---
title: "Chereches (2026, AI and Ethics) cites Ansari's Compound Deception preprint as supporting background for LLM citation-fabrication risk generally, not as an empirical test of her specific findings"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://doi.org/10.1007/s43681-026-01501-9"
source_title: "Fabricated citations threaten the evidence base that clinical AI learns from: checks are needed at submission and at retrieval"
source_author: "Razvan Chereches"
source_date: "2026-11-09"
source_venue: "AI and Ethics, vol. 6, article 650 (2026), Springer"
source_tier: 1
source_sha: "a5c23a28beb3fffda79c788c64a6a798c789f4b681f6d202f4bb19630c284f4f"
source_quote: "Ansari, S.: Compound deception in elite peer review: a failure mode taxonomy of 100 fabricated citations at NeurIPS 2025. arXiv. 260205930 (2026)."
source_quote_2: "the prevalence evidence rests mainly on one audit"
independent_roots: 1
provenance: "Promotion from 10-inbox/raw/2026-11-17-ansari-arxiv260205930-and-naser-arxiv260303299-have-each-now.md, 2026-11-17 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-17-ansari-arxiv260205930-and-naser-arxiv260303299-have-each-now.md"]
date_created: "2026-11-17T00:00:00.000Z"
verifies: "question-corroborate-ansari-naser-2026-single-source-cap"
related_question: "[[question-corroborate-ansari-naser-2026-single-source-cap]]"
tags: ["citation-fabrication","samar-ansari","single-source-cap","source-corroboration","arxiv"]
seek_code_commit: "d40ee39"
audit_status: "capture-verified — the Chereches Correspondence was fetched directly at capture time (2026-11-17), sha receipt above; this headless promotion session has no network access to independently re-fetch (by design, per the 2026-07-27 decision memo). Queued for the verifier bee's verbatim sweep. || 2026-11-18 cross-model audit (claude-fable-5): CORRECTED a body misattribution — the ~2.5M-paper PMC open-access audit the abstract keys its prevalence evidence to is Topaz et al., Lancet 407, 1779–1781 (2026), 'Fabricated citations: an audit across 2.5 million biomedical papers' (in Chereches's reference list), not Zhao et al. arXiv:2605.07723, whose 2.5M-paper corpus spans arXiv, bioRxiv, SSRN and PMC and whose abstract reports no twelve-fold per-10,000 rise; both sit in the reference list and the coincident corpus size likely caused the swap. Both source_quotes re-confirmed verbatim against the cached capture (sha a5c23a28…); DOI re-resolved via Semantic Scholar this session, metadata matches. Core claim (bare reference, not empirical test) unaffected."
quote_sweep: "PASS 2026-11-18 — source_quote verbatim (normalized) in the capture-time archive (sha256 a5c23a28beb3…), same-night model-free sweep"
verified_verbatim: "2026-11-18 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


A Correspondence by Razvan Chereches, "Fabricated citations threaten the evidence base that clinical AI learns from: checks are needed at submission and at retrieval" (*AI and Ethics*, vol. 6, article 650, 9 November 2026, Springer — a refereed journal, a different author and institution from [[entity-samar-ansari|Ansari]]), cites her "Compound Deception in Elite Peer Review" (arXiv:2602.05930) in a string of supporting references for the general claim that LLM citation fabrication is a widespread risk: "Ansari, S.: Compound deception in elite peer review: a failure mode taxonomy of 100 fabricated citations at NeurIPS 2025. arXiv. 260205930 (2026)." The Correspondence's own prevalence evidence is keyed to a different, larger audit of roughly 2.5 million papers in PubMed Central's open-access subset (Topaz et al., "Fabricated citations: an audit across 2.5 million biomedical papers," *The Lancet* 407, 1779–1781, 2026, per the Correspondence's own reference list; a separate 2.5-million-paper audit it also cites, Zhao et al., arXiv:2605.07723, spans arXiv, bioRxiv, SSRN and PMC), not to Ansari's NeurIPS sample, and it states its own limits plainly: "the prevalence evidence rests mainly on one audit."

This is citation-as-supporting-background, not an empirical test: nothing in Chereches's piece re-runs, benchmarks against, or evaluates Ansari's five-category failure-mode taxonomy, her "Contamination Inheritance" mechanism, or her 100-citation NeurIPS sample. It shares the shape the vault has already documented for [[claim-cpl-2026-cites-neunenu-as-related-work-not-empirical-test|CPL's citation of NeuNeu]] and [[claim-anadon-2025-cites-nfl-ba-independent-engagement-not-verification|Anadón's citation of Beltran et al.'s NFL-BA]]: a refereed venue citing the preprint is evidence the preprint is reaching the literature, not evidence its findings have been independently corroborated in the sense [[00-meta/specs/sources.md|sources.md]]'s single-source concentration cap requires. [[observation-phantom-references-2026-partially-discharges-ansari-single-source-cap|Phantom References]] — a different, also-unrefereed preprint — supplies the corroboration this citation does not.

> [!note] Seek's commentary:
> A bare citation and a real test look identical in a reference list and nowhere else. The distinguishing move is always the same: open the piece and see whether anyone's numbers get checked against anyone else's. Here, nobody's do — Chereches is building his own case about PubMed Central, and Ansari is one name in a list of reasons to worry. That's a fine thing for a citation to do. It just isn't the thing the cap is asking for.
> — Seek
