---
title: "Does Simkin & Roychowdhury's copying-driven citation skew — rather than citation-count inflation generally — explain the training-corpus distribution behind GPT-4's citation-popularity bias?"
type: "question"
status: "open"
date_raised: "2026-10-16T00:00:00.000Z"
tags: ["citation-copying","matthew-effect","preferential-attachment","llm-citation-bias","training-data-exposure","mechanism-comparison"]
---


Raised while promoting
[[10-inbox/raw/2026-10-15-does-gpt-4s-citation-selection-bias-toward-highly|a 2026-10-15 capture]],
itself the forager's second run at the same cosine-0.883 seed pair that
[[observation-algaba-2025-simkin-2005-cosine-0883-pairing-is-embedding-false-friend|the
2026-10-14 hop]] already adjudicated as an embedding false friend at the
mechanism level. That note names one live thread it left untested: "whether
Simkin's copying model explains the training-corpus skew Algaba's first
hypothesis takes as given — upstream of the finding, not a competitor to
it." This question operationalizes that thread.

[[claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect|Algaba
et al.]] propose, as "one plausible hypothesis," that GPT-4's citation bias
comes from "the increased frequency of citations to heavily cited papers
within the model's training data" — a claim about the training corpus, not
about why the corpus looks that way.
[[claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties|Simkin
& Roychowdhury's copying model]] explains *one* route to a citation
ecosystem's skew toward already-famous papers: reference-list copying, where
"a majority of scientific citations are copied from the lists of references
used in other papers," mechanically feeding a paper's count forward into its
own future odds. But citation-count inflation has other routes too —
independent merit recognition, media attention, open-access availability,
institutional prestige — and no located source tests whether the
training-corpus skew Algaba's hypothesis takes as a bare fact is itself
*attributable to* Simkin-style copying specifically, as opposed to one of
these other routes.
[[claim-mozafari-2026-pretraining-exposure-explains-llm-popularity-better-than-wikipedia|Mozafari
et al. 2026]] directly confirms the general "LLM preference is inherited
from training-corpus exposure" mechanism-class — for entity popularity
broadly, using OLMo/Dolma's open training data — but for entities, not
papers or citation counts, and without engaging copying as a candidate
cause of the exposure distribution itself.

What would resolve this: a study that (a) takes a citation ecosystem whose
skew is independently shown — e.g. via misprint-fingerprinting, as in
[[claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original|Simkin
& Roychowdhury's own 2002 method]] — to be copying-driven rather than
merit-driven; (b) traces that same ecosystem's text into an LLM's training
corpus (an open-training-data model, per the Mozafari precedent, since a
closed model like GPT-4 cannot be audited this way); and (c) shows the LLM's
output bias tracks the copying-attributable portion of the skew
specifically, not citation count in general. No such study was found across
a targeted search (multiple query formulations, direct inspection of the
Algaba and Mozafari primaries).

Not the same question as
[[question-does-citation-popularity-bias-compound-across-llm-training-generations]],
which asks whether the bias strengthens across successive *model*
generations as LLM-selected citations re-enter training data. This question
is about the *human* ecosystem's citation skew — whether it is
copying-produced in the first place — not about whether LLM output feeds
back into it.

## Progress log

- 2026-10-16: Opened at promotion. No source splices the two mechanisms;
  the gap is real and specific, not a search failure — Simkin &
  Roychowdhury's paper predates LLM citation behavior by two decades and
  makes no claim about machine training, and Algaba et al.'s Discussion
  proposes training-data frequency without citing Simkin, Roychowdhury, or
  any copying-based account of why that frequency distribution looks the
  way it does. [[claim-ghostcite-augmix-propagation-root-cause-is-shared-tool-not-reader-copying|GhostCite's
  AugMix finding]] — propagation traced to a shared tool rather than
  reader copying — is a reminder that "looks like inherited copying" and
  "is inherited copying" have come apart before in this vault's own
  material, which is one more reason this question wants a direct test
  rather than a plausibility argument.
- 2026-10-18: A second, independent search session (one day later, fresh
  query formulations) reconfirms the gap rather than closing it: no source
  located splices the copying mechanism to the training-corpus distribution.
  Two papers new to the vault were surveyed directly —
  [[claim-niimi-2025-citation-count-predicts-llm-bibliographic-memorization-threshold|Niimi
  2025]] (citation count as training-data-redundancy proxy; near-verbatim
  memorization above ~1,000 citations) and
  [[claim-liu-chen-liu-lu-murray-2026-llms-overcite-popular-older-papers-affirms-algaba|Liu
  et al. 2026]] (six-LLM masked-citation audit affirming Algaba's
  popularity-bias finding) — and neither tests, cites, or engages a
  copying-specific account; Liu et al.'s full text does not use the word
  "copying" anywhere. Still open.