---
id: "20261117-0206-ansari-arxiv260205930-and-naser"
title: "Ansari (arXiv:2602.05930) and Naser (arXiv:2603.03299) have each now reached the single-source concentration cap — what would discharge it, and does their mutual non-citation hold up?"
type: "capture"
status: "promoted"
origin: "batch"
promoted_to: ["30-notes/observation-phantom-references-2026-partially-discharges-ansari-single-source-cap.md","30-notes/claim-chereches-2026-cites-ansari-as-bare-reference-not-empirical-test.md","30-notes/claim-chen-liang-wang-2026-cites-naser-as-bare-reference-not-empirical-test.md"]
not_promoted: ["Claim 1 (neither preprint past v1 / no journal-ref as of 2026-11-17): pure arXiv metadata — routed per promote.md 5a to the question's progress log instead of a claim-note (see question-corroborate-ansari-naser-2026-single-source-cap.md's 2026-11-17 entries).","Claim 4 (mutual non-citation holds up on a fresh check; Phantom References bridges but doesn't reconcile the two): the metadata-recheck half routed to the question log alongside Claim 1; the substantive half (Phantom References treats the two on separate axes without resolving the contrast) folded into the Phantom References note's own body rather than duplicated as a separate note. Still blocked from its own claim-note — it rests on both preprints at once, and Naser's cap is not yet discharged even though Ansari's partially is.","Further leads (Ansari's other 21 Semantic Scholar citers, Naser's other 4, Zhao et al.'s ~2.5M-paper PubMed Central audit, GhostCite as a possible second Ansari-branch corroboration candidate): none read this session; left as leads for a future capture, not claims.","Entity candidates: Walters & Wilder (2023) re-flag — already has a watching stub (entity-walters-wilder-2023.md, first_seen 2026-11-16); this capture's claim that it is the founding study both Ansari's AND Naser's audits build on is not actually substantiated by any claim checked this session (only Naser's related-work section was confirmed crediting it), so no dated update was appended to avoid asserting an unverified fact. Russinovich, Kumar & Salem: co-authorship on one joint paper only, no separate vault standing — per the 2026-10-01 Cali ruling, a mention not a hub; named in the new observation note instead. Chereches; Chen, Liang & Wang: minor role, bare citation only — not promoted to hubs. RefChecker: named once in passing, not engaged — too thin even for a watching stub. GhostCite: already recurs in 3 existing notes and arguably clears the hub bar, but this capture didn't substantively engage it (a further lead only) — flagged to 00-meta/seek-flags.md ([entity], 2026-11-17) rather than built here, to keep this capture's scope to what it actually investigated."]
writer_model: "claude-sonnet-5"
date_created: "2026-11-17T00:00:00.000Z"
provenance: "batch run, 2026-11-17"
derived_from: []
verifies: "question-corroborate-ansari-naser-2026-single-source-cap"
tags: ["citation-fabrication","samar-ansari","mz-naser","single-source-cap","corroboration","hallucination-taxonomy","arxiv","phantom-references"]
seek_code_commit: "d40ee39"
---


[[question-corroborate-ansari-naser-2026-single-source-cap]] (raised 2026-11-16) asks two things: what would discharge the single-source concentration cap each preprint has now hit (three claim-notes apiece — [[entity-samar-ansari]], [[entity-mz-naser]]), and whether the two papers' mutual non-citation, already directly verified by an end-to-end reference-list read, still holds. This capture re-checks both arXiv records today and searches outward from each paper via Semantic Scholar's citation graph for independent engagement.

## Claim: As of 2026-11-17, neither preprint has itself progressed past a single (v1) version or acquired a journal-ref, so neither has an available "accepted to a refereed venue" discharge route yet
verifies: question-corroborate-ansari-naser-2026-single-source-cap

A direct re-fetch of both arXiv abstract pages today shows each record is still exactly one version, with no journal-ref field present (compare the precedent that discharged the cap for [[entity-beltran-et-al-2024-v1|Beltran et al.'s NFL-BA]] once its NeurIPS 2025 acceptance was confirmed at the primary record). Ansari's page lists only "[v1] Thu, 5 Feb 2026 17:43:35 UTC (64 KB)"; Naser's lists only "[v1] Sat, 7 Feb 2026 00:14:00 UTC (1,044 KB)." Neither page carries a "Journal-ref:" line of the kind arXiv adds once a paper is confirmed published.

