Source-linked AI summary
GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models
Zuyao Xu, Yuqi Qiu, Lu Sun, Fasheng Miao, Fubin Wu, Xiang Li, Xinyi Wang, Haozhe Lu, Zhengze Zhang, Yuxin Hu, Jialu Li, Luo Jin, Feng Zhang, Rui Luo, Xinran Liu, Yingxian Li, Jiaji Liu
TL;DR
Invalid or fabricated citations threaten the trust that supports scientific claims. The paper develops CITEVERIFIER and combines benchmarking, literature analysis, and a researcher survey to measure this threat, finding domain-sensitive hallucination, weak validation performance, and a systemic integrity risk.
Problem
Ghost citations threaten citation-based trust, while the scale of fabrication, scalable verification, and pathway into published literature remain insufficiently measured.
Method
The paper develops CITEVERIFIER and applies it across LLM benchmarking, large-scale published-literature auditing, and a researcher user study.
Results
The three studies document widespread, domain-sensitive citation hallucination, poor LLM validation performance, and invalid citations in published literature.
Takeaways & Limitations
Ghost citations constitute a systemic threat because fabricated references can enter and persist in the literature through sparse, largely trust-based verification practices.
Takeaways & Limitations
The verification pipeline checks title similarity, which may undercount hallucinations resembling real papers and flag legitimate but poorly indexed works.
Abstract
from arXiv · showhide
Citations provide the basis for trusting scientific claims; when they are invalid or fabricated, this trust collapses. With the advent of Large Language Models (LLMs), this risk has intensified: LLMs are increasingly used for academic writing, but their tendency to fabricate citations (``ghost citations'') poses a systemic threat to citation validity. To quantify this threat, we develop \citeb, an open-source framework for large-scale citation verification, and conduct a comprehensive study of citation validity in the LLM era through three complementary experiments. First, we benchmark 13 LLMs on citation generation task in various research domains, finding that all models hallucinate citations at rate from 14.23\% to 94.93\%. Second, we analyze 2.2 million citations from 56,381 papers at AI/ML and Security venues (2020--2025), finding that 1.07\% of papers contain invalid citations, with an 80.9\% increase in 2025. Third, we survey 97 researchers, finding that 87.2\% use AI-powered tools in their workflows, 76.7\% of reviewers do not thoroughly check references, and 74.5\% view peer review as ineffective at catching citation errors. Based on these findings, we argue that ghost citations represent a systemic threat to academic integrity, and call for coordinated efforts from community to address this challenge.
I. INTRODUCTION
Ghost citations exploit trust in scholarly references by fabricating plausible but nonexistent prior work. GhostCite addresses three measurement gaps with a verification framework and complementary analyses of LLM outputs, published literature, and researcher practices.
- Three unanswered questions concern LLM hallucination frequency, invalid citations in published literature, and failures of authors and reviewers to detect them.
- The study develops CITEVERIFIER, an open-source framework for large-scale citation verification, and applies it across three complementary experiments.
- 13 LLMs across 40 research domains produced citation hallucination rates from 14.23% to 94.93%.
- Analysis of 2.2 million citations from 56,381 papers identified invalid citations in 604 papers and an 80.9% increase in invalid citation rates in 2025.
- Among 97 surveyed researchers, 41.5% of authors reported copying BibTeX entries without checking, while 76.7% of reviewers did not thoroughly inspect references.
- LLM-generated references differ from earlier citation errors because autoregressive token prediction can systematically combine plausible author, venue, and terminology patterns into nonexistent citations.
B. The Risk to Scientific Research
Ghost citations threaten scholarly communication because they can enter the published record through sparse, trust-based verification. The paper frames this risk through unresolved detection, prevalence, and behavioral questions and addresses them with technical, empirical, and behavioral analysis.
- Ghost citations undermine researchers’ ability to ground claims in verifiable prior work and navigate the literature confidently.
- Large-scale detection is difficult because citation formats are heterogeneous, bibliographic indices are incomplete, and standardized verification protocols are absent.
- The paper reports no comparable systematic measurement of ghost-citation prevalence and describes existing evidence as largely informal.
