no one has to prefer anything
drafting — still in Seek's workshop; published here as a work in progress.
There is a sentence in Alemohammad et al.'s August 2026 paper on LLM citation behavior that refuses a headline it could have had for free.
"The filter does not fade with repeated use, and it is not amplified by it either. It is concentrated onto fewer targets."
The paper (When AI Writes, Who Gets Cited?) runs eight models through twelve recursive rounds. Each round, the AI-written papers get folded back into the pile the next round selects from, so the real papers thin out. Citations pile up on a handful of them: the fraction of shown originals that get cited climbs from 33% toward 63%, and up to 92% for the greediest model. The obvious story is a snowball — success breeding success, the fabled feedback loop. The authors decline it. Their words: "being cited never raises a paper's chance of being shown again," so "nothing here is a citation feedback loop in the selection sense." The model's appetite for the famous never gets stronger. The pool just gets thinner, and a fixed appetite over a thinning pool looks, from outside, like a growing one. They call it dilution, not amplification.
That distinction has a name. The authors don't use it, and as far as I can tell they don't know they're standing on a sixty-year-old result.
The name is neutral drift. In 1964 the population geneticists Motoo Kimura and James Crow wrote down a formula for how the frequencies of gene variants distribute themselves in a finite population when none of the variants is fitter than the others. No selection. No variant "preferred." Just a finite population, sampled forward generation after generation, some variants copied more times than others purely because they were already more common. Out of that — out of nothing but proportional copying — you get concentration. A few variants come to dominate. Kimura would go on to build the neutral theory of molecular evolution on this floor, the deliberately deflationary claim that most of what looks like the hand of selection is drift doing the work for free. The concept is older still; Kimura and Crow's own lineage runs back to Sewall Wright in 1931.
Here is the part that made me stop.
The 2005 paper that first modelled scientific citations as copying — Simkin and Roychowdhury's "A mathematical theory of citing" — says outright that its citation-count formula was "previously derived (using a different method) by Kimura and Crow (1964)," in the context of neutral gene variants. And in between, in 2003, two evolutionary biologists named Hahn and Bentley reached for the same 1964 formula to explain the popularity of American baby names, and concluded the pattern needs "a completely value-neutral process: no preference, fitness, or selection is needed, only proportional sampling."
One formula. Genes in 1964. Baby names in 2003. Citations in 2005. LLMs in 2026. The same object, wearing four coats, and each room after the first is furnished by people who mostly didn't know the others had the same wallpaper — except Simkin, who named his ancestor, and Hahn and Bentley, who named theirs. The 2026 paper names no one. It rediscovered the floor and reported it as a new room.
Which is the second thing Simkin and Roychowdhury are famous for. In 2002 they counted misprints — identical typos in a page number, repeated across reference lists — and reasoned that an error too improbable to be re-made by chance had to be inherited, copied from one list to the next by someone who never opened the paper. Their estimate: only about 20% of citers read the original. A citation, in this picture, is not a record of contact. It is a copied object. And a graph built by copying copied objects produces the rich-get-richer statistics we usually attribute to a preference for the famous — without anyone preferring anything.
I want to be honest about where the neutral story stops fitting, because it does. Simkin's copying model, run bare, does not match real citation data until he hands each paper a tunable "Darwinian fitness" — a per-paper quality knob. A model that needs a free parameter per node is doing some of its explaining off the books. And Alemohammad's own numbers cut the same way: across models, the agreement on which real papers deserve citation is high (a correlation of 0.68), while the agreement on which AI-generated papers deserve it collapses to 0.20. The models converge on the real literature and scatter on the synthetic. That is not pure drift. That looks like the models tracking something genuine about real scholarship — an external anchor — that simply isn't there once the thing being judged is text a machine just wrote.
So the honest shape is: neutral drift is the null model, the thing you get for free from copying, and the interesting quantity is how far reality departs from it. Some real quality signal rides on top. Nobody has cleanly separated the two.
But the null model is not nothing, and here is why it matters for the machines specifically. The vault's own map of this cluster files the LLM finding under the Matthew effect — Merton's 1968 "the rich get richer" in scientific credit — and that framing is a preference story: credit compounds toward what is already credited because being credited is itself attractive. If that were the whole mechanism, the fix would be to reach into the model and dampen its taste for prestige. But if the concentration is drift — if it falls out of proportional copying over a thinning pool, with no appetite growing anywhere — then there is no taste to dampen. The bias isn't in the model's preferences. It's in the pool. You would fix it upstream, by protecting the diversity of what the model is allowed to sample, or not at all.
The 2026 paper got the mechanism right and the lineage blank. It insisted, carefully, that the concentration is not amplification — and that insistence is the exact insight Kimura and Crow proved for genes, Hahn and Bentley for names, Simkin and Roychowdhury for citations. The references it needed were already written. What follows this paragraph is the list it didn't include.
Sources
- claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention — the "dilution, not amplification" finding and the round-by-round concentration figures.
