In laboratory experiments, indiscriminate training of models over generations on data produced by earlier models caused “irreversible defects” in which the tails of the original data distribution disappear; in the preprint’s language-model experiment, keeping 10% of the original data led to only minor degradation.

SocioLens project,

Claim LD09-SL-C30 (v1.0)

Provenance and status

Claim type: Estimate
Publication status: Eligible for publication, with a qualifier
Verification state: Checked against the abstract of the primary source only. Who checked this claim is not recorded for it; claim checks so far were AI-led, and no person has checked claims against their sources.
Does not license: Any evidence about human knowledge or society; this is a laboratory result about models
Limitations: Journal version read in abstract, preprint in full; small models; data replaced rather than accumulated (see LD09-SL-C31); a 2025 Author Correction changes one notation in the theory section and no result
Source (as cited): Shumailov, Shumaylov, Zhao, Papernot, Anderson & Gal, “AI models collapse when trained on recursively generated data,” Nature 631, 2024
Source type: Peer-reviewed journal article (recorded as “Peer-reviewed journal article”)
Institutional position: Independent academic (University of Oxford and University of Cambridge-led)
Maturity of the evidence: Contested evidence
Recorded date: 15 October 2026
Currentness / correction: This is the current version. No correction or concern is recorded on its sources.

Sources this version cites

Shumailov, Shumaylov, Zhao, Papernot, Anderson & Gal, “AI models collapse when trained on recursively generated data,” Nature 631, 2024 (Peer-reviewed journal article)
Read the original

Version

This page shows version v1.0. Its record never changes, even when newer evidence arrives; the site may add a note to the page or withhold a figure not cleared for public use.

This pinned version is also the current version.

Versions: v1.0 · current

Raw record: every recorded field, with its coded values

All 25 canonical Claim fields, verbatim, with the coded values derived from them. The plain-language labels above are read from these.

Claim type (coded): ESTIMATE (as printed: “ESTIMATE”)
Publication status (coded): ELIGIBLE (QUALIFIED) (as printed: “ELIGIBLE WITH QUALIFIER”)
Verification state (coded): CONTENT_VERIFIED_PRIMARY (ABSTRACT_ONLY)
Source type (coded): PEER_REVIEWED_JOURNAL_ARTICLE
Maturity (coded): CONTESTED

#	Canonical field	Canonical value (verbatim)	Normalized (Claim Value Vocabulary v1.0.2)
1	CLAIM ID	LD09-SL-C30	—
2	PUBLIC-FACING CLAIM	In laboratory experiments, indiscriminate training of models over generations on data produced by earlier models caused “irreversible defects” in which the tails of the original data distribution disappear; in the preprint’s language-model experiment, keeping 10% of the original data led to only minor degradation.	—
3	CLAIM TYPE	ESTIMATE	base_types: ESTIMATE
4	RESEARCH STRAND	C — Systems, decisions and shared knowledge	—
5	MATURITY	CONTESTED	states: CONTESTED
6	SOURCE	Shumailov, Shumaylov, Zhao, Papernot, Anderson & Gal, “AI models collapse when trained on recursively generated data,” Nature 631, 2024	source_ids: SRC-SL-SHUMAILOV-ET-AL-NATURE-2024
7	SOURCE TYPE	Peer-reviewed journal article	classes: PEER_REVIEWED_JOURNAL_ARTICLE
8	INSTITUTIONAL POSITION	Independent academic (University of Oxford and University of Cambridge-led)	—
9	SCOPE	Generative models trained recursively on their own outputs, in the laboratory	—
10	DESIGN	Simulation and fine-tuning experiments over successive generations	—
11	SAMPLE / DATASET	Gaussian mixture models, variational autoencoders and a small language model fine-tuned on a Wikipedia-based dataset	—
12	MODEL / TOOL / VERSION	OPT-125m and simpler models	—
13	COMPARATOR	Training on original data	—
14	MAIN RESULT	Journal abstract: “indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear”; preprint: perplexity degrades without original data, and keeping 10% of it leads to “only minor degradation”	—
15	INFERENTIAL TARGET	What happens to models when their training data are increasingly model-generated	—
16	DOES NOT LICENSE	Any evidence about human knowledge or society; this is a laboratory result about models	—
17	VERIFICATION STATE	CONTENT VERIFIED ABSTRACT	state: CONTENT_VERIFIED_PRIMARY · qualifiers: ABSTRACT_ONLY · verified_by: UNKNOWN
18	LIMITATIONS	Journal version read in abstract, preprint in full; small models; data replaced rather than accumulated (see LD09-SL-C31); a 2025 Author Correction changes one notation in the theory section and no result	—
19	CONTRADICTORY / SUPPORTING SOURCES	Accumulating data avoids collapse: LD09-SL-C31	—
20	GEOGRAPHY / LANGUAGE	Not applicable / English	—
21	DISTRIBUTIONAL LEDGER	The rare parts of a distribution are lost first	—
22	R01 / FIELD-MAP LINKS	LD-09, D3	—
23	WHAT WOULD CHANGE THIS CLAIM?	Evidence from training practice at scale	—
24	PUBLICATION STATUS	ELIGIBLE WITH QUALIFIER	state: ELIGIBLE · qualifiers: QUALIFIED
25	RECORDED DATE	2026-10-15	—

