Language models learn a probability distribution from human text. What happens when much of the available text was written by other models? Shumailov et al. (2024) ran the experiment: they fine-tuned a small model (OPT-125m) on Wikipedia articles, had it generate a corpus of the same size, trained the next generation on it, and so on. Given a passage on medieval church towers, generation 0 still talked about architecture; generation 9 listed "black-tailed jackrabbits, white-tailed jackrabbits, blue-tailed jackrabbits, red-tailed jackrabbits…"
The authors distinguish early collapse, in which the tails of the distribution are lost (the rare, the minority, the eccentric), from late collapse, in which the model converges to something that bears little resemblance to the original and has much less variance. The main cause is statistical: with finite samples, what is infrequent has some chance of not appearing, and what does not appear in one generation can no longer be learned by the next; each network's limits of expressivity and learning add further errors. Alemohammad et al. (2024) observed the same in images and, in honor of mad cow disease, called it model autophagy disorder.
It is not an inevitable fate. Gerstgrasser et al. (2024) showed that collapse happens when synthetic data replace real data, but is avoided when they accumulate alongside them; and the industry uses filtered synthetic data on purpose, in distillation or in reasoning training. The underlying worry remains: high-quality human text is finite (Muennighoff et al., 2025), and the web is filling up with slop. Not to be confused with mode collapse, which happens within a single training run.
For the distributional hypothesis the case is curious: if meaning is learned from use, what anchors meaning when the use is, more and more, that of machines?