What model collapse is
Model collapse is a well-documented risk where a model's outputs degrade in quality and diversity over successive generations of training on synthetic data, especially data generated by other AI models. The core mechanism is that any generative model produces an imperfect approximation of the real data distribution, missing some of the rarer patterns and subtly amplifying its own biases. If a new model is then trained heavily on that generated output, it inherits those distortions, and if that process repeats across generations, the errors can compound.
The practical effect researchers have described is a narrowing: outputs become more generic, rare or unusual patterns get lost, and the model's sense of the true diversity of the real world erodes. This isn't a hypothetical concern, it's been studied and demonstrated in controlled settings where models were trained recursively on their own or each other's outputs.