Model collapse
Etymology
From "collapse" in the sense of a distribution narrowing and falling in on itself. The term comes from research around 2023 showing that models trained recursively on synthetic data drift toward blandness and error, forgetting the rare cases in the original human data.
noun phrase · /MOD-ul kuh-LAPS/
The degradation that occurs when AI models are trained on data generated by earlier AI models, losing accuracy and diversity over generations.
What it names
A self-poisoning failure mode of the AI era: as the internet fills with machine-made content, the systems trained on it can decay by feeding on their own output.
Example
If future models scrape a web flooded with AI text, researchers warn of model collapse — each generation blander than the last.
Related idea
A generational analogue of overfitting: overfitting fails on new data within one model, while model collapse degrades a lineage of models trained on each other’s output.
See also