Failures / Research term
Model collapse
A family of training failures in which recursively using generated data causes a model to lose parts of the original data distribution or amplify errors.
Studies of recursive training show collapse under particular assumptions about how synthetic samples replace or mix with real data. Rare events can disappear and later generations can drift. The result is not a law that synthetic data always harms models: generated data can help when its source, coverage, filtering, weighting, and connection to real observations are controlled.
Builder example
A training or fine-tuning pipeline should track where each example came from, which model generated synthetic records, how they were filtered, and what real evaluation distribution remains protected. The relevant risk is feedback without sufficient independent data and quality control, not machine authorship by itself.
A company generates thousands of synthetic tickets from model-written examples and trains on them.
Mix in real resolved cases, track provenance, sample rare cases deliberately, and audit for narrowing.
Common confusion: Model collapse concerns recursive data and distribution coverage. It differs from mode collapse in a generator and from generic prose that happens to sound repetitive.