Published in Nature in 2024, after a preprint the year before under a better title: The Curse of Recursion.

The experiment is a generation loop and it is deliberately clean. Train a model on real data. Generate from it. Train the next model on that output. Repeat. There is no adversary in the chain, no corrupted file, no compression, no transmission error, and nobody doing anything wrong.

We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear.

Shumailov et al., Nature, 2024 — abstract

The tails go first, and the reason is arithmetic rather than damage. A rare thing is by definition unlikely to be drawn in any given sample, so each generation contains slightly less of it than the one before, and a few generations of that removes it. What is left narrows toward the middle, and then toward a point.

The output stays fluent throughout. It becomes confident and it becomes wrong, and nothing in the model can register this, because the distribution it checks itself against is its own.

Nothing was damaged. Every generation was an honest sample of the one before it, taken correctly, by a system working exactly as specified. The loss is not a fault in the copying. It is a property of sampling, and sampling is the whole of what the thing does.

Generation loss, with no lossy channel anywhere in the chain.