Logical-fallacy detection in scientific writing

Last updated: 2026-05-17

Synthesis

A fallacy is an error in reasoning that produces an argument whose conclusion does not in fact follow from its premises, even though the inference may look plausible. Logicians distinguish formal fallacies — invalid by virtue of their syntactic structure (e.g. affirming the consequent) — from informal fallacies, whose problems lie in the content, context, or pragmatic use of the argument (e.g. appeal to authority, false dichotomy, hasty generalization). Scientific manuscripts overwhelmingly fail in informal ways: a paper’s syntactic logic is rarely the issue; its inferential warrants almost always are.

Two empirical bodies of work matter here. First, the “spin” literature (Boutron et al. 2010 ¹ and follow-ons) documents that interpretive fallacies — overstating effects, conflating significance with importance, attributing causation from correlation — appear in roughly half of RCT discussion sections even when the underlying analysis is competently executed. Second, the methodological-critique literature (Kerr 1998 on HARKing ²; Nuzzo 2014 on p-value misunderstanding ³; the 2016 ASA statement on p-values) documents that several specific fallacies have become so endemic that they are better treated as reportable patterns than as occasional lapses.

For scriptorium, the practical implication is that fallacy detection is pattern matching against a small canonical list, not an attempt at general-purpose logical analysis. The list is short — perhaps twelve to twenty patterns account for the vast majority of in-science inferential errors — and each pattern has both a definition and a canonical surface form an LLM can flag. See argument-mapping for the structural underpinnings and citation-claim-alignment for the closely related problem of overclaim through citation.

Techniques and tools

The canonical taxonomies

• Bo Bennett, Logically Fallacious (2012 / 2021 Academic Edition, ISBN 9781456607524 [hardback]) ⁴. The book and the companion site https://www.logicallyfallacious.com/ document ~300 fallacies with definition, example, and logical form. The site is the most comprehensive open inventory; the book imposes a usable taxonomy.
• Wikipedia “List of fallacies” — a community-maintained inventory organised by formal / informal / pragmatic categories; useful as a cross-check, less rigorous than Bennett.
• Walton, Reed & Macagno Argumentation Schemes (see argument-mapping) — frames each “fallacy” as a scheme misuse paired with a known critical-question audit. This is the most productive frame for pattern detection because it forces the fallacy to point at a missing answer.

The high-frequency in-science fallacies

The patterns below are the ones that surface most often in biomedical peer review and post-publication critique. Each has a recognizable surface form an automated check can match against.

1. Post hoc ergo propter hoc / correlation as causation. Inferring that A causes B from the observation that A precedes B, or from a correlation in observational data. The default fallback in observational epidemiology and increasingly in single-cell / association studies. Surface form: “X was associated with Y, suggesting that X drives / causes / leads to Y.”

2. Affirming the consequent. Formal fallacy: from “if H then D” and “D” inferring “H.” In science, frequently dressed as “our model predicted this pattern; we observed this pattern; therefore our model is correct.” Without an alternative-hypothesis check, this is a fallacy.

3. False dichotomy. Presenting two options as exhaustive when more exist. Common in discussion sections: “either X mechanism or Y mechanism.”

4. Hasty generalisation. Drawing a population-level conclusion from a small or non-representative sample. Surface form: small-N or single-site studies that conclude “this is true of [broad population].”

5. Appeal to authority. Citing a senior researcher or major journal as warrant rather than the underlying data. Walton’s critical questions for “argument from expert opinion” make the audit concrete (see argument-mapping). Note: in science, citation of established results is not an automatic appeal-to-authority fallacy — the test is whether the citation does inferential work the cited evidence can support.

6. Base-rate neglect. Ignoring prior probability when interpreting a positive test result or association. Endemic in diagnostic accuracy reporting, screening papers, and rare-disease research. Nuzzo (2014) ³ gives the canonical worked example: a p<0.05 result on a low-prior hypothesis still has high posterior probability of being a false positive.

7. p-hacking as reasoning. Inferring an effect by exploiting analytic flexibility — multiple outcomes, multiple subgroups, multiple covariate adjustments — until p < 0.05 ³. The cognitive move that makes this a fallacy (and not merely a procedural error) is treating the surviving result as if it had been pre-specified.

8. HARKing — Hypothesizing After the Results are Known. Kerr 1998 ² defined HARKing as “presenting a post hoc hypothesis (i.e., one based on or informed by one’s results) in one’s research report as if it were, in fact, an a priori hypothesis.” Kerr’s survey found HARKing widely practiced and widely viewed as inappropriate. HARKing is the upstream cause of many spurious “novel” findings and is detectable from internal evidence — see internal-consistency and Yarkoni 2019 ⁵.

9. Overgeneralisation from a model organism / model system. Inferring human relevance from cells in a dish, a single mouse strain, or a single cohort.

10. Equivocation / shifting definitions. Using a key term in one sense in the methods and a different sense in the conclusion. Particularly common with terms like “significant,” “validated,” “associated.”

