{"paper":{"title":"GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Large language models fabricate citations at rates from 14 to 95 percent, and the fraction of papers containing such errors rose 81 percent in 2025.","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Fasheng Miao, Feng Zhang, Fubin Wu, Haozhe Lu, Jiaji Liu, Jialu Li, Luo Jin, Lu Sun, Rui Luo, Xiang Li, Xinran Liu, Xinyi Wang, Yingxian Li, Yuqi Qiu, Yuxin Hu, Zhengze Zhang, Zuyao Xu","submitted_at":"2026-02-06T14:08:34Z","abstract_excerpt":"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"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"ghost citations represent a systemic threat to academic integrity, with all 13 tested LLMs hallucinating citations at rates from 14.23% to 94.93%, 1.07% of 56,381 papers containing invalid citations, and an 80.9% increase in 2025.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The GhostCite framework accurately identifies invalid citations without substantial false positives or negatives, and the sampled AI/ML and Security venues from 2020-2025 represent broader trends in citation validity.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLMs hallucinate citations at rates from 14.23% to 94.93%, with 1.07% of papers containing invalid citations and an 80.9% increase in 2025.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Large language models fabricate citations at rates from 14 to 95 percent, and the fraction of papers containing such errors rose 81 percent in 2025.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9e053f0c124c464b49c9a65df16b7f83db2e1faa6222fb14d7ce6d684cd27228"},"source":{"id":"2602.06718","kind":"arxiv","version":2},"verdict":{"id":"dcd5df92-19a9-47fc-b762-2d24e55354ca","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T06:57:42.158691Z","strongest_claim":"ghost citations represent a systemic threat to academic integrity, with all 13 tested LLMs hallucinating citations at rates from 14.23% to 94.93%, 1.07% of 56,381 papers containing invalid citations, and an 80.9% increase in 2025.","one_line_summary":"LLMs hallucinate citations at rates from 14.23% to 94.93%, with 1.07% of papers containing invalid citations and an 80.9% increase in 2025.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The GhostCite framework accurately identifies invalid citations without substantial false positives or negatives, and the sampled AI/ML and Security venues from 2020-2025 represent broader trends in citation validity.","pith_extraction_headline":"Large language models fabricate citations at rates from 14 to 95 percent, and the fraction of papers containing such errors rose 81 percent in 2025."},"references":{"count":52,"sample":[{"doi":"","year":null,"title":"URL: https://www.consul tmu.co.uk/ghost-references-cause-many-gen ai-errors/","work_id":"e6c7b5eb-13de-4411-bbab-de9b9d5315c9","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"URL:https://openrouter.ai","work_id":"06d4cbcc-b119-4769-8a39-c6eed503f972","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Scrape.do: Which API Is Better? URL: https://scrape.do/compare/scrapingdog-vs-s crapedo/","work_id":"066826c9-f2b4-4050-a30c-f681ab802c5b","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"URL: https://fo rums.zotero.org/discussion/75155/ghost-cit ations","work_id":"99d7d9b7-83b0-4158-ab66-b46096b37785","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"arxiv e-print archive: Computer science subject classes. https://arxiv.org/archive/cs , 2025. Accessed: 2025-09-24","work_id":"ab9a7f2c-4852-4b9c-9233-6222ad04f07c","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":52,"snapshot_sha256":"2570bb09c177f97827f026c6538306c0a6c2c52eef3a6351c5961406e30b7f8f","internal_anchors":3},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}