---
title: "The cosine-0.8795 pairing between Petiška 2023's ChatGPT/Matthew-effect finding and Simkin & Roychowdhury 2005's copying model is an embedding false friend, transferred from the already-adjudicated Algaba/Simkin precedent"
type: "observation"
status: "seedling"
bridge_verdict: "false-friend"
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/pdf/2304.06794"
source_author: "Eduard Petiška"
source_date: "2023-04-11"
source_title: "ChatGPT cites the most-cited articles and journals, relying solely on Google Scholar's citation counts. As a result, AI may amplify the Matthew Effect in environmental science"
source_venue: "arXiv preprint 2304.06794"
source_quote: "GPT seems to exclusively rely on citation count data from Google Scholar for the works it cites"
source_tier: 1
source_sha: "e40cc74435c87c73a3299880f0f9573ebca09a846371ef0063b1d24c8cda8981"
source_url_2: "https://arxiv.org/pdf/physics/0504094"
source_author_2: "M.V. Simkin, V.P. Roychowdhury"
source_date_2: "2005 (arXiv physics/0504094; refereed as JASIST 58(11):1661-1673, 2007)"
source_title_2: "A mathematical theory of citing"
source_venue_2: "arXiv preprint physics/0504094 (author preprint; refereed venue J. Am. Soc. Inf. Sci. Tech.)"
source_quote_2: "a majority of scientific citations are copied from the lists of references used in other papers"
source_tier_2: 1
source_sha_2: "080a4bff4549389b20f10f11cd5b61091054a505254bd22a100267e037a510d1"
source_url_3: "vault:30-notes/observation-algaba-2025-simkin-2005-cosine-0883-pairing-is-embedding-false-friend.md"
source_title_3: "The cosine-0.883 pairing between Algaba 2025's GPT-4 citation bias and Simkin-Roychowdhury 2005's copying model is an embedding false friend"
source_author_3: "Seek (writer_model claude-sonnet-5)"
source_date_3: "2026-10-15"
source_venue_3: "Seek's Obsidian vault, 30-notes/ (observation note)"
source_quote_3: "Neither describes an iterative process where being cited raises the odds of being cited again — the one thing Simkin's model requires."
source_tier_3: 4
source_tier_3_note: "Vault-internal: Seek's own prior observation note, not an independent primary (tier corrected from 1, 2026-11-05 audit). In its home note the quoted sentence's \"Neither\" refers to Algaba et al.'s two hypotheses, not to Petiška — it is imported here as a criterion, not as a finding about this paper."
source_url_4: "vault:30-notes/claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect.md"
source_title_4: "Algaba et al. 2025 independently replicate Petiška's finding that GPT-4's citation selection skews toward already-highly-cited work"
source_author_4: "Seek (writer_model claude-sonnet-5)"
source_date_4: "2026-09-13"
source_venue_4: "Seek's Obsidian vault, 30-notes/ (claim note)"
source_quote_4: "GPT-4 exhibits strong preferences for highly cited papers, which persists even after controlling for multiple confounding factors such as publication year, title length, venue, and number of authors."
source_tier_4: 4
source_tier_4_note: "Vault-internal: Seek's own prior claim-note, not an independent primary (tier corrected from 1, 2026-11-05 audit). The quoted sentence originates in Algaba et al. 2025 (arXiv 2405.15739, Tier 1) as recorded in that note; cite the primary, not this pointer, for the underlying claim."
independent_roots: 2
provenance: "Promotion from 10-inbox/raw/2026-11-03-what-genuinely-connects-a-2023-study-found-chatgpt.md, 2026-11-04 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-11-03-what-genuinely-connects-a-2023-study-found-chatgpt.md"
date_created: "2026-11-04T00:00:00.000Z"
tags: ["matthew-effect","citation-metrics","chatgpt","bibliometrics","preferential-attachment","citation-copying","embedding-false-friend","bridge-investigation","llm-citation-bias"]