- source_url: "https://arxiv.org/abs/2602.05930"
- source_title: "Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025"
- source_author: "Samar Ansari"
- source_date: "2026-02-05"
- source_venue: "arXiv:2602.05930 [cs.DL], abstract page"
- source_sha: "0ffa26a7f4c69bfbd81312a9a83e56ba06848f3708e251a40e9b1f81fdadfdd6"
- source_quote: "[v1]\n        Thu, 5 Feb 2026 17:43:35 UTC (64 KB)"
- source_tier: 1

- source_url: "https://arxiv.org/abs/2603.03299"
- source_title: "How LLMs Cite and Why It Matters: A Cross-Model Audit of Reference Fabrication in AI-Assisted Academic Writing and Methods to Detect Phantom Citations"
- source_author: "MZ Naser"
- source_date: "2026-02-07"
- source_venue: "arXiv:2603.03299 [cs.CL], abstract page"
- source_sha: "6bc8883314791d505818fc1523fcc17fe3e42afc98adb91c36e2055984c70993"
- source_quote: "[v1]\n        Sat, 7 Feb 2026 00:14:00 UTC (1,044 KB)"
- source_tier: 1

## Claim: A third, independent author group's own large-scale audit — Russinovich, Kumar & Salem's "Phantom References" (arXiv:2607.00738, Microsoft) — names Ansari's finding specifically and frames its own convergent result as corroborating evidence, making it the strongest candidate so far for discharging Ansari's cap, though the candidate source is itself an unrefereed preprint
verifies: question-corroborate-ansari-naser-2026-single-source-cap

Semantic Scholar's citation graph lists 22 citing records for Ansari's preprint (re-queried today). Among them is "Phantom References: Hallucinated Citations That Survive Peer Review at Top-Tier Conferences" (Mark Russinovich, Ram Shankar Siva Kumar, Ahmed Salem — all Microsoft, no author overlap with Ansari or Naser), which independently built its own verification pipeline (RefChecker) and audited camera-ready proceedings from ICLR, ICML, NeurIPS, and USENIX Security directly — a different corpus, different method, and different institution from Ansari's 100-citation NeurIPS sample. Its related-work section names Ansari by name as a within-community precedent: "Other studies confirm the trend within single communities, e.g., Ansari [14] on NeurIPS 2025 and Sakai et al. [15] on the ACL conference family (2024 and 2025)." Its own independent measurement reaches a result consistent with Ansari's headline finding that hallucinated citations survive elite peer review and reach NeurIPS specifically: "In 2025, roughly one in twenty NeurIPS and USENIX Security papers contains at least two likely hallucinated academic-paper-like references under our strict definition." The paper states explicitly that it treats this agreement as corroborative: "Like both, we report only in aggregate and read the convergence of independent measurements as evidence that the phenomenon is real and growing."

This meets the cap's stated independence test — "a different author group and a different venue" (sources.md) — on the same terms that discharged the [[observation-daryani-independently-corroborates-chronoculture-discharging-single-source-cap|ChronoCulture/Daryani pairing]]: different people, different institution, independent data and method, converging on the same real-world phenomenon. The caveat is the same one noted there, sharpened: Phantom References is itself a second unrefereed arXiv preprint (v2, 6 July 2026; no journal-ref), so even if it counts as independent corroboration of Ansari's *general* finding, the branch would rest on two unreviewed papers agreeing rather than on the field's verdict. It also does not reach Ansari's more specific claims — her five-category failure-mode taxonomy, the "Contamination Inheritance" mechanism, or the Zhu/NeuralRGB-D boundary case — which Phantom References' identity-level pass/fail definition does not attempt to classify.

The same paper also names Naser, but on a separate, non-corroborating axis: "Fabrication under controlled prompting. ... Naser [17] audits 69,557 model-generated citations across ten deployed models and reports hallucination rates of 11.4%–56.8% depending on model and domain." It cites this as a distinct phenomenon (prompted generation in isolation) from its own target (what survives into peer-reviewed proceedings), not as a test of Naser's specific invention-vs-distortion resemblance scoring — so it does not supply a discharge candidate for Naser's cap in the way it does for Ansari's.

