Glossary Updates6 new terms added to the glossaries · October 11, 2026, 13:38 CEST
AI TechDocKnowledge

Glossary · 6 · Quality, risk and governance

Model collapse

Also known as: Recursive training collapse, Model autophagy

Model collapse is a degenerative process in which generative models trained on data produced by earlier models lose information about the original data distribution over successive generations — first its rare cases, then its variety — and produce increasingly narrow or distorted output.

  • Expert
  • Technical writers
  • Technical project managers
  • Developers

In one sentence

Model collapse explained: generative models trained on output of earlier models lose rare cases and variety over generations.

Example

A language model fine-tuned again and again on its own summaries of service reports gradually stops mentioning rare fault patterns that appeared only a few times in the original reports.

Why it matters on your learning path

  • Technical writers: Released documentation written from engineering facts is the kind of original data that keeps models useful. Keep it apart from AI-generated drafts and answers.
  • Technical marketers: Large volumes of generated web text dilute the original material later models learn from; distinctive, verified content keeps its value.
  • Technical project managers: When data is collected for training or fine-tuning, record its origin and limit the share of model-generated material.
  • Developers: Label generated data, keep original data sets, and test for lost rare cases when models are retrained on mixed data.

How it applies

  • Findings: Shumailov et al. showed for language models, variational autoencoders and Gaussian mixture models that the tails of the original distribution disappear when models are trained recursively on generated data. They stress the value of data from genuine human interactions.
  • Documentation: Rare cases — unusual faults, special operating modes, warnings that apply to a few variants — are exactly what collapses first. Keeping AI-derived content out of the source repository and the training data protects them.
  • Limits: The effect depends on how much generated data replaces original data; later studies discuss that accumulating rather than replacing original data can reduce it.

Model collapse vs. synthetic data

Synthetic data is generated on purpose and in a controlled way to complement real data. Model collapse is the unwanted result of training on generated data that replaces original data without control.

External references