What Is Model Collapse?
Model collapse is a phenomenon where AI models trained on data that includes outputs from previous AI models progressively degrade in quality and diversity, losing the ability to represent the full distribution of the original training data and converging on a narrow, repetitive subset.
How Model Collapse Works
As AI-generated content floods the internet, future AI models increasingly risk training on AI-generated data rather than human-created data. Research has shown that this recursive training cycle leads to model collapse: each generation of models loses information about the tails of the data distribution, becoming less diverse and less capable over time. Imagine repeatedly photocopying a photocopy — each generation degrades further. This is a growing concern as the internet becomes saturated with AI content. Mitigations include curating verified human-generated training data, watermarking AI content to enable filtering, and maintaining archives of pre-AI internet data. Model collapse has been demonstrated mathematically and experimentally, making it one of the most discussed risks in AI training methodology.
Real-World Examples
A study showing that a language model trained on text from a previous AI model produces increasingly repetitive and generic output after 5 generations
An image generation model trained on AI-generated images gradually losing the ability to produce diverse faces, converging on a generic 'average' face
AI researchers archiving pre-2023 internet data to preserve human-generated training data before widespread AI content contamination