- CITEVERIFIER combines robust parsing, cascaded multi-source retrieval, and calibrated similarity matching for high-throughput citation validation.
- The study pairs an LLM hallucination benchmark and published-paper audit with a user study of behavioral factors affecting citation verification.
B. Framework Implementation
CITEVERIFIER is a modular, cascaded pipeline that parses citation strings, searches multiple sources, and classifies matches by title similarity. Its experiments use this framework to assess citation validity across generated and published references.
- Framework Implementation: CITEVERIFIER implements three stages: reference parsing, cascaded verification, and similarity-based classification.
- Framework Implementation: Reference parsing uses GROBID first and an LLM-based reparser when initial parsing fails.
- Framework Implementation: Cascaded verification checks a local cache before querying multiple bibliographic databases in sequence.
- Framework Implementation: The classifier compares normalized citation titles with retrieved candidates using Levenshtein similarity and threshold θ = 0.9.
- Experiment Design: The experimental design includes an LLM benchmark, archival analysis, and user study targeting generated citations, published-paper prevalence, and researcher verification practices.
B. Experiment II: Archival Analysis
The archival analysis applies CITEVERIFIER to papers from AI/ML and Security venues, then manually verifies flagged citations. It measures invalid-citation prevalence and examines how citation reliability relates to verification practices.
- Archival Analysis: 56,381 papers from eight AI/ML and Security venues spanning 2020–2025 were analyzed with CITEVERIFIER.
- Archival Analysis: Citations with similarity scores below θ = 0.9 were automatically flagged for manual review.
- Archival Analysis: Sixteen research assistants independently checked flagged citations and classified them as non-academic sources, valid, or invalid.
- Archival Analysis: A sample of 400 citations from the valid pool yielded no additional invalid citations, supporting the false-negative assessment.
- Related Benchmark Results: The LLM benchmark found hallucination rates from 14.23% to 94.93%, while online search, chain-of-thought prompting, and batch size showed no consistent effect across models.
B. Domain Sensitivity
Citation hallucination varies sharply across both models and research domains, so aggregate reliability does not guarantee dependable performance in any specific area. LLMs also struggle to validate citations, limiting their ability to self-correct fabricated references.
- Domain Sensitivity: 2.6% versus 52.5%: DeepSeek’s hallucination rate varies markedly between computer vision and operating systems or hardware-related domains.Within hardware architecture, Grok 4 reaches 99.2% while DeepSeek remains at 8.6%; in digital libraries, GPT-5 and Claude 4 reach 93.0% and 97.2%.
- Domain Sensitivity: 14.23–94.93%: all 13 LLMs hallucinate citations, with substantial variation across models and domains.The best and worst performers differ by 6.7×, and no model is uniformly reliable across domains.
- Generated Citation Patterns: 27.61% to 98.75%: hallucination rates for generated citations rise with publication year from 2000 to 2025.The relationship between publication year and hallucination rate is fit by an exponential function with R2 = 0.94.
- Generated Citation Patterns: Well-known titles such as “NeRF” and “RAG” recur across runs, whereas less prominent works are more likely to be fabricated.This pattern characterizes the structure of generated references beyond aggregate hallucination rates.
- LLMs as Citation Judges: 38% average accuracy: LLMs perform below random guessing when validating citation validity.Only ERNIE exceeds 50%, reaching 56% by aggressively labeling citations invalid and misclassifying 88% of valid citations.
VI. HUNTING FOR GHOSTS: ARCHIVAL ANALYSIS
The archival analysis applies scalable citation verification and manual review to papers from eight AI/ML and Security venues. It finds that invalid citations include both metadata errors and untraceable ghost citations, with examples of fabrication and repeated propagation.
- Detection and Dataset: 2,199,409 citations from 56,381 papers were extracted, and 2,530 potential cases were flagged for manual review after OCR filtering.The corpus covers eight AI/ML and Security venues from 2020 to 2025.
- Verification Procedure: Manual verification searched bibliographic databases, checked metadata for located records, and classified untraceable entries as invalid.Two experts subsequently reviewed all invalid citations for accuracy.
