CoChat

Evaluation of Large Language Model Performance and Reliability for Citations and References in Scholarly Writing: Cross-Disciplinary Study

Open access
Mugaanyi, Joseph, Cai, Liuying, Cheng, Sumei, Lu, Caide, Huang, Jing · 2024 · Journal of Medical Internet Research

Published version of a preprint

The DOI matches its Crossref record; 18 of the 18 references it cites were found in a registry or an index; 10 of the 13 sources its claims cite were checked for support.

Abstract

BACKGROUND: Large language models (LLMs) have gained prominence since the release of ChatGPT in late 2022. OBJECTIVE: The aim of this study was to assess the accuracy of citations and references generated by ChatGPT (GPT-3.5) in two distinct academic domains: the natural sciences and humanities. METHODS: Two researchers independently prompted ChatGPT to write an introduction section for a manuscript and include citations; they then evaluated the accuracy of the citations and Digital Object Identifiers (DOIs). Results were compared between the two disciplines. RESULTS: Ten topics were included, including 5 in the natural sciences and 5 in the humanities. A total of 102 citations were generated, with 55 in the natural sciences and 47 in the humanities. Among these, 40 citations (72.7%) in the natural sciences and 36 citations (76.6%) in the humanities were confirmed to exist (P=.42). There were significant disparities found in DOI presence in the natural sciences (39/55, 70.9%) and the humanities (18/47, 38.3%), along with significant differences in accuracy between the two disciplines (18/55, 32.7% vs 4/47, 8.5%). DOI hallucination was more prevalent in the humanities (42/55, 89.4%). The Levenshtein distance was significantly higher in the humanities than in the natural sciences, reflecting the lower DOI accuracy. CONCLUSIONS: ChatGPT's performance in generating citations and references varies across disciplines. Differences in DOI standards and disciplinary nuances contribute to performance variations. Researchers should consider the strengths and limitations of artificial intelligence writing tools with respect to citation accuracy. The use of domain-specific models may enhance accuracy.

Overview

CoChat Summary

TL;DR

In a cross-disciplinary test, ChatGPT (GPT-3.5) generated 102 citations across 10 natural sciences and humanities topics; citation existence, accuracy, and relevance were similar across domains, but DOI existence and accuracy were significantly worse in the humanities, with more DOI hallucination. The authors conclude that LLM citation reliability varies by discipline and requires validation.

Key findings

  • Across 10 topics (5 natural sciences, 5 humanities), ChatGPT generated 102 citations: 55 in natural sciences and 47 in humanities. Citation existence was 40/55 (72.7%) in natural sciences vs 36/47 (76.6%) in humanities (P=.42); citation accuracy was 37/55 (67.3%) vs 29/47 (61.7%) (P=.35); relevance was 39/55 (70.9%) vs 35/47 (74.5%) (P=.43).
  • DOI existence differed significantly: 39/55 (70.9%) in natural sciences vs 18/47 (38.3%) in humanities (P=.001). DOI accuracy was 18/55 (32.7%) vs 4/47 (8.5%) (P=.003).
  • DOI hallucination was more prevalent in the humanities: Table 2 reports 42/47 (89.4%) in humanities vs 34/55 (61.8%) in natural sciences (P=.001). The abstract states “42/55, 89.4%” for humanities, which conflicts with Table 2’s denominator.
  • Levenshtein distance was reported as significantly different (P=.009), but the direction is inconsistent: Table 2 and the Results text report a higher mean in natural sciences (64.13, SD 42.26) than humanities (42.15, SD 40.23), whereas the abstract says it was higher in the humanities.

Methods

  • Two researchers independently prompted ChatGPT (GPT-3.5) to write introduction sections for 10 English-language topics (5 clinical/biomedical natural sciences; 5 philosophy/psychology humanities), adhering to APA style and including DOIs, between July 10 and August 15, 2023.
  • References were validated via Google Scholar, PubMed, and Google Search; DOI existence and validity were checked using the DOI Foundation website. Hallucination was defined as a nonexistent DOI/citation or a DOI linked to a different source; disagreements were resolved by discussion and consensus.
  • Analysis used SPSS 26 and Python; Levenshtein distance was calculated with the thefuzz package. Continuous variables were reported as mean (SD), categorical variables as n (%), with independent-sample t tests and Fisher exact tests; P<.05 was significant.

Limitations

  • Stated: only 10 topics were included, limiting coverage of disciplinary complexity; only ChatGPT 3.5 was tested, so newer models (GPT-4, Claude+, Gemini) may differ; the study did not assess user feedback or subjective quality/coherence of generated content.
  • Evident from the text: the abstract and Table 2/Results disagree on the direction of the Levenshtein distance finding and on the denominator for humanities DOI hallucination, which complicates interpretation.

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

Integrity signals

Registry

Crossref · 10.2196/52935

Cites

18 of 18 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Transformer model architecture (left) and GPT architecture (right).
Figure 1 Transformer model architecture (left) and GPT architecture (right).CC BY · PubMed Central

Verification

Claims the paper makes

2 Valid · 3 Not valid · 1 Not checked · 25 Own results and interpretations

8 more claims

Create a free account to see all 31 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 11
  • Probably not a substantial claim. 7
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

14 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

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.

3 of 3

Citations

Cites

Research works printed but not linked: 4

  1. Vaswani, A, arXiv
  2. Bahdanau, D, arXiv
  3. DOI Foundation2024-04-01https://doi.org/
  4. Zhang, Y, arXiv

Cited by

See all 74

Metadata

Authors
Mugaanyi, J., Cai, L., Cheng, S., Lu, C., & Huang, J.
Affiliations
  • Ningbo University (CN)
  • Shanghai Academy of Social Sciences (CN)
Corresponding author
Huang, Jing
Venue
Journal of Medical Internet Research
Published
2024-04-05
Volume
26
Pages
e52935
ISSN
1438-8871 (ISSN-L)
OpenAlex
W4392710953
Cited by
74
References
17
Citations by year
2026: 45 · 2025: 21 · 2024: 8
FWCI
109.16
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Journal of Medical Internet Research (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Topics
  • Social Sciences › Library and Information Sciences › Academic Writing and Publishing
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Social Sciences › Library and Information Sciences › scientometrics and bibliometrics research
Keywords
large language models, ChatGPT, citation accuracy, reference generation, digital object identifiers, hallucinations, academic writing, Levenshtein distance, AI writing tools, cross-disciplinary comparison
MeSH terms
Artificial Intelligence, Humans, Language, Research Personnel, Writing, Reproducibility of Results

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 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (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

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.