CoChat

Quantifying the prevalence and impact of overreaching causal claims in social science

Open access
Isch, Calvin, Dörr, Timothy, Fasching, Neil, Jennings, Grace, Watts, Duncan J. · 2026 · Nature Human Behaviour

The DOI matches its Crossref record; 100 of the 100 references it cites were found in a registry or an index; 18 of the 35 sources its claims cite were checked for support.

Abstract

Abstract Across the social sciences, many studies use cross-sectional designs that reveal associations but are generally unable to support direct causal claims, yet authors of such articles may make or imply causal claims anyway. Here, to examine the prevalence of such ‘overreaching’ causal language, we analysed 194,631 cross-sectional articles using large language models. Over the period 1980–2024, an average of 46% of articles contained causal language in their titles or abstracts, where the annual rate has risen almost threefold since 2000 from 20% to 60%. To examine the effects of such language, we conducted a human-subjects experiment ( N = 1, 105), finding that readers frequently indicate abstracts with this phrasing provide causal evidence but that methodological labels ( β = −0.4, 95% confidence interval −0.56 to −0.19) and associational wording ( β = −0.3, 95% confidence interval −0.43 to −0.07) reduce this tendency. Experiments with five LLMs revealed that model summaries of these articles ( N = 100 each) can amplify causal overstatement, removing hedges and introducing causal claims where articles used strictly associational phrasing; however, prompting caution diminishes this pattern.

Overview

CoChat Summary

TL;DR

Across 194,631 cross-sectional social-science articles, 46.3% used causal language in titles or abstracts, rising from about 20% in 2000 to over 60% by 2024. A preregistered human experiment (N = 1,105) and experiments with five LLMs show that readers and models often infer or amplify causality, though associational rewrites, methodological labels, and careful prompts reduce this tendency.

Key findings

  • Prevalence: 46.3% of 194,631 cross-sectional articles contained causal language in titles/abstracts; the annual rate rose almost threefold since 2000 from 20% to over 60%. By 2024, business was highest (84%) and sociology lowest (46%), with economics 67%, psychology 55%, and political science 53%. Higher-impact journals showed more causal language (54.4% above median impact factor vs 43.3% below); elite journals were 43.2%, and the Journal of Quantitative Description: Digital Media was 12.9%.
  • Human experiment: with N = 1,105, participants tended to agree that original abstracts provided causal evidence; associational rewrites and methodological labels reduced this (abstract reports β = -0.3, 95% CI -0.43 to -0.07 for associational wording; β = -0.4, 95% CI -0.56 to -0.19 for methodological labels). Causal language in participant summaries remained common across conditions: original 71.3%, rewrite 49.6%, note 64.2%, AI feedback 59.4%.
  • LLM summaries: GPT-4.1 summaries based on original abstracts almost always included causal claims (99.3%), reduced to 89.8% with rewrite, 41.7% with methodological note, and 30.3% with AI feedback (all P < 0.001). In a second experiment with five LLMs and 100 articles each, basic, simplifying, and practical prompts often produced more unhedged causal claims than the abstracts, while a “careful” prompt reduced causal language across all models except GPT-4.1 (mean 5.0%).
  • Full text: among 69,580 cross-sectional articles with full text, 81.8% contained at least one causal snippet. In a claim-level subsample (N = 100), 26.8% of matched claims were overreaching; unhedged causal claims appeared in 60.0% of introductions, 55.1% of discussions, 42.7% of results, and 6.1% of methods sections.

Methods

  • Descriptive analysis: keyword search of five ProQuest TDM Studio databases yielded 321,322 unique scholarly abstracts; GPT-4o prompts validated against 200 human-coded articles were used to train BERT classifiers (trained on 30,000, validated on 10,000), identifying 194,631 cross-sectional articles and coding causal language in titles/abstracts.
  • Human experiment: preregistered between-subjects study with 1,105 Prolific participants (US adults with at least a bachelor’s degree), randomly assigned to original abstract, associational rewrite, methodological note, or AI feedback conditions (plus an exploratory No Info condition); outcomes analyzed with multilevel linear and logistic models with random intercepts for article.
  • LLM experiments: first experiment presented the same stimuli to GPT-4.1; second experiment tested GPT-4.1, GPT-5.1, GPT-5.4, DeepSeek Chat-V3.2, and Claude Sonnet 4.6 under four summarization prompts, using 100 articles per model/prompt and comparing abstract-only versus full-text inputs with McNemar/Stuart-Maxwell tests and multilevel logistic regression.

