What Is Model Collapse?

Model Collapse — a degenerative process in which generative models trained on data produced by earlier models progressively lose information about the true data distribution — narrowing diversity and degrading quality over successive generations.

As AI-generated content fills the web, models risk being trained on other models' outputs, which research has shown can cause them to "forget" rare cases and converge on blander, less accurate results. The governance implication is about data provenance and curation: knowing what is in training data, and preserving genuine human data, becomes a quality and fairness safeguard.

Source: Shumailov et al. (2024), Nature; machine-learning literature