AI glossary · Advanced · Data & evaluation · Language models
Model collapse
Model collapse is a phenomenon in which, as each new AI is repeatedly trained on data made by AI, the model gradually forgets the diverse reality of the original world and its outputs become monotonous or strange.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Foundations & applications
- Fields
- Data & evaluation, Language models
- Added
- 2026-10-09
How it works
A generative model learns patterns that appear often in its training data better and learns rare patterns less well. The outputs of such a model therefore contain more common expressions and fewer rare ones. If these outputs are used again as training data, rare cases shrink further, and the model's errors are passed straight on to the next model. When this process repeats over many generations, the 'tails' of the data distribution disappear, outputs narrow toward one side, and in severe cases the model may produce meaningless output.
How model collapse builds up
Diversity declines with each generation
- 1Model training
- 2Data generation
- 3Retraining on generated data
- 4Distribution narrows
- 1. Model training Learns common patterns better
- 2. Data generation Rare cases appear less
- 3. Retraining on generated data Errors are inherited too
- 4. Distribution narrows Tails disappear
Why it matters
AI-generated text and images are growing quickly on the internet, so training data collected from the web will increasingly contain AI-made content instead of human-made content. Training on it without filtering it out can cause model collapse unintentionally. This raises the value of original human-made data and makes recording data provenance important. It is also a risk that must always be weighed in today's training methods, which make heavy use of synthetic data.
Good to know
Using synthetic data does not automatically cause collapse. The general understanding is that mixing it with real data, applying quality checks and filtering, or keeping the original data in use rather than discarding it can greatly reduce the risk. The problem grows when models are trained generation after generation on AI output alone without verification. In short, the key is not 'ban synthetic data' but 'how to mix and filter it'.
Example
When AI companies train their next model, filtering out web documents that appear to be AI-written, or keeping the share of human-written original data at a set level, are typical practical measures against model collapse. When generating large volumes of math problems or code as synthetic data, quality is also managed by keeping only items that pass answer verification.