---
title: "Bias amplification"
type: "entity"
entity_kind: "concept"
status: "watching"
canonical_name: "Bias amplification"
aliases: []
first_seen: "2026-09-14T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["model collapse","iterated learning","Matthew effect","citation metrics"]
seek_code_commit: "546fa57"
---


In iterated model retraining, the property of a *specific* social or
statistical bias growing stronger across generations by a mechanism
argued to be separable from generic model collapse —
[[claim-wang-2024-bias-amplification-persists-independent-of-model-collapse|Wang
et al. 2024/2025]] report "largely distinct neuron populations" driving
the two effects across a ten-generation GPT-2 chain, for political-lean
bias specifically.

First substantial appearance in this vault as a dedicated concept,
arriving already load-bearing: it names the exact mechanism this vault's
[[question-does-citation-popularity-bias-compound-across-llm-training-generations|open
citation-popularity-bias question]] asks whether anyone has pointed at
citations rather than politics. Held at `watching` rather than `hub` on
first appearance, per this same research thread's own same-day precedent
([[entity-citation-monoculture]]) of not promoting a first-appearance
term to a full hub before it recurs, even when it is central to the claim
that introduces it.

## References
- [[claim-wang-2024-bias-amplification-persists-independent-of-model-collapse]]
- Capture: 10-inbox/raw/2026-09-14-has-any-study-run-a-genuine-multi-generation.md