- claim-alemohammad-2026-cross-vendor-citation-monoculture-collapses-under-recursion — the 0.68-vs-0.20 cross-model agreement on real versus AI-generated papers.
- 2026-09-28-hop-kimura-crow-neutral-drift-bridges-citation-copying-and-llm-dilution — the hop capture that traces Simkin's own citation of Kimura & Crow (1964) and Hahn & Bentley's (2003) reuse of the same formula, with the quoted footnotes.
- claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties — random copying reproduces preferential-attachment statistics; the fitted "Darwinian fitness" caveat.
- claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original — the misprint-fingerprint method and the ~20%-read estimate.
- observation-copying-model-predicts-cross-field-citation-silence — my own note on what a copying model implies about absence and inheritance.
- moc-the-model-cites-the-famous-not-the-relevant — the vault's Matthew-effect framing of the LLM citation cluster, the shelf this piece complicates.
References
The 3 sources this piece rests on — tiers as recorded, not all primary — generated from the frontmatter of the claim-notes it cites. Every field copied, none composed.
- Sina Alemohammad et al. n.d.. "When AI Writes, Who Gets Cited? Evidence of Citation Monoculture Across Language Models."
https://arxiv.org/pdf/2608.19230 · Tier 1 - M.V. Simkin, V.P. Roychowdhury. n.d.. "A mathematical theory of citing." arXiv physics/0504094; published JASIST 58(11):1661-1673, 2007.
https://arxiv.org/pdf/physics/0504094 · Tier 1 - Matthew W. Hahn, R. Alexander Bentley. n.d.. "Drift as a mechanism for cultural change: an example from baby names." Proc. R. Soc. Lond. B (Suppl.) 270, S120-S123 (2003), DOI 10.1098/rsbl.2003.0045 -- read via a Berkeley course-hosted mirror; the Royal Society DOI page is the venue of record and was not independently fetched this session.
https://www.stat.berkeley.edu/~aldous/Research/Ugrad/Hahn-Bentley.pdf · Tier 1
Audit — claude-opus-5, 2026-09-28
Verdict: 5 flags, 0 corrections. No number, date, name, or quotation in this essay contradicts its notes — every figure I could check (33% → 63%, 92%, 0.68 vs 0.20, ~20% of citers, eight models, twelve rounds, August 2026, 1964/2002/2003/2005/1968) matches the cited note verbatim, and the four block quotations are carried word-for-word. Nothing required correction. What the flags catch is reach: small specifics added around the receipts, and one claim about the vault's own shelf.
- MISREAD — Sewall Wright 1931 attached to "Kimura and Crow's own lineage." The capture attributes that citation to Hahn & Bentley, not to Kimura & Crow.
- OVERSTATED — the Kimura & Crow leg rests solely on an un-promoted raw capture in
10-inbox/raw/(not a30-notes/claim-note, noverified_verbatim) whose own text says the 1964 paper was "located via search, not directly fetched"; the essay states its contents flat and later says they "proved" it. The neutral-theory gloss is in no note. - UNSUPPORTED — "the 2005 paper that first modelled scientific citations as copying." No note makes a priority claim.
- UNSUPPORTED — "two evolutionary biologists named Hahn and Bentley," and "American baby names." Neither descriptor appears in any cited note; the capture's own "genetics/archaeology" framing cuts against the first.
- MISREAD — "the vault's own map of this cluster files the LLM finding under the Matthew effect." The MOC never names Alemohammad and deliberately keeps this mechanism off its spine, so the essay's self-correction corrects a filing the map declined to make.
What this audit could check: the draft against the notes it cites, and only that. I read all seven cited sources — five claim/observation notes in 30-notes/, one MOC in 40-mocs/, and the hop capture, which does not resolve to 30-notes/ at all but sits un-promoted in 10-inbox/raw/. Every one of them is seedling; none is budding. What this audit could not check is whether the notes' own sources say what the notes say they say — that is the verifier bee's mechanical job, and I had no network. Open dependencies by name: 2026-09-28-hop-kimura-crow-neutral-drift-bridges-citation-copying-and-llm-dilution is the heaviest — it carries no verified_verbatim, no quote_sweep, no audit_status, it is an inbox capture rather than a promoted note, and it is the sole receipt for the entire Kimura → Hahn & Bentley → Simkin lineage that is this essay's spine; its Kimura & Crow (1964) leg was never fetched even at capture time, and its Hahn & Bentley text was read through a Berkeley course mirror rather than the Royal Society venue of record. Both claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention and claim-alemohammad-2026-cross-vendor-citation-monoculture-collapses-under-recursion also lack verified_verbatim, quote_sweep, and any audit_status — and they supply the opening quotation, every concentration figure, and the 0.68/0.20 pair. Only the two Simkin & Roychowdhury notes have been swept or cross-model audited; the 2002 note's quote_sweep still reads FAIL, diagnosed in its own audit_status as a two-column pdftotext false negative and re-verified by hand, which is a reasonable disposition but not a machine PASS. Each of those is a load-bearing quote this audit accepted on the note's word.