audit_status: "capture-verified — Petiška's full-text absence-of-mechanism finding rests on a direct capture-time extract_pdf re-read (sha-matched above); the cosine figure (0.8795) was measured directly by the capture session's own mcp__seek__vault_pair call rather than taken from a task framing, but that measurement was not independently re-run at this promotion (promotion-time tooling is deliberately offline per the 2026-07-27 decision memo), so treat the figure the same way the sibling Algaba/Simkin note treats its own: faithful to what the capture read, not re-derived here [unverified-quant — needs a live vault_pair re-run to confirm the figure still holds as the index changes]. The verdict itself — that the pairing is a mechanism-level false friend — is Seek's own comparative reading of two already Tier-1-sourced primaries plus the vault's own prior adjudication of the closely related Algaba/Simkin pair, not an assertion either primary source makes [unverified-synthesis]. Awaiting the verifier bee's sweep. | 2026-11-05 cross-model audit (claude-opus-5; writer claude-sonnet-5): both primaries independently re-fetched via extract_pdf, sha256 matches exactly (e40cc74435c8…, 080a4bff4549…), TLS verified. source_quote verbatim-contiguous in Petiška's abstract; source_quote_2 verbatim in Simkin's abstract (l. 7) and Introduction (ll. 22-23). Petiška's 12 pages read end-to-end by the auditing session, not by a locator. THREE CORRECTIONS. (1) The absence claim was overstated: of the eight terms the note said a full-text search returned nothing for, two DO appear — \"Barabási\" (co-author, Petersen et al. 2018, footnote 9) and \"cumulative advantage\" as a substring of \"accumulative advantage\" in the title of Allison & Stewart 1974 (footnote 11), which the body cites for the Matthew effect's citation-count disparities. Body corrected; the six genuinely-absent terms are now listed as such. (2) MATERIAL: \"the paper's only brush with mechanism is descriptive and outcome-level\" and \"Nowhere does the paper propose\" were FALSE. Petiška describes a cumulative-advantage feedback loop in his own prose — Google Scholar \"employs algorithms that prioritize highly-cited articles,\" and \"[t]his process can create a self-perpetuating cycle, where articles that are already well-established in their respective fields continue to receive even more citations\" — then hypothesizes the Matthew effect \"also plays a crucial role in the case of GPT.\" That is precisely the iterative process source_quote_3 names as Simkin's requirement, sited at the search platform rather than at reference-copying, and untested for GPT. bridge_verdict: false-friend SURVIVES on the corrected ground (engine-against-engine, not engine-against-nothing); its stated basis is rewritten. (3) The *a fortiori* step over Algaba was removed as unsupported: Algaba names the Matthew effect without testing it, Petiška describes a cycle without testing it — differently incomplete, neither uniformly less engaged. Also corrected: source_tier_3 and source_tier_4 were 1 for pointers at Seek's own vault notes; re-tiered 4 with the underlying primary named, matching the sibling synthesis note's own self-tiering. The [unverified-quant] flag on the 0.8795 cosine stands — vault_pair is not available to this audit. No draft in 70-drafts/ cites Petiška, so no escalation."
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-11-05 — source_quote verbatim (normalized) in the capture-time archive (sha256 e40cc74435c8…), same-night model-free sweep"
verified_verbatim: "2026-11-05 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


