions had spin (Boutron 2010). Grade: measured.

2. Structure does not guarantee readability

Readability of abstracts fell from 1881 to 2015, including the decades in which IMRaD and structured abstracts spread (Plavén-Sigray 2017). Grade: measured as a trend; no causal link to structure is claimed.
The same authors note that readability formulas ignore layout, coherence and the complexity of ideas. Grade: supported with caveat.
Abstracts disagree with their own articles: 18% to 68% were deficient across six medical journals (Pitkin 1999). Grade: measured.
Fewer than 10% of the articles surveyed (about 1,000, of which 500 were scored) met her three criteria of engaging, accessible prose in Sword 2012 (secondary). Grade: supported with caveat.
About half of the common style advice is associated with fewer citations, as associations and not effects (Weinberger 2015). Grade: supported with caveat.
IMRaD is not the default in many fields, so the advice to follow the standard structure holds mainly for biomedicine (Lin 2012, Moskovitz 2024). Grade: measured.

3. Structured abstracts: benefits mixed or field-dependent

Length: structured abstracts run 21% to 35% longer, so part of each gain may come from more words (Hartley 2004). Grade: measured.
Recall: the advantage disappeared once length and readability were equated (Hartley 2004, study 18). Grade: measured.
Search speed: an advantage with simplified abstracts, none with real MEDLINE abstracts (Hartley 2004). Grade: measured.
Retrieval by labelled field raised precision and cut recall (Booth 1997, full text). The authors conclude higher precision with the same or lower recall in three of five questions per database, and one question per database retrieved nothing (p. 165); their Table 4 shows higher precision in only two of five (B1, B2). Grade: measured.
Readability: no formula difference when journals switched format; higher scores only when authors rewrote (Hartley 1997). Grade: measured.
Accuracy: no better than traditional abstracts; psychology abstracts held few errors, unlike medicine, which the author ties to the discipline (Hartley 2000). Grade: measured.
Most early studies used abstracts rewritten by the investigators and undergraduates as judges (Hartley 2014). Grade: supported with caveat.
Results-free review suited experimental designs, drew no qualitative or interpretive submissions, and left null results hard to interpret (Findley 2016). Grade: supported with caveat.

4. Visual abstracts and highlights

In one journal, papers with a graphical abstract had fewer downloads, views and citations (Pferschy-Wenzig 2016). Grade: supported with caveat (observational).
The 7.7-fold impressions came from a study whose authors sat on the journal's editorial board, against title-only tweets (Ibrahim 2017). Grade: supported with caveat.
Public engagement stayed low whatever the format (Chapman 2019). Grade: measured.
Highlights exist for search engines by the publisher's own account (Elsevier highlights, example of practice only, 9.0/20); no reader benefit was found. Grade: convention, untested.
The Elsevier graphical-abstract page states aims (a quick takeaway, discovery) and gives no outcome data (Elsevier GA 2026, example of practice only, 9.0/20). Grade: convention, untested.

5. The bundle alternatives

Notebooks as papers: of 863,878 re-runs of Python notebooks with a valid execution order, 24.11% completed and 4.03% reproduced the stored results (Pimentel 2019). Grade: measured.
Distill paused in 2021: its editors found running an interactive journal unsustainable for volunteers (Distill 2021). Grade: supported with caveat.
Nanopublications: we found a model and its infrastructure (Groth 2010), no measure of reader benefit.

6. Reading advice is untested

The three-pass method (Keshav 2007) and the ten rules for structure (Mensh 2017) are advice from experience. Grade: convention, untested.

7. Checklists are self-reported and lightly used

Of authors who said they had code, reviewers confirmed 71.5% (Pineau 2021). Grade: measured.
34% of 2019 NeurIPS reviewers found the checklist useful; the share who read it is unknown (Pineau 2021). Grade: measured, descriptive.
Submissions that answer yes to every item are accepted less than expected, which the authors read as inaccurate answers (Magnusson 2023). Grade: supported with caveat.
Reporting rates were flat across the year after the first gain; the before data are 50 papers checked by hand against later self-reports (Magnusson 2023). Grade: supported with caveat.
The organisers of the NeurIPS programme state they have no evidence of a quality effect (Pineau 2021). Grade: supported with caveat.
Broader impact statements were dropped after one year (Ashurst 2022), and their content was uneven (Nanayakkara 2021). Grade: supported with caveat.
Checklists catch missing items, not wrong methods: leakage affected 294 papers across 17 fields (Kapoor 2023). Grade: measured.
Outside CONSORT, endorsing a guideline did not significantly raise completeness on nearly every item: 26 evaluations of nine of 101 guidelines (Stevens 2014). Grade: measured.
Across 56,800 papers at five AI venues, code and data sharing rose from 11% to 64% in 2014-2024, and the rise began before the checklists (Coakley 2026, preprint). Grade: supported with caveat.
Only 11 of 31 interventions to raise adherence to reporting guidelines had been evaluated (Blanco 2019). Grade: measured, descriptive.