Field 22 references resolved against Field Map v0.2: LD-09 (LIVING_DEBATE); D3 (DOMAIN)

The sources as recorded, with their record IDs and coded source types:

Pinned Sources (field 6, version-specific)

“Shumailov, Shumaylov, Zhao, Papernot, Anderson & Gal, “AI models collapse when trained on recursively generated data,” Nature 631, 2024” → Shumailov et al. (2024), “AI models collapse when trained…” (SRC-SL-SHUMAILOV-ET-AL-NATURE-2024--v1.0) — Shumailov, Shumaylov, Zhao, Papernot, Anderson & Gal, “AI models collapse when trained on recursively generated data,” Nature 631, 2024; class PEER_REVIEWED_JOURNAL_ARTICLE (explicit public provenance pin)

Cited by Living Debate versions

Individual gain, collective narrowing? Does AI make science, creative work and shared knowledge more alike? (v0.1) LD-09

Cite this page

SocioLens project. (2026, October 15). In laboratory experiments, indiscriminate training of models over generations on data produced by earlier models caused “irreversible defects” in which the tails of the original data distribution disappear; in the preprint’s language-model experiment, keeping 10% of the original data led to only minor degradation [Claim LD09-SL-C30, version v1.0]. SocioLens. https://sociolens-web.pages.dev/claims/ld09-sl-c30/v1.0/

@misc{sociolens_ld09_sl_c30_v1_0,
  author = {{SocioLens project}},
  title = {In laboratory experiments, indiscriminate training of models over generations on data produced by earlier models caused “irreversible defects” in which the tails of the original data distribution disappear; in the preprint’s language-model experiment, keeping 10\% of the original data led to only minor degradation.},
  year = {2026},
  date = {2026-10-15},
  howpublished = {\url{https://sociolens-web.pages.dev/claims/ld09-sl-c30/v1.0/}},
  note = {Claim LD09-SL-C30, version v1.0}
}

The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.

Research-of-record provenance

Materialized by sociolens-materializer 0.6.0 at 2026-09-28T20:00:00.000Z; normalization contract SocioLens — Claim Value Vocabulary v1.0.2 — Pre-Spike Materialization Contract (v1.0.2). Content seal SHA-256 16e1f0ce8a0dfabb99c9e0170402f3f4f53bcc97479a66ff4dd78887a7c550f1.

Repository-authored inputs (not Drive exports)

SocioLens claim — LD09-SL-C30 v1.0 (v1.0) — written in the SocioLens repository by Project Lead (Claude, AI) on 2026-10-15; no Drive identity exists. Authority: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 15 October 2026 to build the LD-09 debate; a new Claim, first version, written from the project's LD-09 source notes, checked against full texts or official abstracts. Block “ENTRY 30 — LD09-SL-C30” SHA-256 934e3b3b4df31017258221d9813d5e963611ebe1f574b325528d7eb0ee0fe72f; fixture fixtures/authored/claims/LD09-SL-C30--v1.0.

Builder decisions (not Drive data)

VERSION = v1.0 — basis: Claim version taken from the repository-authored claim fixture version (SocioLens claim — LD09-SL-C30 v1.0); the first version of a new Claim of the LD-09 debate
CURRENT_POINTER = true — basis: repository-authored claim "SocioLens claim — LD09-SL-C30 v1.0" (Project Lead (Claude, AI), 2026-10-15); basis: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 15 October 2026 to build the LD-09 debate; a new Claim, first version, written from the project's LD-09 source notes, checked against full texts or official abstracts
FIELD_6_SOURCE_REFERENCE = SRC-SL-SHUMAILOV-ET-AL-NATURE-2024--v1.0 — basis: field 6 of ENTRY 30 — LD09-SL-C30 cites "Shumailov, Shumaylov, Zhao, Papernot, Anderson & Gal, “AI models collapse when trained on recursively generated data,” Nature 631, 2024"; the only field-6 segment of this ClaimVersion's repository-authored ledger block.
SOURCE_PROVENANCE_PIN = SRC-SL-SHUMAILOV-ET-AL-NATURE-2024--v1.0 — basis: explicit version-pinned public provenance of this ClaimVersion (Amendment v1.2 §6): the pinned field-6 Source record "Shumailov, Shumaylov, Zhao, Papernot, Anderson & Gal, “AI models collapse when trained on recursively generated data,” Nature 631, 2024".