Limitations

  • Stated: the study focuses on causal language as only one form of “narrative license”; cross-sectional designs are not the only ones susceptible to overclaiming; treating all causal claims from cross-sectional studies as overreach may ignore genuine causal signals; the human sample was college-educated and not practicing social scientists; experiments used abstracts rather than full texts; LLM results cover only a few models and prompts; full-text claim-level analysis at the sentence level can miss contextual nuance; classifier precision exceeded recall, likely underestimating causal language.
  • Evident: the descriptive corpus was a purposive keyword-based sample, not a random or representative sample of all social-science publications; ProQuest licensing prevents redistribution of underlying article texts; the LLM experiments were not preregistered; no demographic information beyond education and US residence was collected, and the Prolific sample was self-selected.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41562-026-02553-x

Cites

100 of 100 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

11 Valid · 2 Not valid · 5 Not checked · 234 Own results and interpretations

107 more claims

Create a free account to see all 252 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 109
  • It may state two separate claims as one. 63
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Consistentz = -2.1reported p = .039 · recomputed p = .036

    “Z = -2.1, p = .039”

  • Consistentz = -6.73reported p < .001 · recomputed p < .001

    “Z = -6.73, p < .001”

  • Consistentz = -22.2reported p < .001 · recomputed p < .001

    “z = -22.2, p < .001”

  • Consistentt(217) = 2.17reported p = .031 · recomputed p = .031

    “t (217) = 2.17, p = .031”

  • Consistentχ²(6) = 81reported p < .001 · recomputed p < .001

    “χ 2 (6) = 81.0, p < .001”

5 of 5

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 100 references

Research works printed but not linked: 6

  1. Wooldridge, J. M. Econometric Analysis of Cross Section and Panel Data (MIT Press, 2010).
  2. Hewitt, L., Ashokkumar, A., Ghezae, I. & Willer, R. Predicting results of social science experiments using large language models. Preprint at Samim.io https://samim.io/dl/Predicting%20results%20of%20social%20science%20experiments%20using%20large%20language%20models.pdf (2024).
  3. Bleske-Rechek, A., Gunseor, M. M. & Maly, J. R. Does the language fit the evidence? Unwarranted causal language in psychological scientists’ scholarly work. Behav. Ther. 41, 341–352 (2018).
  4. Rawat, S. & Meena, S. Publish or perish: where are we heading? J. Res. Med. Sci. 19, 87 (2014).
  5. Important Notice No. 121: New Criteria for NSF Proposals Announces Adoption of the Two Merit Review Criteria: Intellectual Merit and Broader Impacts (National Science Foundation, 1997).
  6. Bates, D. et al. Package ‘lme4’. Convergence 12, 2 (2015).

Metadata

Authors
Isch, C., Dörr, T., Fasching, N., Jennings, G., & Watts, D. J.
Affiliations
University of Pennsylvania (US)
Venue
Nature Human Behaviour
Published
2026-08-24
ISSN
2397-3374 (ISSN-L)
OpenAlex
W7204139589
Cited by
0
References
84
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Human Behaviour (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • Mason | Institute for Humane Studies, George Mason University (IHS020582)
  • United States Department of Defense | Defense Advanced Research Projects Agency (HR00112520300)
Topics
  • Social Sciences › Decision Sciences › Statistics, Probability and Uncertainty › Meta-analysis and systematic reviews
  • Social Sciences › Library and Information Sciences › scientometrics and bibliometrics research
  • Social Sciences › General Social Sciences › Computational and Text Analysis Methods
Keywords
causal language, cross-section design, association versus causation, research reporting practices, abstracts, large language models, hedges

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Quantifying the prevalence and impact of overreaching causal claims in social science. https://app.cochat.ai/papers/10.1038/s41562-026-02553-x
The paper
Isch, C., Dörr, T., Fasching, N., Jennings, G., & Watts, D. J. (2026). Quantifying the prevalence and impact of overreaching causal claims in social science. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02553-x

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.