- source_url: "https://arxiv.org/pdf/2607.00738"
- source_title: "Phantom References: Hallucinated Citations That Survive Peer Review at Top-Tier Conferences"
- source_author: "Mark Russinovich, Ram Shankar Siva Kumar, Ahmed Salem"
- source_date: "2026-07-01"
- source_venue: "arXiv preprint 2607.00738v2 [cs.DL]; Microsoft / Microsoft Azure"
- source_sha: "e3457c4260c92e062ea423f548805ce0da334f34b088d29b7eba1731499af5e2"
- source_quote: "Other studies confirm the trend within single communities, e.g., Ansari [14] on NeurIPS 2025 and Sakai et al. [15] on the ACL conference family (2024 and 2025)."
- source_quote_2: "In 2025, roughly one in twenty NeurIPS and USENIX Security papers contains at least two likely hallucinated academic-paper-like references under our strict definition."
- source_quote_3: "Like both, we report only in aggregate and read the convergence of independent measurements as evidence that the phenomenon is real and growing."
- source_quote_4: "Naser [17] audits 69,557 model-generated citations across ten deployed models and reports hallucination rates of 11.4%–56.8% depending on model and domain."
- source_tier: 1

## Claim: Each preprint has also drawn at least one citation from an already-refereed, published venue, but in the one case fully checked the citation is a bare related-work mention rather than an empirical test — so publication-adjacent citation alone does not yet discharge either cap
verifies: question-corroborate-ansari-naser-2026-single-source-cap

Among Ansari's 22 Semantic-Scholar-listed citers is Razvan Chereches's "Fabricated citations threaten the evidence base that clinical AI learns from" — a Correspondence published in *AI and Ethics* (Springer, vol. 6, article 650, 9 November 2026), a refereed journal, different author, different institution. Its reference list cites Ansari 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 piece's own prevalence evidence is explicitly keyed to a different, larger audit of ~2.5 million PubMed Central papers, not to Ansari's NeurIPS sample, and it states its own limits plainly: "the prevalence evidence rests mainly on one audit." This is a bare citation, not an empirical replication — the same shape the vault has already documented for [[claim-cpl-2026-cites-neunenu-as-related-work-not-empirical-test|CPL-2026's citation of Neunenu]].

The parallel case on Naser's side is Chen, Liang & Wang's "Beyond existence checks: a cross-database benchmark of bibliographic metadata consistency across academic domains and LLM-generated citations" (*Scientometrics*, vol. 131, pp. 5965–5984, published 22 July 2026 — a refereed journal, authors at National Taiwan University of Science and Technology and Providence University, no overlap with Naser). Its reference list cites Naser's preprint by title: "Naser, M. Z. (2026). How LLMs cite and why it matters: A cross-model audit of reference fabrication in AI-assisted academic writing. https://arxiv.org/abs/2603.03299." The paper's own empirical target — cross-database metadata *disagreement* (66.7% of 1,246 citations show at least one inter-database field conflict) — is a different measurement from Naser's hallucination-resemblance taxonomy; nothing in the visible abstract, data-availability note, or reference list indicates it tests or replicates Naser's 11.4%–56.8% hallucination-rate finding. Both citations are therefore evidence that each preprint is reaching refereed literature as a *citation*, not evidence that either has been independently *corroborated* in the cap's sense.

- 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_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"
- source_tier: 1

- source_url: "https://doi.org/10.1007/s11192-026-05727-6"
- source_title: "Beyond existence checks: a cross-database benchmark of bibliographic metadata consistency across academic domains and LLM-generated citations"
- source_author: "Chung-Kuang Chen, Shu-Hao Liang, Teng-Shih Wang"
- source_date: "2026-07-22"
- source_venue: "Scientometrics, vol. 131, pp. 5965–5984 (2026), Springer"
- source_sha: "fc3429652277732f1cda94fe4d8a47c1c753f00a440eae3bf3bd2d2eb5ba710f"
- source_quote: "Naser, M. Z. (2026). How LLMs cite and why it matters: A cross-model audit of reference fabrication in AI-assisted academic writing. https://arxiv.org/abs/2603.03299"
- source_tier: 1

## Claim: The mutual non-citation holds up on a fresh check today, and outside literature now bridges the two papers without reconciling them — a third source cites both as separate, non-overlapping strands rather than as counterparts
verifies: question-corroborate-ansari-naser-2026-single-source-cap

Today's re-fetch of both arXiv abstract pages (above) confirms neither preprint has posted a revision since the 2026-11-16 capture that checked both reference lists end to end — so no opportunity for either to add a citation to the other has arisen. Separately, "Phantom References" (Claim 2 above) is the first source found in this search that discusses Ansari and Naser in the same document: it places them in different related-work subsections — Ansari under "Auditing hallucination" (what survives into published proceedings) and Naser under "Fabrication under controlled prompting" (what models produce when prompted in isolation) — treating them as answering different questions rather than as each other's counterpart on the invention-vs-distortion question the vault's own [[observation-ansari-naser-2026-cosine-080-pairing-is-analytical-contrast|analytical-contrast observation]] already identified. No source found in this search discusses the two papers' opposite headline conclusions as being in tension with one another; the contrast remains, as that observation states, something only a direct read of both primaries surfaces.

