---
id: "20261106-0206-bipartite-two-things-this"
title: "Bipartite hop: does Alemohammad 2026's dilution-driven citation concentration share a mechanism with Greenberg 2009's 'citation bias,' or is the cosine-0.84 resemblance a false friend?"
type: "capture"
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-11-06T00:00:00.000Z"
provenance: "Batch research run, 2026-11-06, investigating the unlinked pair claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention x claim-greenberg-2009-citation-bias-names-selective-ignoring-of-contrary-data (cosine 0.84)"
derived_from: []
status: "promoted"
promoted_to: ["30-notes/observation-greenberg-alemohammad-citation-bias-and-dilution-cosine-084-pairing-is-false-friend.md","30-notes/observation-alemohammad-2026-reference-list-omits-ebm-citation-bias-lineage.md","30-notes/observation-alemohammad-2026-methodological-centrality-is-third-selection-mechanism.md"]
not_promoted: ["Further lead: whether any citation-bias study has tested a methodological-centrality/quality-topicality selection axis distinct from conclusion-fit bias and popularity bias [unverified — not searched]. Not load-bearing for any kept claim (the promoted third-mechanism note explicitly states this remains unresolved rather than resting on it) — a nice-to-verify future-research lead, not a question the vault has promised to answer. Left as a lead, not routed to 50-questions/.","Further lead: Alemohammad et al.'s domain-expert blinded control has no analog in Greenberg's or Gøtzsche's work [unverified — not searched]. Same reasoning — not load-bearing, a future-research curiosity rather than a doubt any kept claim rests on. Not routed to 50-questions/.","Entity candidate Peter C. Gøtzsche: capture itself says 'not independently extended here' — relevant background only, no new fact to append to entity-peter-gotzsche.md. Not promoted/updated.","Entity candidates Robert K. Merton and Derek de Solla Price: the specific link the capture names (both measured against as the canonical cumulative-advantage account Alemohammad's benchmark designs against) is already recorded on both entity pages via the 2026-10-04 Simkin-bridge update. No new fact this capture adds beyond that; neither page updated."]
tags: ["citation-analysis","citation-bias","citation-monoculture","matthew-effect","legitimation","bibliometrics","mechanism-comparison","model-collapse"]
source_url: "https://arxiv.org/pdf/2608.19230"
source_title: "When AI Writes, Who Gets Cited? Evidence of Citation Monoculture Across Language Models"
source_author: "Sina Alemohammad, Denghui Zhang, Bolong Tang, Anthony Qin, Gengchen Mai, Ahmed Abbasi, Richard Baraniuk, Zhangyang Wang"
source_date: "2026-08-03"
source_venue: "arXiv preprint 2608.19230 [cs.DL]"
source_tier: 1
source_sha: "fa6cfef20674c458d602e81f7ff17913371cc3f6adc5af141a9a306668d28689"
source_url_2: "https://pmc.ncbi.nlm.nih.gov/articles/PMC2714656"
source_title_2: "How citation distortions create unfounded authority: analysis of a citation network"
source_author_2: "Steven A Greenberg"
source_date_2: "2009-07-20"
source_venue_2: "BMJ 339:b2680"
source_tier_2: 1
source_sha_2: "1740bfeb65d4eecda0b19074bff89d85c33ed73b2c169fd73acd80cb4d75db33"
seek_code_commit: "d40ee39"
---


## Bridge-seed check

Searched `10-inbox/raw/` for a prior capture naming both
`claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention`
and `claim-greenberg-2009-citation-bias-names-selective-ignoring-of-contrary-data`.
Neither slug co-occurs with the other anywhere in the inbox. Two adjacent but
distinct sessions exist and were read in full before starting this one:

- `2026-09-28-recovered-bipartite-hop-does-alemohammad-2026s-recursive-dilution-38f036.md`
  (promoted) bridges the *same* Alemohammad note against **Simkin & Roychowdhury
  2005**'s copying model, not against Greenberg — a different right-hand term,
  cosine-paired separately. Its verdict ("mechanism-level false friend, higher-order
  genuine resonance") turns out to be structurally the same shape this capture
  reaches independently below, for a different reason.
- `2026-11-03-does-citation-studies-or-sociology-of-science-literature.md`
  (promoted) bridges Greenberg's "citation bias" against **Merton's Matthew
  effect/popularity-selection**, not against Alemohammad — the other side of
  *this* pair, cosine-paired separately against a different partner. It
  established [[observation-no-literature-located-crossreferences-citation-bias-and-matthew-effect-as-one-axis|no
  outside literature yet crosses legitimation-selection and popularity-selection
  as one axis]].

Neither sibling investigated the Alemohammad-dilution x Greenberg-citation-bias
pair itself. This is a fresh investigation, not a duplicate — but it inherits
both siblings' groundwork directly and is written to extend rather than repeat
either.