11. Composition / division. Inferring properties of the whole from parts, or vice versa. Surface form: aggregating subgroup effects into a population-level claim without test for heterogeneity.

12. Survivorship / selection bias as reasoning. Drawing a conclusion from a sample defined by an outcome rather than by enrolment, then interpreting the sample as representative.

Spin as a special class

Boutron et al. (2010) ¹ analysed 72 RCTs with nonsignificant primary outcomes published in December 2006 and found spin — the systematic distortion of interpretation to favour the experimental treatment — in 18.0% of titles, 37.5% of abstract Results sections, 58.3% of abstract Conclusions, and 50.0% of main-text Conclusions. The detailed spin taxonomy (focus on within-group analyses, focus on secondary outcomes, focus on subgroup analyses, downplay of nonsignificance) is in effect a catalogue of interpretive fallacies specific to clinical trials. Yavchitz et al. (2012) showed that this spin propagates linearly into press releases and news coverage. See citation-overreach-research for the full citation chain.

Cognitive-bias literature

Kahneman’s Thinking, Fast and Slow (2011) and the broader heuristics-and-biases programme are the substrate behind several of the above patterns: anchoring (the first number reported sets the frame), availability (vivid examples dominate inference), and confirmation bias (researchers preferentially see evidence consistent with their hypothesis). Mlodinow’s The Drunkard’s Walk (2008) is the canonical popularisation of base-rate neglect. The substantive evidence base for “researchers exhibit the same biases as anyone else” is robust enough that scriptorium should treat this as background, not a finding to be debated.

How this informs scriptorium

• reviewer-simulation can be primed with the 12-pattern list above as an explicit checklist. Each persona walks the conclusions section against the list and reports either “no instance flagged” or “instance flagged at [passage], pattern: [name], reason: [why].”
• argumentative-flow can detect surface forms of correlation→ cause and overgeneralisation through targeted prompts; high-recall, modest-precision is the realistic operating point.
• A future spin-audit could implement the Boutron taxonomy on abstract Conclusions sections; this is a narrow, high-value target because spin is both prevalent and identifiable from text alone.
• Make the patterns named, not vibes: every flagged fallacy should cite the pattern by name (and ideally a Bennett / Walton reference) so the author can argue back if the LLM is wrong.

LLM limits — be honest

• LLMs detect surface forms reliably but cannot reliably distinguish between a valid appeal to expertise (a domain-appropriate citation doing real work) and a fallacious appeal to authority. This is a warrant-validity question, not a textual one.
• HARKing detection from a single manuscript is partial at best (see internal-consistency); full detection requires preregistration comparison.
• Base-rate fallacies require numerical reasoning the model often hallucinates. For prevalence/PPV calculations the skill should call out to a script (see statistical-inconsistency) rather than recompute in-band.

Limits and caveats

• The fallacy list above is prescriptive, not exhaustive. Several authors (notably Boudry) have argued that the fallacy-theoretic frame itself is overused — that what looks like an appeal-to- authority is often a legitimate appeal to consensus, and that fallacy spotting can substitute for engagement with substance.
• Distinguishing a fallacy from a defeasible-but-permissible inference requires context. Scriptorium critiques should always flag with hedging (“possible appeal to authority — does the cited paper actually support the claim?”) rather than verdict (“appeal to authority”).
• The “spin” taxonomy is well-established for RCTs and less well- developed for observational studies, basic-science papers, and computational methods papers. Adapting it to those genres remains ongoing methodological work.

References

Footnotes

1. Boutron I, Dutton S, Ravaud P, Altman DG. Reporting and Interpretation of Randomized Controlled Trials With Statistically Nonsignificant Results for Primary Outcomes. JAMA. 2010; 303(20):2058–2064. DOI: 10.1001/jama.2010.651. PMID: 20501928. ↩ ↩²

2. Kerr NL. HARKing: Hypothesizing After the Results are Known. Personality and Social Psychology Review. 1998; 2(3):196–217. DOI: 10.1207/s15327957pspr0203_4. Definition: “presenting a post hoc hypothesis (i.e., one based on or informed by one’s results) in one’s research report as if it were, in fact, an a priori hypothesis.” Note: the original prompt referenced DOI suffix 0204_1; the correct DOI suffix is 0203_4 (verified). ↩ ↩²

3. Nuzzo R. Scientific method: statistical errors. Nature. 2014; 506(7487):150–152. DOI: 10.1038/506150a. ↩ ↩² ↩³

4. Bennett B. Logically Fallacious: The Ultimate Collection of Over 300 Logical Fallacies (Academic Edition). eBookIt.com, 2012; expanded 2021. ISBN 9781456607524 (Academic Edition hardback). Companion site: https://www.logicallyfallacious.com/. ↩

5. Yarkoni T. The generalizability crisis. Behavioral and Brain Sciences. 2019 (published online 2020); 45:e1. DOI: 10.1017/S0140525X20001685. ↩