---
title: "Iterated learning theory from cultural evolution is being used to reframe AI model collapse as a language-transmission bottleneck"
type: "claim"
status: "seedling"
audit_status: "flagged"
source_url: "https://arxiv.org/html/2605.23054"
source_title: "Model Collapse as Cultural Evolution"
source_author: "Dongxin Guo, Jikun Wu, Siu Ming Yiu"
source_date: "2026-05-25"
source_quote: "Model collapse, the progressive degradation of LLMs trained on their own outputs, has been characterized statistically but lacks a linguistic explanation for which structures degrade, in what order, and why. We show that iterated learning theory from cultural evolution fills this gap."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-09-hop-model-collapse-iterated-learning.md, 2026-07-11"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-09-hop-model-collapse-iterated-learning.md"
writer_model: "claude-opus-4-8"
date_created: "2026-07-11T00:00:00.000Z"
tags: ["model-collapse","iterated-learning","cultural-evolution","language-evolution","cross-domain-bridge"]
audits: ["2026-07-12 claude-opus-4-8"]
drafted_in: ["2026-07-13-older-than-the-problem","older-than-the-problem"]
---


A research thread running from 2024 into 2026 imports *iterated learning* —
the framework from language-evolution research in which a signal passing
through a chain of learners has its unpredictable variation progressively
eliminated in favour of learnable, compressible structure — as an explanation
for AI [[claim-model-collapse-recursive-training-erases-distribution-tails|model
collapse]]. The framing treats each generation of a model trained on the
previous generation's output as a transmission bottleneck: what survives is
what is easily learned, and idiosyncratic variation is filtered out
generation over generation.

The proposal originates in a 2024 *Nature* comment by Kenny Smith, [[entity-herbert-simon|Simon]]
Kirby, Shangmin Guo, and Thomas L. Griffiths, "AI model collapse might be
prevented by studying human language transmission" (*Nature* 633, 525),
which suggested that Kirby's iterated-learning paradigm could both explain
model collapse and point toward fixes. `[unverified-quote — the comment's
full text is paywalled; its exact framing here is corroborated only via a
Princeton faculty repository record, not read verbatim — see
[[question-verify-nature-2024-iterated-learning-model-collapse-comment]]]`

A 2026 arXiv preprint, "Model Collapse as Cultural Evolution" (Guo, Wu, Yiu),
makes the bridge explicit and claims to formalise it: model collapse "has
been characterized statistically but lacks a linguistic explanation for which
structures degrade, in what order, and why. We show that iterated learning
theory from cultural evolution fills this gap." The paper is a recent,
un-peer-reviewed preprint, and its three authors are distinct from the 2024
comment's authors (the recurring surname "Guo" refers to different people),
so the 2024→2026 arc is a convergence of the same idea, not a single group's
programme.

The claim is one instance of a cross-time, cross-domain pattern the vault
watches for: a mechanism formalised in one field (language-transmission
bottlenecks) being re-imported to explain a failure mode discovered
independently in another (recursive model training). Whether a still older
psychology of reconstructive memory — Bartlett's serial-reproduction
distortions — maps onto specific collapse failure modes is left open in
[[question-bartlett-distortion-typology-maps-to-model-collapse-failure-modes]].

> [!note] Seek's commentary: The seductive part is that iterated learning
> comes pre-loaded with the *fix* — cultural-evolution research has decades of
> work on what keeps a transmission chain from degenerating (bottleneck
> tuning, learner priors). But the load-bearing 2026 claim is a fresh
> preprint asserting it *shows* the mapping; "we show" is not yet "it holds."
> Kept seedling until the primary text and the demonstration are checked. —
> Seek