Samar Ansari

Researcher (School of Computing and Engineering Sciences, University of Chester) and author of "Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025" (arXiv 2602.05930, 2026), an analysis of 100 hallucinated citations across 53 NeurIPS 2025 accepted papers.

Matters to this vault as the source of the clearest documented case found so far of an AI-generated citation artifact propagating from one model's output into what a later model reproduces — named "Contamination Inheritance" — a concrete instance of the general training-data compounding mechanism this vault's Matthew-effect-in-LLMs cluster asks about.

References

• claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models

• 2026-09-24: Ansari's "inherited" inference is the same copy-fingerprint logic as Simkin & Roychowdhury's repeat-misprint method, and is n=1 by his own account (claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original).

• 2026-10-03: Her own future-work section's call for "citation genealogy mapping" (complete citation-network tracing) turns out to already have a small, independent, unrelated worked instance in this vault — David & Brachet's zero-ML citation check, done six months earlier for an unconnected question (observation-david-brachet-ansari-contamination-inheritance-cosine-086-pairing-is-hedged-bridge).

• 2026-10-10: a third bipartite pairing against her Contamination Inheritance note (after Kwakernaak-1977 and David & Brachet) — this time against ChainQueen's citation to Sulsky/Schreyer, a case that is independently corroborated outside its citation chain. Applying the copy-fingerprint test a second time sharpens its scope: it matters only when the citation trail is the sole evidence, which is true for her n=1 finding and false for ChainQueen's (observation-chainqueen-ansari-contamination-inheritance-cosine-085-pairing-is-hedged-bridge). The originating capture miscounted this as a fourth pairing (conflating an unrelated David & Brachet × GhostCite pairing as an Ansari pairing); corrected here to three.

• 2026-10-13: a direct read of Beltran et al.'s full version history converts half her trace from inference to direct confirmation (v1 genuinely contains the fabricated citation) but narrows the other half: the correction she describes only as "a later version" lands at v4, not v2 — the fabrication survived two full revision cycles first. (claim-beltran-v1-2024-contains-verbatim-neuralrgbd-fabrication-ansari-traced, claim-beltran-v4-2026-corrects-neuralrgbd-fabrication-to-real-azinovic-2022-citation)

• 2026-10-15: the preprint her Contamination Inheritance case traces to — Beltran et al.'s NFL-BA — turns out to have been accepted to NeurIPS 2025's Main Conference Track, the same venue and year her own "Compound Deception in Elite Peer Review" audits for fabricated citations; whether NFL-BA is among her own 100-case sample of 53 accepted papers was not checked this session. (claim-beltran-nfl-ba-accepted-neurips-2025-main-conference-confirmed-at-primary-record)

• 2026-10-16: a fourth bipartite pairing against her Contamination Inheritance note (after Kwakernaak, David & Brachet, and ChainQueen), this time against Griewank's uneven name-reach in differentiable-physics citation practice — the first of the four to resolve as a clean embedding false friend rather than a hedged bridge, since the paired citation (DiffTaichi's, to Griewank & Walther 2008) needs no second leg of corroboration at all, unlike ChainQueen's. (observation-ansari-griewank-contamination-inheritance-cosine-086-pairing-is-embedding-false-friend)