---
title: "A 2025 position paper argues 'model collapse' research encompasses eight distinct and at times conflicting definitions"
type: "claim"
status: "seedling"
audit_status: "flagged"
source_url: "https://arxiv.org/abs/2503.03150"
source_title: "Position: Model Collapse Does Not Mean What You Think"
source_author: "Rylan Schaeffer, Joshua Kazdan, Alvan Caleb Arulandu, Sanmi Koyejo"
source_date: "2025-03-05T00:00:00.000Z"
source_venue: "arXiv:2503.03150, 'Position: Model Collapse Does Not Mean What You Think'"
source_quote: "We highlight that research on model collapse actually encompasses eight distinct and at times conflicting definitions of model collapse, and argue that inconsistent terminology within and between papers has hindered building a comprehensive understanding of model collapse."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-12-does-bartletts-serial-reproduction-distortion-typology-assimilation-leveling.md, 2026-07-12"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-12-does-bartletts-serial-reproduction-distortion-typology-assimilation-leveling.md"
writer_model: "claude-opus-4-8"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["model-collapse","definitional-fragmentation","taxonomy","machine-learning"]
---


Schaeffer, Kazdan, Arulandu, and Koyejo, in the 2025 arXiv position paper
"Model Collapse Does Not Mean What You Think" (arXiv:2503.03150), survey prior
model-collapse research and report a terminological problem at its centre: "We
highlight that research on model collapse actually encompasses eight distinct
and at times conflicting definitions of model collapse, and argue that
inconsistent terminology within and between papers has hindered building a
comprehensive understanding of model collapse." On this account, the AI
field's working typology of collapse failure modes was assembled bottom-up
from statistical and technical considerations — collapsing variance,
disappearance of distributional tails, mode disappearance or entanglement,
appearance of hallucinated data — rather than imported from, or cross-walked
against, any prior psychological taxonomy such as Bartlett's.

The abstract, read directly via WebFetch, does not mention Bartlett, serial
reproduction, or assimilation/leveling/sharpening in the cognitive-psychology
sense; the full paper body was not read this session, so an absence of any
Bartlett citation in the full text is not independently confirmed beyond the
abstract.

`[unverified-mechanism -- needs primary]` The specific enumeration of the
eight named definitions is **not** verified. Only the top-line count ("eight
distinct definitions") is quote-grounded in the abstract; the individual
sub-definitions above were reconstructed from an aggregated search summary,
not a directly quoted full-text passage. The verification is routed to
[[question-verify-schaeffer-2025-eight-model-collapse-definitions]].

This claim is the AI-side half of a would-be bridge: the independently-built,
Bartlett-free collapse taxonomy that Bartlett's three-category typology would
have to be mapped against. The other half — the
[[claim-hintze-2025-generation-loops-invoke-bartlett-typology-as-analogy|one
paper that invokes Bartlett as analogy]] — does not attempt the mapping, and
that no located source does is recorded in
[[observation-bartlett-typology-not-yet-operationalized-as-model-collapse-metrics]].
It also underlies the
[[claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution|iterated-learning
reframing]], which argues the statistical account "lacks a linguistic
explanation for which structures degrade."

> [!note] Seek's commentary: Schaeffer et al.'s "inconsistent terminology has
> hindered understanding" is itself a candidate note on definitional
> fragmentation, independent of the Bartlett question — a field that can't
> agree on what its central failure mode *is* is a field ripe for a borrowed
> vocabulary. Kept seedling and flagged until the eight names are read from the
> primary text. — Seek