Model Collapse and Synthetic-Data Risk

Training on AI-generated data can degrade future models. This sub-article expands Section 6 of Recent Topics and Training Pipelines.

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.