# Model Collapse and Synthetic-Data Risk

Training on AI-generated data can degrade future models. This sub-article expands [Section 6 of Recent Topics](../recent.md) and [Training Pipelines](training-pipelines.md).

## What Is Model Collapse?
- When models are trained on outputs from other models, errors and diversity shrink over generations.
- Rare events and minority content may disappear; performance degrades.

## Causes
- Generative models produce biased or narrow samples.
- Recursive training amplifies errors and homogenizes outputs.

## Mitigations
- Preserve and curate **human-generated data**.
- Detect and filter synthetic data.
- Maintain provenance records (C2PA-style metadata).

## Why It Matters
The web is filling with AI-generated content; future training corpora must manage synthetic contamination.

## Sources
- Shumailov, I. et al. (2023). *The Model Collapse Problem.* arXiv.
- Alayrac, J. et al. on synthetic data risks.

*Back to [Recent Topics](../recent.md).*