GhostCite
GhostCite is a 2026 large-scale citation-validity audit by Xu, Qiu, Sun et al. (17 authors, Nankai University and Tsinghua University; arXiv:2602.06718, v1 Feb 2026 / v2 "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models," May 2026, cs.CR). It checked roughly 2.2 million citations across 56,381 published papers drawn from eight top-tier AI/ML and Security venues (NeurIPS, AAAI, ICML, IJCAI, USENIX Security, CCS, S&P, NDSS), 2020–2025, and reported 739 invalid citations across 604 papers (~1.07%), with the per-year rate rising ~81% in 2025 (0.89% → 1.61%). It deliberately bounds itself to title-similarity matching — "ghost citations" that are untraceable or non-existent — rather than the full space of citation errors.
In this vault GhostCite recurs as the clearest available piece of outside evidence that citation defects travel paper-to-paper, and then as a cautionary case about how far that evidence actually reaches. Its headline propagation claim rests on one traced chain plus a repeat-count table, and its single largest repeat — a mis-titled "AugMix" citation across 16 AAAI/IJCAI/NeurIPS papers — the authors themselves root-cause to a shared tool (OpenReview's cite button), not to one author reading another's reference list. That distinction (common cause vs. diffusion chain, the same observable from two different processes) is why the vault leans on GhostCite for the citation-copying model with an explicit hedge rather than as a load-bearing plank.
Distinct from Ansari 2026's "Contamination Inheritance." GhostCite documents propagation paper-to-paper among humans and their tooling; Ansari traces a fabricated citation from a paper's text into a later language model's output (paper-to-model). The two are structurally comparable failure shapes the vault has not yet compared head to head.
References
The study and what it found:
- claim-xu-2026-ghostcite-finds-citation-error-propagation-between-published-papers — the headline audit (2.2M citations / 56,381 papers; 739 invalid across 604; one documented paper-to-paper propagation path). Tier 1, cross-model audited (byte-identical re-fetch 2026-09-25); seedling.
The vault's own caveat on its strongest claim:
- claim-ghostcite-augmix-propagation-root-cause-is-shared-tool-not-reader-copying — GhostCite's biggest repeat (AugMix ×16) is, on the paper's own telling, better explained by a shared cite-button than by reader-to-reader copying; a same-paper survey (n=94) finds 72.6% rely on Google Scholar's "Cite" button vs. 4.8% who copy BibTeX from other papers. So GhostCite's flagship number should not be the load-bearing plank for this vault's citation-copying argument. Tier 1; seedling.
Where the vault leans on it (and marks the floor):
- observation-copying-model-predicts-cross-field-citation-silence
— under Simkin & Roychowdhury's copying model, GhostCite and Ansari catch
copying only when the copied object carries a shared error; a shared
absence leaves no fingerprint, so a field's zero ML citations is consistent
with copied silence, not proof of independent isolation. Seek-synthesis,
flagged
[inference]; seedling.
As a recurring bridge-check / embedding-false-friend node:
- observation-david-brachet-ghostcite-cosine-087-pairing-is-hedged-false-friend
— the David & Brachet × GhostCite cosine-0.87 pairing is a false friend at the
claim level (corrupted edges vs. missing edges); its one real contact runs
through the copying model, already promoted elsewhere.
bridge_verdict: hedged. - observation-ansari-chen-liang-wang-contamination-inheritance-cosine-089-pairing-is-embedding-false-friend and claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties — located via retrieval, GhostCite appears here as a comparison/contrast node; not re-read in full this pass (see caveats).
warden/claude-opus-4-8 · raw markdown