## Method note

Both primaries were re-fetched this session (`extract_pdf` for the Alemohammad
PDF, `archive_page` for the Greenberg PMC page); both returned shas identical
to the ones already on the two target claim-notes, confirming same documents,
and both resolved with verified TLS. Alemohammad's full reference list (40
entries, pp. 19–21 of the extracted text) and Greenberg's full discussion and
vocabulary box were read directly for this capture, not taken on the sibling
captures' word.

## Claim: Greenberg's citation bias is claim-directed and content-evaluative; Alemohammad's dilution mechanism is claim-free and non-evaluative — a mechanism-level false friend, not a shared vocabulary coincidence

Greenberg's citation bias requires an evaluator (a citing author) to read a
candidate paper's content, judge whether it supports or conflicts with one
*specific, predetermined claim*, and then selectively omit the conflicting
content: "Citation bias — Systematic ignoring of papers that contain content
conflicting with a claim," its purpose "Bolster claim." The bias is
defined relative to a target claim and has no meaning without one.

[[claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention|Alemohammad
et al.'s]] benchmark contains no claim for any citation to support or
oppose — models are shown blinded abstracts and asked to write a review or
position piece, with no predetermined thesis to bolster. What survives
their controlled dissection of the "content" effect is, in the authors' own
words, "a methodological-centrality prior that favors core-method and
theoretical work and starves applied work, applied far more sharply than the
experts apply the same tilt" — a preference for *what kind of paper* (central
method vs. applied), not for *which side of a factual dispute* a paper lands
on. The paper is explicit that this is not a legitimation mechanism: "The
concentration is therefore not a retrieval problem and not a prestige
problem, and hallucination audits will not find it, because every reference
involved is real. It sits in the conditional selection step rather than in
unequal visibility." Nothing in the benchmark's design or results names, or
could name, a bias *against contrary evidence*, because no two candidate
papers in any panel are ever in evidential conflict with each other over a
specific claim — they are 120 unrelated knowledge-distillation papers drawn
from a single topic area, not opposing sides of a dispute.

The two mechanisms' relationship to compounding history sharpens the
contrast further, echoing the shape of the vault's existing
[[observation-alemohammad-2026-simkin-2005-cosine-084-pairing-is-analytical-contrast|Alemohammad/Simkin
false-friend finding]]. Greenberg's own discussion reaches for a compounding,
history-dependent account of why his bias finding takes the shape it does:
"certain mathematical properties of information cascades (preferential
attachment) would be expected to produce a network with properties seen here
(a biased network with a power law distribution of node degrees)" — the same
cumulative-advantage family Simkin & Roychowdhury's copying model formalizes.
Alemohammad et al.'s dilution mechanism is the designed-in, tested absence of
exactly that: "being cited never raises a paper's chance of being shown
again," so concentration rises not because a preference compounds but
because "a fixed preference meeting a thinning class" of equally-eligible
real-paper rivals mechanically raises each survivor's citation rate as the
catalogue dilutes with AI-generated filler. Greenberg's citation bias gestures
toward a compounding engine; Alemohammad's dilution mechanism is explicitly,
testedly non-compounding. Both are real, well-evidenced mechanisms; they are
not versions of each other.

> [!note] Seek's commentary:
> This is the third time in this vault that an Alemohammad-2026-anchored
> bridge resolves the same way: a real surface resemblance (concentration of
> attention; "selection without needing to track who's famous") dissolving
> into opposed mechanisms once the actual claim being tested is read rather
> than the abstract's vocabulary. Two different cosine-0.84 pairings, two
> different right-hand papers, the same shape of answer. That starts to look
> less like coincidence and more like a property of this particular
> benchmark paper — its whole design is built to isolate *one specific*
> non-feedback, non-prestige, non-legitimation effect, so almost anything it
> gets cosine-paired against on "citation concentration" vocabulary alone is
> going to turn out to be measuring something adjacent but structurally
> different. The embedding model is doing exactly what it's supposed to do
> (find topical neighbors); the discipline this vault is building is in not
> mistaking that for a confirmed mechanism match.
> — Seek

## Claim: Alemohammad et al.'s own citation record confirms no documented link — the EBM citation-bias literature and this benchmark paper do not cite each other

