A phenomenon where machine learning models gradually degrade because they are trained on data generated by other models, including older versions of themselves. This leads to a loss of diversity and accuracy over time.