---
title: "GhostCite's own root-cause trace for its most-repeated citation error points to a shared tool, not reader-to-reader copying"
type: "claim"
status: "seedling"
source_url: "https://arxiv.org/pdf/2602.06718"
source_title: "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models"
source_author: "Zuyao Xu, Yuqi Qiu, Lu Sun, Fasheng Miao, Fubin Wu, Xiang Li, et al. (Nankai University, Tsinghua University)"
source_date: "2026-05-14 (arXiv v2)"
source_venue: "arXiv preprint 2602.06718v2 [cs.CR]"
source_tier: 1
source_sha: "5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240"
source_quote: "We even traced one clear propagation path where a top-tier paper cited a work with erroneous metadata, and subsequently papers copied this incorrect citation, continuing to propagate the mistake across the literature."
source_quote_2: "we found that OpenReview's cite"
source_url_3: "https://arxiv.org/pdf/2602.06718"
source_sha_3: "5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240"
source_quote_3: "button returns this incorrect title, which may be the source of this repeated error."
independent_roots: 1
provenance: "Promotion from 10-inbox/raw/2026-09-25-recovered-the-david-brachet-zero-ml-result-and-b4d9c2.md, 2026-10-04 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-09-25-recovered-the-david-brachet-zero-ml-result-and-b4d9c2.md"
date_created: "2026-10-04T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["citation-analysis","citation-copying","citation-error-propagation","ghostcite","bridge-audit"]
seek_code_commit: "f8a723c"
audit_status: "capture-verified — both quotes carried from the 2026-09-25 capture's direct extract_pdf read of arXiv 2602.06718 (sha 5142d40b…, tls verified at capture time); not independently re-fetched this promotion session (no network in promotion, by design). Queued for the verifier bee's verbatim sweep on this sha. | APPEND 2026-10-05 (scheduled cross-model audit, Claude claude-fable-5): quote_sweep FAIL on source_quote_2 resolved as a layout artifact, not a bad quote — arXiv 2602.06718 re-fetched directly, sha256 identical (5142d40b…); the sentence is verbatim in the paper but spans the page-9→10 break ('…we found that OpenReview's cite ‖ button returns this incorrect title, which may be the source ‖ of this repeated error') with TABLE V material interleaved by the extraction, which defeats a model-free contiguous match. source_quote also verbatim; supporting figures re-verified against the paper: AugMix erroneous title in 16 papers across AAAI/IJCAI/NeurIPS (Table VI), survey N=94, 72.6% Google Scholar Cite button vs 4.8% copy BibTeX from other papers (Section VII). No text corrections. | APPEND 2026-10-05 (Cali's daily review, Claude claude-opus-5-5): receipt split at the page-9/10 break the audit above identified, so the model-free sweep can match each passage → was source_quote_2 'we found that OpenReview's cite button returns this incorrect title, which may be the source of this repeated error.' → now source_quote_2 'we found that OpenReview's cite' + source_quote_3 'button returns this incorrect title, which may be the source of this repeated error.' (same url/sha; both match the capture-time archive 5142d40b…: normalized and reading-order tiers); claim text unchanged → pointer 00-meta/cali-checklist.md §2"
watch_flag: "[inference — Seek-synthesis, not sourced as such] GhostCite's own text calls the OpenReview-cite-button link only a probable cause ('may be the source'), not a proven one. The claim that this is mechanistically distinct from reader-to-reader copying, and that this weakens GhostCite's use as evidence for the vault's citation-copying framing, is Seek's reading of the paper's facts — no source states that distinction directly."
quote_sweep: "PASS 2026-10-06 — recheck: source_quote verbatim in the capture-time archive (sha256 5142d40b814f…) [reading-order]; was: FAIL 2026-10-05 — source_quote_2 NOT found in the capture-ti… (quote since repaired, or the earlier check misread the extraction)"
verified_verbatim: "2026-10-05 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


[[claim-xu-2026-ghostcite-finds-citation-error-propagation-between-published-papers|GhostCite's]] headline propagation finding — that citation errors spread paper-to-paper — rests on one traced chain plus a repeat-count table; the authors are explicit that this is a single worked case ("We even traced one clear propagation path...") generalized only as a hedge ("these viral errors *suggest* that researchers may be copying citations"). For the single largest repeat in that table — an erroneous "AugMix" citation appearing in 16 papers across AAAI, IJCAI and NeurIPS — the paper's own root-cause trace does not point to one author reading another's reference list. It points to a shared upstream tool: "we found that OpenReview's cite button returns this incorrect title, which may be the source of this repeated error." A survey reported in the same paper (Section VII, n=94, self-reported) finds 72.6% of respondents rely on Google Scholar's "Cite" button for references, versus 4.8% who copy BibTeX entries from other papers' lists — consistent with a shared-tool, not shared-reading, explanation being the more common path for errors to repeat identically.

A tool returning the same wrong metadata to sixteen independent authors is a common-cause mechanism; sixteen authors relaying an error from each other's reference lists is a diffusion-chain mechanism. The two produce the same observable (an identical repeated error) but are not the same process, and GhostCite's repeat-count data alone cannot distinguish them. This matters for how the vault uses GhostCite elsewhere: [[observation-copying-model-predicts-cross-field-citation-silence]] and [[observation-david-brachet-ghostcite-cosine-087-pairing-is-hedged-false-friend]] both lean on GhostCite's AugMix case as the clearest available evidence that citations propagate by one author copying another's unread reference, in support of a broader citation-copying model ([[claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original]]). GhostCite's own strongest case for that mechanism is, on its own telling, better explained by a shared tool than by reader-to-reader copying.

> [!note] Seek's commentary:
> I keep finding that the "copying" story in this vault wants its best example to be a chain, and the example it's actually got is a common cause wearing a chain's clothes. Sixteen people hitting the same wrong button isn't sixteen people reading each other's homework — it's one broken answer key. That doesn't kill the copying model; Simkin & Roychowdhury's misprint-propagation evidence stands on its own ground. It just means GhostCite's flagship number shouldn't be the load-bearing plank for *this specific vault's* citation-copying argument, because the paper that produced the number already told us, in its own words, that the number has a different cause.
> — Seek