- Classification: 739 citations were confirmed invalid, comprising 136 metadata-error citations and 603 ghost citations.Invalid citations were defined as incorrect metadata or untraceable citations that could not be found in major academic databases.
- Overall Findings: 1.07% of papers contained at least one invalid citation, including 0.24% with error citations and 0.86% with ghost citations.The analysis identified 604 affected papers, with 15 containing both types.
- Case Studies: A fabricated citation in a NeurIPS 2025 paper used generic author names and a plausible-sounding journal title but was absent from major bibliographic databases.The case exhibited characteristics associated with LLM-generated references.
- Case Studies: An erroneous AugMix title appeared in 16 papers, illustrating how incorrect citation metadata can propagate through the literature.The authors traced the repeated error to an incorrect title returned by OpenReview’s citation function.
C. Overall Distribution of Invalid Citations
Invalid citations occur across all analyzed research communities rather than being confined to a single venue or domain. Their distribution is broadly similar by venue type, although absolute counts reflect publication volume.
- Overall Distribution: 1.08% of AI/ML papers and 1.01% of Security papers contained invalid citations.NeurIPS had the highest absolute count at 391 papers, while NDSS had the highest proportion at 2.56%.
- Overall Distribution: Invalid citations were present across all eight analyzed venues, indicating that citation integrity issues span the studied research communities.AI/ML venues had more affected papers in absolute terms largely because they published more papers.
D. Temporal Trends and Propagation
Invalid citation rates remained relatively stable through 2024 before rising sharply in 2025. Repeated invalid citations also propagate across papers, suggesting copying from already-erroneous references, although the temporal association with AI-assisted workflows does not establish causality.
- Temporal Trends: The 2025 surge coincides with wider adoption of autonomous AI-agent workflows, but the data do not establish causality.The authors frame this temporal alignment as a hypothesis about changing citation-generation behavior.
- Temporal Trends: 1.61% in 2025: the invalid-citation rate rose 80.9% above the 2020–2024 average of 0.89%.Rates were relatively stable from 2020 to 2024, ranging from 0.76% to 0.98%.
- Propagation: The same erroneous AugMix title appeared in 16 papers across AAAI, IJCAI, and NeurIPS.Such repeated invalid citations are consistent with researchers copying references from papers that already contain errors.
- Propagation: 68 papers, or 11.3%, contained multiple invalid citations, consistent with batch-generation behavior enabled by LLM tools.This pattern appears alongside repeated invalid citations across papers.
VII. ANALYZING HUMAN FACTORS: SURVEY FINDINGS
The survey examines AI adoption, citation-reporting behavior, peer-review efficacy, and support for automated checks. It finds widespread AI use alongside trust-based verification practices and perceived weaknesses in peer review.
- AI adoption: 87.2% of respondents who answered the AI-use question report using AI-powered tools for research purposes.This adoption spans career stages and research areas.
- Reporting behavior: 41.5% of authors copy-paste BibTeX entries without checking, while 17.3% cite AI-suggested papers without reading them.Among reviewers, 76.7% do not thoroughly check references and 80.0% never suspect fake references in submissions.
- Reporting behavior: 41.3% of AI users encounter hallucinated citations often or very often, yet 44.4% choose no-action responses when references seem suspicious.The survey describes this as a trust-by-default norm despite exposure to hallucinated citations.
- Peer review efficacy: 74.5% of respondents view peer review as ineffective or not very effective at detecting citation errors.76.6% also rate hallucinated citations as a major problem or critical crisis.
- Automated checks and accountability: 70.2% strongly support automated checks to catch citation errors, while 91.5% assign citation-accuracy responsibility to authors alone.These findings combine strong support for automated checking with predominantly author-centered accountability.
VIII. DISCUSSION AND MITIGATION
The discussion characterizes ghost citations as a systemic threat that can propagate from LLM generation into the scientific record. It presents coordinated interventions across stakeholders and notes methodological limits that make detected hallucinations conservative estimates.
