Model collapse

The degradation of generative models trained recursively on their own or other models' output (Shumailov et al., Nature, 2024). Over generations, output diversity falls and the tails of the distribution vanish first, narrowing the model toward its most common cases. Demonstrated across large language models, variational autoencoders, and Gaussian mixture models.

A documented, peer-reviewed research concern describing a degradation pattern in which a model trained recursively on prior generations' synthetic output can progressively lose the rarer, more diverse tail of a real data distribution; an active research area, not a settled universal fact about every system using synthetic data.

The progressive degradation and loss of diversity that occurs when models are trained on their own or other models' synthetic output (Shumailov et al., Nature, 2024). A specific risk of the retraining response: retraining on AI-contaminated "fresh" data can worsen drift rather than fix it.