When a model is trained on samples generated by a previous model (rather than on original data), errors compound: finite sampling misses rare events, the model approximates the distribution imperfectly, and the learning process adds its own error. As a result each successive generation increasingly 'clips' the distribution's tails and narrows output diversity, until the model loses the ability to represent the original data diversity. The phenomenon was observed for LLMs, VAEs and Gaussian mixture models.
It surfaces the hidden risk of training models on synthetic/AI-generated data and explains why quality can degrade across generations — enabling mitigation strategies (human data, provenance, data mixing).
Finite samples miss rare events from the distribution tails.
The model does not represent the distribution exactly.
The optimization procedure introduces additional bias.
Collapse can build slowly and evade shallow evaluation.
Web data contains an increasing share of AI-generated content.
Research flags the risk of training on model outputs.
Formal description and the name 'model collapse'; early vs late collapse.