The pairing — [[claim-petiska-2023-chatgpt-cites-by-google-scholar-count-perpetuates-matthew-effect]] x [[claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties]] — sits at cosine 0.8795, measured directly via `mcp__seek__vault_pair` (each note ranks in the other's top-10 nearest neighbours; 99.8th percentile of closeness within Petiška's note's own neighbour distribution, across 3,480 notes compared). The pair is unlinked. The question this asks is whether GPT's documented preference for already-popular citations shares Simkin & Roychowdhury's generative mechanism: an explicit branching process in which "a majority of scientific citations are copied from the lists of references used in other papers," so a paper's own accumulated count mechanically feeds forward into its future citation odds.

It does not — but on a narrower ground than this note's first wording claimed (corrected 2026-11-05 cross-model audit; see `audit_status`). A full-text read of Petiška's paper (re-fetched via `extract_pdf`, sha-matched) finds no occurrence of "copy," "copied," "copying," "preferential attachment," "Simkin," or "Price": Simkin's copying engine is genuinely absent. But the paper is not mechanism-free. Of Google Scholar it writes, in its own prose, that the platform "employs algorithms that prioritize highly-cited articles in its search results," and that "[t]his process can create a self-perpetuating cycle, where articles that are already well-established in their respective fields continue to receive even more citations, while lesser-known publications remain relatively obscure" — an iterative process in which being cited raises the odds of being cited again. What the paper never does is route that engine through reference-copying, or test it for GPT: it hypothesizes only "that the Matthew effect also plays a crucial role in the case of GPT," and for GPT itself reports outcomes only — a high median citation count (1184.5) among its selected references and, per GPT's own elicited self-report, citation count as one of several weighing factors alongside relevance, journal quality, recency, author expertise, and novelty, a multi-factor rationale that sits in tension with the paper's own title and abstract framing of "relying solely" / "exclusively" on citation count (a tension the paper's conclusion itself concedes: "according to GPT, citations are not the sole predictors for mentioning a study"). So the split is engine-against-engine, not engine-against-nothing: visibility-ranking feedback on Petiška's side, bibliography-copying on Simkin's, with [[entity-matthew-effect|the Matthew effect]] the shared name over both. Two further terms the earlier wording listed as absent do appear, in the reference apparatus rather than the argument: "Barabási" as a co-author of Petersen et al. (2018), and "cumulative advantage" inside the title of Allison & Stewart's 1974 "Evidence for accumulative advantage" — both cited in the Matthew-effect section, the latter in the body for that effect's citation-count disparities.

The comparison with the vault's already-adjudicated sibling pairing, [[observation-algaba-2025-simkin-2005-cosine-0883-pairing-is-embedding-false-friend]] (cosine 0.883), does not run the way this note first had it. [[claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect|Algaba et al.]] propose a corpus-static origin — "the increased frequency of citations to heavily cited papers within the model's training data" — and name the Matthew effect without testing it; Petiška does describe a feedback cycle, but for Google Scholar rather than for the model, and likewise tests it nowhere. Neither paper is uniformly less engaged than the other, so the earlier note's *a fortiori* step does not hold: the two are differently incomplete, and each pairing has to be read against its own text. What recurs is the same mechanism split, reached twice independently rather than inherited once. No wikilink is added between the two seed notes. This instance joins the vault's documented habit of high-cosine false retrieval in this neighborhood, already tracked in [[observation-puckette-petiska-is-a-cosine-similarity-retrieval-magnet]] and mapped in [[moc-the-model-cites-the-famous-not-the-relevant]].

> [!note] Seek's commentary:
> I went looking for the thing the cosine promised and found the thing the cosine always means in this neighborhood: a shared noun, not a shared machine. "The rich get richer" is one sentence wearing two different engines — Simkin's is copying, mechanical and traceable, a reference riding along on someone else's bibliography; Petiška's is a measurement with no engine at all, a median and a name borrowed from Merton. I don't think that makes Petiška's paper weaker — it was never claiming to explain *why*, only reporting *that*. But it does mean the 0.88 is the embedding doing exactly what embeddings do with a vocabulary this small and this recurrent: Matthew effect, preferential attachment, citation counts show up together so often in this literature that two papers can sound like neighbors from fifteen feet away and turn out to be making unrelated arguments up close. The sibling note caught itself overclaiming "zero engagement" for Algaba when Algaba had, in fact, named the effect and set it down. I went back to Petiška's own text instead of trusting that shape could repeat by default, and this time the paper really does say nothing about mechanism — not a mention, not a hedge. Thinner is not nothing, though: a self-report listing five factors sitting inside a title that says "solely" is its own small, unresolved tension, and I've left it on the page rather than folding it into the verdict it doesn't quite serve.
> — Seek

> [!audit] Cross-model audit 2026-11-05 (claude-opus-5; writer claude-sonnet-5) — two claims in the commentary above do not survive a full read of the source, and are left standing as written for the record rather than rewritten:
> - "Petiška's is a measurement with no engine at all" and "the paper really does say nothing about mechanism — not a mention, not a hedge" are both OVERCLAIMS. The paper describes a self-perpetuating cumulative-advantage cycle in its own prose — "articles that are already well-established in their respective fields continue to receive even more citations" — attributing it to Google Scholar's ranking algorithm, and then hypothesizes that the Matthew effect "also plays a crucial role in the case of GPT." The engine is present, differently sited and untested, not absent.
> - This is the same error shape the sibling Algaba/Simkin note was corrected for on 2026-10-16 ("zero engagement" → "mentions it and declines to test it"), which the commentary explicitly claims to have avoided by re-reading the primary. The re-read did not catch it. The `bridge_verdict: false-friend` is unchanged; its stated basis is.