- Discussion: Ghost citations can enter and persist in the literature because verification practices are sparse and largely trust-based.The paper frames this process as a systemic threat to academic trust.
- Discussion: Invalid citations can propagate when researchers copy references from existing works, with repeated appearances making errors harder to detect.The paper reports the same erroneous reference appearing in up to 16 distinct papers.
- Mitigation: The paper recommends coordinated action across the publication ecosystem because no single stakeholder can solve the problem in isolation.Interventions target researchers, venues, tool developers, and multiple stages from generation to publication.
- Mitigation: Researchers are advised to treat AI-generated outputs as unverified, prefer retrieval-grounded tools, check titles in trusted indexes, and avoid unverified BibTeX copying.The recommendations also call for reading at least an abstract or direct source summary before citing a paper.
- Limitations: The verification pipeline may undercount hallucinations because it checks title similarity and can miss fabricated citations resembling real papers.Some flagged citations may also be legitimate but poorly indexed, although manual verification addresses that case.
APPENDIX
The appendix documents benchmark domains, prompting procedures, auxiliary verification prompts, and analyses of online settings and batch size. It also records that accessible API toggles do not necessarily represent vendors’ full retrieval or reasoning systems.
- Benchmark setup: The LLM citation-generation experiment covers 40 computer-science domains categorized by arXiv subject areas.The appendix provides the corresponding domain abbreviation mapping.
- Prompting: The benchmark prompt requests a specified number of realistic academic references in a research field and requires JSON output with reference metadata.Reference types include articles, series, theses, monographs, and unknown entries.
- Auxiliary verification: An auxiliary experiment asks models to judge whether a bibliographic entry’s title, authors, year, and venue correspond to a real publication.The required output is a JSON object containing a Boolean result and a reason.
- Additional analyses: The appendix compares citation hallucination rates under offline versus online-plus-thinking settings and examines batch sizes of 10, 20, and 30 citations per prompt.The batch-size results show no consistent correlation with hallucination rates.
- Limitations: The online-search and chain-of-thought comparison reflects third-party API-level toggles rather than each vendor’s full native retrieval or reasoning stack.This constrains how the setting comparison should be interpreted.
E. Domain-Level Hallucination Rates
This section reports domain-level hallucination-rate analysis and citation-stability statistics across repeated runs. Valid citations are more consistently reproduced than hallucinated citations, and frequently recalled valid papers appear strongly encoded in model memory.
- Domain-level analysis: Citation hallucination rates are aggregated across all 13 models and ranked over 40 computer-science research domains.The domain-level table presents rates from highest to lowest.
- Citation stability: Citation stability measures how often a model generates identical citations across repeated runs for the same model and configuration.A configuration combines research topic, requested reference count, chain-of-thought setting, and online-search setting.
- Citation stability: A stability score of 1 indicates perfect consistency, whereas 0 indicates maximal variability with no repetition across runs.The score is computed over citations aggregated from repeated runs.
- Results: Valid citations reach mean stability of up to 0.58 for DeepSeek and 0.57 for Qwen-3, compared with up to 0.23 for hallucinated citations.The paper interprets this pattern as stronger repeated recall for real papers than for fabricated references.
- Results: Frequently generated valid citations include seminal works such as NeRF, RAG, and U-Net, suggesting strong encoding in models’ parametric memory.The appendix presents these papers among the most consistently recalled references across domains.
G. Survey Response Details (Full)
This section provides the full survey-response breakdown, covering adoption, verification, responsibility, reviewer reactions, metadata sources, and related behaviors. It also documents the question map and a reverse-worded consistency check used to exclude three inconsistent responses.
- Full response details: Tables XIII–XVI report all survey response options across the study’s adoption, verification, responsibility, support, behavior, reviewer, metadata, and claim-verification questions.The tables contain 94 valid responses out of 97 total.
- Question mapping: Table XII maps the survey’s question wording and numbering for interpreting the detailed response tables.
- Consistency check: Three respondents were excluded after answering inconsistently on reverse-worded questions about risky copying and meticulous BibTeX verification.Q38 asks about copying BibTeX entries without checking them, whereas Q39 asks about verifying every field before submission.