8. Artifacts and badges

Artifacts did not raise citations in a decade of evaluations (Winter 2022). Grade: measured.
Committees disagree on what an artifact must meet (Hermann 2020), and creators and reviewers see little reward (Timperley 2021). Grade: supported with caveat.
In 2017 only 58.5% of ICSE papers that built an artifact released it (Heumüller 2020). Grade: measured.
The 32.3% quick-build rate (the technical report behind Collberg 2016, secondary) was disputed in public by authors of papers it classed as not repeatable (secondary). Grade: supported with caveat.

9. Documentation cards

Environmental impact is filled in 2.0% of model cards and limitations 17.4% (Liang 2024b). Grade: measured, descriptive.
The 29.0% download gain appeared in one of two batches (42 models in all); the other batch was null (Liang 2024b). Grade: supported with caveat.
Two authors of Liang 2024b work at the hub it studies, and the preprint gives 44.2% and 42.8% for the share of repositories with cards in different places. Grade: supported with caveat.
The datasheet test has 23 participants and one dataset (Boyd 2021). Grade: supported with caveat.

10. Plain-language formats

None of 1,738 Cochrane plain language summaries met every measured standard (Jelicic Kadic 2016). Grade: measured.
Their readability stays far above the recommended level (Karačić 2019). Grade: measured.
An infographic and a text summary gave the same knowledge in three trials (Buljan 2018). Grade: measured.
Bredbenner 2019 states that abstracts are less successful at producing comprehension, while its own results show all groups did well and preference did not track comprehension. Grade: supported with caveat.
Significance statements differ little from abstracts (readability d = 0.036, Markowitz 2024). Grade: measured.
The eLife reader survey is the publisher's own and self-selected (eLife 2017). Grade: convention, untested.
Cell Press and Elsevier pages describing STAR Methods are the owner's claim; this search found no independent evaluation. Grade: convention, untested.
The Elsevier page on the key resources table offers one anecdote and no study (Elsevier KRT 2026, example of practice only, 9.5/20). Grade: convention, untested.
Across 90 articles on plain-language summaries, empirical evidence on criteria for a good one is scarce (Stoll 2022). Grade: supported with caveat.

11. Availability statements

0.6% of declared clinical-trial datasets were public in three leading journals (Danchev 2021). Grade: measured.
Even where data were reusable, 64 of 1,324 values did not reproduce within 10% (Hardwicke 2018). Grade: measured.
Request-based sharing worked for 6.8% to 44% depending on the study (Gabelica 2022, Stodden 2018). Grade: measured.
Only eight of 105 meta-research studies had low risk of bias (Hamilton 2023). Grade: measured.

12. Industry release and AI-use estimates

A company report is its owner's account; the GPT-4 report states what it withholds (OpenAI 2023). Grade: convention, untested.
The transparency index is run by academics who score companies, and the follow-up score rose partly because developers sent the information (Bommasani 2024). Grade: supported with caveat.
The company-versus-university brief gives no counts (Rodriguez 2021), and the funding study is an abstract-level source here (Gnewuch 2025). Grade: supported with caveat.
The excess-vocabulary estimate is a lower bound and counts abstracts processed, not authors' intent (Kobak 2025). Grade: measured.
The mixture-model estimate depends on reference texts and covers preprints and one publisher (Liang 2025). Grade: supported with caveat.
The hidden-prompt method flagged 506 reviewers, and the organisers say it is easy to evade and catches only careless use (ICML 2026b). Grade: measured, descriptive.
Under a ban, 22.5% of surveyed reviewers said they used an LLM (Kim 2026, preprint, organisers among the authors). Grade: supported with caveat.