Model collapse: training a model recursively on model-generated data erases the tails of the original data distribution

"Model collapse" is the progressive, degenerative process by which a generative model trained on data produced by earlier generations of models — rather than on the original human-generated distribution — loses fidelity to that original distribution. In the Nature study that named the effect, Shumailov and colleagues showed that as synthetic output is fed back as training data across successive generations, "the tails of the original content distribution disappear": rare-but-informative events are sampled ever less often, are under-represented in each next generation's training set, and eventually vanish. The model converges toward a lower-variance, more homogeneous output that no longer reflects the diversity of the data it was originally meant to model.

The mechanism is statistical rather than architectural: it is driven by finite-sample error compounding across generations (rare events are missed when sampling), functional approximation error, and functional expressivity error, so it appears across model families, not just large language models. The paper characterises collapse as irreversible once the tails are gone — the lost information cannot be recovered by continued training on the degraded distribution.

The finding is the mechanistic content behind the "model-collapse objections" flagged as an unexplained critique term in claim-karpathy-llm-wiki-gist-canonized-compounding-pattern, whose watch flag asks whether that critique literature would land. It is also the statistical baseline that a later cultural-evolution reframing sets out to explain: see claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution, which argues the statistical description of collapse "lacks a linguistic explanation for which structures degrade, in what order, and why."