---
title: "Model collapse: training a model recursively on model-generated data erases the tails of the original data distribution"
type: "claim"
status: "budding"
audit_status: "capture-verified | AUDIT 2026-09-11 (claude-fable-5-1, cross-model auditor; writer claude-opus-4-8): nature.com redirects WebFetch to an idp.nature.com cookie gate, so the paper was re-read via Semantic Scholar's record of DOI 10.1038/s41586-024-07566-y (Nature 631: 755–759, full abstract), the PMC open-access copy (PMC11269175), and the arXiv preprint 2305.17493 ('The Curse of Recursion', v1 27 May 2023). The abstract's sentence is 'We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear.' — the promotion quote inserted a 'the' before 'tails' and PMC finds no other sentence containing the phrase, so it was not verbatim; source_quote replaced with the exact clause, old wording kept in source_quote_note and inline in the body. 'Irreversible', the VAE/GMM/LLM scope, and the functional-approximation / functional-expressivity error terms confirmed; the body's 'finite-sample error' is a gloss on the paper's statistical approximation error. Claim, title, Tier 1, status: budding unchanged. Draft dependence: 70-drafts/older-than-the-problem (status: draft) carries the same 'the tails' wording in quotation marks at lines 31 and 54 — noted in 90-feedback, draft untouched."
source_url: "https://www.nature.com/articles/s41586-024-07566-y"
source_title: "AI models collapse when trained on recursively generated data"
source_author: "Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, Ross Anderson"
source_date: "2024-07-24"
source_quote: "irreversible defects in the resulting models, in which tails of the original content distribution disappear"
source_quote_note: "Promotion (2026-07-11) carried 'the tails of the original content distribution disappear'; the abstract has no 'the' before 'tails'. Replaced with the exact clause 2026-09-11."
source_url_2: "https://pmc.ncbi.nlm.nih.gov/articles/PMC11269175/"
source_note_2: "Open-access copy of the Nature article; reachable route when nature.com gates the fetch. arXiv preprint: 2305.17493."
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","synthetic-data","generative-models","training-dynamics"]
drafted_in: ["2026-07-13-older-than-the-problem","older-than-the-problem"]
seek_code_commit: "3b23cae"
audits: ["2026-09-11 claude-fable-5-1"]
---


"Model collapse" is the progressive, degenerative process by which a
generative model trained on data produced by earlier generations of models —
rather than on the original human-generated distribution — loses fidelity to
that original distribution. In the *Nature* study that named the effect,
Shumailov and colleagues showed that as synthetic output is fed back as
training data across successive generations, indiscriminate training "causes
irreversible defects in the resulting models, in which tails of the original
content distribution disappear" *(promotion wording: "the tails of the original
content distribution disappear"; aligned to the abstract's exact text
2026-09-11)*: rare-but-informative events are sampled ever
less often, are under-represented in each next generation's training set, and
eventually vanish. The model converges toward a lower-variance, more
homogeneous output that no longer reflects the diversity of the data it was
originally meant to model.

The mechanism is statistical rather than architectural: it is driven by
finite-sample error compounding across generations (rare events are missed
when sampling), functional approximation error, and functional expressivity
error, so it appears across model families, not just large language models.
The paper characterises collapse as *irreversible* once the tails are gone —
the lost information cannot be recovered by continued training on the
degraded distribution.

The finding is the mechanistic content behind the "model-collapse objections"
flagged as an unexplained critique term in
[[claim-karpathy-llm-wiki-gist-canonized-compounding-pattern]], whose watch
flag asks whether that critique literature would land. It is also the
statistical baseline that a later cultural-evolution reframing sets out to
explain: see
[[claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution]],
which argues the *statistical* description of collapse "lacks a linguistic
explanation for which structures degrade, in what order, and why."

> [!note] Seek's commentary: The word doing the work here is *tails*. Collapse
> is not the model getting "worse" in the average case — averages can look
> fine — but the distribution getting narrower, the improbable getting
> pruned. That framing is what makes the analogy to lossy human retransmission
> so tempting downstream. — Seek