(Sourcing for this claim rests on the same primaries already cited above: the two abstract-page re-fetches and the Phantom References PDF.)

> [!note] Seek's commentary:
> The cleanest discharge candidate in this search is Phantom References, and it is clean mostly by accident: three Microsoft researchers built an unrelated verification pipeline for an unrelated purpose (tooling for publication-integrity policy, not a response to Ansari) and happened to measure the same thing she did, at a different venue set, with a different method, and landed in the same place. That is closer to the independence the cap actually wants than a citation chain ever is — but it buys discharge only for Ansari's broadest claim, not for the sharper ones (Contamination Inheritance, the five-category split) that make her paper worth three notes in the first place. Meanwhile the two "refereed venue" citations I found are the thing sources.md already has a name for: a citation is not a test. I'd rather report two real but partial candidates than round either one up to "discharged."
> — Seek

## Further leads

- Ansari's Semantic Scholar citer list (22 total) includes several not checked in depth here and worth a closer look: "Distributed Denial of Science: How Indirect Data Poisoning of AI Systems Can Industrialize Scientific Fraud" (Gyevnar, Kasirzadeh & Shah, arXiv:2607.10712); "AI and The Enclosure of the Scientific Commons" (P. Williams, arXiv:2610.27251); "Automatic Reference Scanner and Explainer" (Staudinger, Kusa, Peikos & Hanbury, ACM CIKM 2026 proceedings, DOI 10.1145/3799682.3840278); and "Inspectable AI for Science" (Binkytė et al., IEEE SPW 2026) — none confirmed read this session.
- Naser's Semantic Scholar citer list (5 total) includes "Principles and Techniques for Agentic Research in Experimental Physics" (arXiv:2610.17193) and "PaperDoctor" (arXiv:2609.16995) — likely background citations in survey-style papers, not checked.
- Zhao et al.'s "LLM hallucinations in the wild: large-scale evidence from non-existent citations" (arXiv:2605.07723, ~2.5M PubMed Central papers) recurs across both Phantom References' and Chereches's reference lists as a third, much larger independent audit of the same general phenomenon in a different (biomedical) corpus — a candidate worth its own capture, not pursued here.
- "GhostCite" (Xu et al., arXiv:2602.06718, 2.2M citations across 56,381 AI/ML and security-venue papers) is named by Phantom References as a second "concurrent work" alongside Ansari — another possible independent-measurement candidate for the Ansari branch, not checked directly this session.

## Entity candidates

- Walters & Wilder (2023) — paper/concept — already has an entity stub ([[entity-walters-wilder-2023]]); re-flagged here because it is the founding empirical study both Ansari's and Naser's 2026 audits build on, and the "earlier foundational figure" this whole comparison implicitly measures itself against — raised first per the standing blind-spot instruction, not last.
- Mark Russinovich, Ram Shankar Siva Kumar, Ahmed Salem — persons — authors of "Phantom References" (arXiv:2607.00738), the independent Microsoft-affiliated study whose own convergent measurement is this capture's strongest discharge candidate for Ansari's cap; no entity page yet.
- Razvan Chereches — person — author of the AI and Ethics Correspondence citing Ansari; no entity page yet, minor role (bare citation, not tested further).
- Chung-Kuang Chen, Shu-Hao Liang, Teng-Shih Wang — persons — authors of the Scientometrics paper citing Naser; no entity page yet, minor role (bare citation, not tested further).
- RefChecker — concept/tool — Phantom References' open-sourced citation-verification pipeline (resolves against arXiv, Semantic Scholar, OpenAlex, CrossRef, DBLP, ACL Anthology, escalating to web search); may warrant its own concept note alongside the vault's other citation-verification tools.
- GhostCite — concept/tool — named by Phantom References as a second concurrent large-scale audit (2.2M citations, 56,381 papers); appears in the vault's existing source lists already but not yet checked as a corroboration candidate for the Ansari branch specifically.
