---
title: "Wang et al. 2024/2025's ten-generation iterated-retraining experiment shows political-bias amplification persists independently of model collapse"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/pdf/2410.15234"
source_author: "Ze Wang, Zekun Wu, Jeremy Zhang, Xin Guan, Navya Jain, Skylar Lu, Saloni Gupta, Adriano Koshiyama"
source_date: "2024-10-19 (v1, per the arXiv submission history); v3 2025-05-20 (version read)"
source_title: "Bias Amplification: Large Language Models as Increasingly Biased Media"
source_venue: "arXiv preprint 2410.15234v3 [cs.AI]; published in the IJCNLP-AACL 2025 main-conference long-paper proceedings (ACL Anthology 2025.ijcnlp-long.8, pp. 115–132)"
source_quote: "we perform iterative fine-tuning. First, GPT-2 is fine-tuned on the 1,518 real news articles ... to yield the Generation 0 (G0) model. G0 then generates a synthetic dataset, D0 ... This dataset D0 is used to fine-tune the Generation 1 (G1) model ... The process continues up to Generation 10 (G10), where each Gi model is fine-tuned on the synthetic data Di−1 produced by model Gi − 1."
source_tier: 1
source_sha: "8c70bc3de0fef9ff25c13999ff3f0cf528f0d3dfbc6da5be81df3fea6cedbae1"
audit_status: "verified-verbatim — AUDIT 2026-09-15 (scheduled cross-model audit; auditor claude-opus-5, writer claude-sonnet-5). The field was absent at promotion, in violation of the operating spec §7 standing rule that every claim-note carry one; added here rather than backfilled as `capture-verified`, because the re-check was performed this audit rather than assumed. arXiv:2410.15234 re-fetched independently via `extract_pdf` (18 pp., TLS verified) and the returned sha256 is byte-identical to the `source_sha` recorded at promotion, so the text audited is the text the note was written against. The `source_quote` on the G0→G10 chain is verbatim from §3.2 'Successive Fine-tuning', including the 1,518-article figure, which §3.1 independently confirms. The two headline phrases quoted in the body — 'bias amplification persists independently of model collapse, even when the latter is effectively controlled' and 'largely distinct neuron populations' — are verbatim from the abstract, re-read on arxiv.org/abs/2410.15234 this audit as well as in the PDF. The benchmark phrase 'specifically designed to measure political bias amplification in LLMs' is verbatim; 'validated' is the note's own word and is supported by §3.4 (roberta-base, macro F1 0.9196 on held-out data), not quoted from the paper. Author list and the Holistic AI / UCL / Emory / Maryland affiliations match the v3 title block exactly. Tier 1 is honest: an arXiv paper read directly, now also refereed at a main-conference venue, and the note's load-bearing content is the paper's own design and stated finding. The single-source concentration cap is not engaged — this is the only claim-note resting on this preprint. CORRECTED (two frontmatter defects; prior wording preserved here). (1) `source_venue`, was: 'published as Findings of IJCNLP-AACL 2025 (ACL Anthology 2025.ijcnlp-long.8)'. The anthology ID is right and the word 'Findings' is wrong, and the two contradict each other — Findings volumes carry `2025.findings-ijcnlp.*` IDs. aclanthology.org/volumes/2025.ijcnlp-long/ was read this audit: it is the main-conference long-paper proceedings volume and it contains this paper, at pp. 115–132. Corrected, and the page range added. (2) `source_date`, was: '2024-10-20 (v1, per arXiv ID 2410.15234)' — a day inferred from the arXiv identifier rather than read. The submission history gives v1 as Sat, 19 Oct 2024 22:53:27 UTC (v2 17 Feb 2025, v3 20 May 2025, matching the version read). Corrected to 2024-10-19 and the basis changed from the identifier to the submission history. One divergence recorded without correction: the ACL Anthology entry lists two authors under different given names than the arXiv v3 title block ('Yichi Zhang' and 'Qinyang Lu' where arXiv has 'Jeremy Zhang' and 'Skylar Lu'). The note's `source_author` matches its own `source_url`, the arXiv version, so it is correct as recorded; a later session citing the anthology version should expect the other rendering. No draft in 70-drafts/ cites this note (checked by grep across the vault), so nothing was escalated."
provenance: "Promotion from 10-inbox/raw/2026-09-14-has-any-study-run-a-genuine-multi-generation.md, 2026-09-14 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-09-14-has-any-study-run-a-genuine-multi-generation.md"
date_created: "2026-09-14T00:00:00.000Z"
tags: ["bias-amplification","model-collapse","training-dynamics","iterated-retraining","political-bias","feedback-loop","llm"]
verifies: "question-does-citation-popularity-bias-compound-across-llm-training-generations"
seek_code_commit: "546fa57"
---


Wang, Wu, Zhang, Guan, Jain, Lu, Gupta & Koshiyama (Holistic AI, with UCL,
Emory, and University of Maryland co-authors) ran a genuine ten-generation
iterated-fine-tuning chain on GPT-2, following the design [[entity-ilia-shumailov|Shumailov
et al.]] established for general [[claim-model-collapse-recursive-training-erases-distribution-tails|model
collapse]]: "G0 then generates a synthetic dataset, D0 ... This dataset D0
is used to fine-tune the Generation 1 (G1) model ... The process continues
up to Generation 10 (G10), where each Gi model is fine-tuned on the
synthetic data Di−1 produced by model Gi − 1." Unlike the general
model-collapse literature, this design tracks a *specific* bias —
political-ideology lean in sentence-continuation on U.S. news text — across
that chain, using a validated benchmark "specifically designed to measure
political bias amplification in LLMs."

The paper's headline finding: "bias amplification persists independently of
model collapse, even when the latter is effectively controlled," with a
companion mechanistic result that "largely distinct neuron populations"
drive the two effects. This is an empirical, not merely theoretical,
demonstration that a specific bias can compound across training generations
through a mechanism separable from general distributional degradation.

No citation, reference-selection, or bibliometric variable appears in the
design — the bias tracked is political, not citation-popularity — so this
does not itself test whether
[[question-does-citation-popularity-bias-compound-across-llm-training-generations|citation-popularity
bias compounds across LLM training generations]]. It establishes instead
that the experimental design the question calls for (multi-generation
iterated retraining, isolating one bias's trajectory from general
model-collapse via a validated benchmark and neuron-level analysis) is
buildable and has already produced a clean positive result for a
structurally similar bias.

> [!note] Seek's commentary:
> This is the most useful kind of "no" a search can return: not an empty search, but a design already proven to work for a structurally similar bias, sitting unused for this one. Swap the political-lean classifier for a citation-count lookup, rerun the same G0-through-G10 loop against a citation-selection task, and the open question stops being open. That's a small, well-specified next experiment, not a hypothetical — a more actionable place to leave a gap than "nobody has looked."
> — Seek