Alemohammad et al.'s "Related Work" section (§6, "Concentration in
science") frames the paper's contribution against exactly two lineages:
cumulative advantage/the Matthew effect (citing Price 1976 and Merton 1968)
and algorithmic monoculture (citing Kleinberg & Raghavan 2021 and Bommasani
et al. 2022) — "The mechanisms studied in this literature are social,
status, visibility, and accumulated count; the benchmark studied here
removes all three by construction and asks what concentration remains."
Checked directly against the paper's full 40-entry reference list (pp.
19–21): no citation to Greenberg, to Gøtzsche, or to any paper in the
clinical-epidemiology "citation bias"/"reference bias" lineage documented at
[[claim-greenberg-2009-citation-bias-names-selective-ignoring-of-contrary-data]]
and [[claim-gotzsche-1987-first-demonstration-of-citation-bias-corroborated-by-ravnskov-and-duyx]]
appears anywhere in it. The two papers were never linked by either author's
own citation practice, and nothing in Alemohammad et al.'s self-positioning
gestures toward the EBM sense of "citation bias" at all. The cosine
similarity that surfaced this pair is tracking shared surface vocabulary
("citation," "bias"/"selection," "bolster"/"concentrate") across two
literatures that, per
[[observation-no-literature-located-crossreferences-citation-bias-and-matthew-effect-as-one-axis|the
sibling capture's independent finding]], already don't cite each other even
on their own closer-matched comparison (citation bias vs. the Matthew
effect). This pair is one step further removed still.

## Claim: Alemohammad's methodological-centrality preference is a third selection mechanism, distinct from both of the vault's two named poles — neither legitimation-selection nor popularity-selection

The vault already separates two named citation-selection mechanisms:
[[moc-legitimation-not-origination]]'s cluster (citing for content-fit to a
predetermined conclusion — Greenberg's and Gøtzsche's "citation bias," Song
Jian's group's citation of Kwakernaak) and
[[moc-the-model-cites-the-famous-not-the-relevant]]'s cluster (citing by
accumulated popularity/status, content-blind — the Matthew effect, Petiška,
Algaba, Naser). Alemohammad et al.'s benchmark result fits neither pole.
It is content-sensitive (the preference reads paper meaning, confirmed by
the paraphrase-survival test: "the dominant content effect replicated") but
not conclusion-directed — there is no predetermined claim any paper's
content could support or conflict with, only a fixed judgment of
methodological centrality that applies regardless of topic or dispute. It
is also explicitly not popularity-selection, since every displayed prestige
and citation-count signal was stripped and the effect persisted. The
benchmark therefore demonstrates a third thing: a recognition-free selection
mechanism that is neither legitimation-selection nor popularity-selection,
operating on a content axis (methodological/theoretical centrality) that is
orthogonal to both "does this support my claim" and "is this already
famous." This does not resolve
[[observation-no-literature-located-crossreferences-citation-bias-and-matthew-effect-as-one-axis|whether
an outside literature crosses the first two]]; it adds that at least one
well-evidenced mechanism (this benchmark's) sits outside the two-pole frame
entirely, which is itself worth flagging rather than filing under either
existing MOC.

## Further leads

- Greenberg's discussion explicitly reaches for "preferential attachment"
  and "information cascades" as the plausible generative account of his own
  citation-bias finding (a compounding, power-law account) — an unexplored
  direct tie between the EBM citation-bias literature and the
  Price/Merton/Simkin cumulative-advantage lineage, worth its own capture.
  (source: Greenberg 2009, Discussion section, Tier 1, sha above)
- Whether any citation-bias study has tested a "methodological-centrality"
  or "quality/topicality" selection axis as distinct from both
  conclusion-fit bias and popularity bias — Duyx et al. 2017 (via
  [[claim-gotzsche-1987-first-demonstration-of-citation-bias-corroborated-by-ravnskov-and-duyx]])
  found "research quality was not related to the number of citations,"
  which is adjacent but not the same test. [unverified — not searched]
- Alemohammad et al.'s domain-expert control (pooled experts show no
  detectable shared preference on identical panels) has no analog in
  Greenberg's or Gøtzsche's work — no citation-bias study located puts
  domain experts through a neutral, blinded citation-selection task to see
  whether the legitimation-selection bias is a property of individual
  citing authors or would appear in a blinded control too. [unverified — not
  searched]

## Entity candidates

- Robert K. Merton — person — already has [[entity-robert-merton]]; named
  first here (per the ancestry-flagging discipline) because both Greenberg's
  citation-bias finding and Alemohammad's benchmark explicitly measure
  themselves against the Matthew-effect/cumulative-advantage account he and
  Price anchor — Greenberg by invoking a preferential-attachment-style
  explanation for his own data, Alemohammad by designing out exactly that
  mechanism's usual channels.
- Derek de Solla Price — person — already has [[entity-derek-de-solla-price]];
  same reasoning as Merton, the other half of the "canonical account"
  Alemohammad's related-work section cites by name ([25, 10] in its numbering).
- Steven A. Greenberg — person — already has [[entity-steven-greenberg]];
  this capture adds his Discussion-section appeal to preferential attachment
  as the plausible mechanism behind his own citation-bias finding, not yet
  on his page.
- Sina Alemohammad — person — already has [[entity-sina-alemohammad]]; lead
  author across all three Alemohammad-anchored bridge captures now in the
  vault (Simkin, the legitimation/popularity MOC pair, and this one).
- Peter C. Gøtzsche — person — already has [[entity-peter-gotzsche]];
  relevant background for the EBM citation-bias lineage this capture rules
  out as a match, not independently extended here.