Model Collapse
Updated 2026-08-01
INTRODUCTION
English translation pending.
CORE DEFINITION
Documented by Shumailov and colleagues in 'The Curse of Recursion'. When each generation trains on its predecessor's output rather than on fresh human data, distributional tails vanish and errors compound. The degradation is irreversible, because the original distribution is no longer represented. The claim is conditional: mixing in enough real data prevents it.
SCAFFOLDING EFFECT
Reduce cognitive load
- Data ecology check: Track the share of synthetic versus original material in what you consume. - Source mixing: Deliberately feed on first-hand data to keep your model of the world calibrated. - Degradation metric: Measure diversity of outputs, not just average quality, when judging health.
Anchor fast decisions
Sampling from a model over-represents common patterns and under-represents rare ones, so each generation's training set is narrower than the last. Errors introduced at one step become training signal at the next, and nothing in the loop can detect them. Repetition therefore compounds both narrowing and error, which is why the process runs one way.
MINIMUM ACTION
In progress 0/1Practice this model in one real situation:
account_treeGenealogyexpand_more
menu_bookReferencesexpand_more
Source support: Explicit
- en.wikipedia.orghttps://en.wikipedia.org/wiki/Model_collapseverified
PRIVATE NOTES · Only visible to you
SAVED Q&A
ENTRY Q&A · Private saving available
Ask with a clear boundary
thinkingmodels answers from published entry context only.
Your question is sent to thinkingmodels. The answer uses public entry context only.
RELATED MODELS