Model Collapse
What happens when models train on their own output for generations — a real effect, and the version you've heard depends on an assumption nobody makes in practice.
Reviewed July 16, 2026Stable
When not to use it
- As an argument against synthetic data generally. Some of the field's best recent results — phi, distillation, RLVR pipelines — depend on it, under accumulation and filtering.
- As a prediction about the open web. The mechanism needs replacement; the web accumulates, and the honest answer at internet scale is that nobody has measured it.
- To explain a model that's simply undertrained. Blandness has many causes, and collapse is a specific distributional claim you can test for.
Reach for something else instead
- Data filtering and dedup address the real risk with none of the drama, and are what mature pipelines actually do.
- Accumulation — keep the real data, add synthetic — is the finding that matters, and it's a one-line policy.
- Verification — RLVR-style checkable signals — sidesteps the problem where a verifier exists, because you're not learning from output, you're learning from correctness.
Sources & further reading
- Shumailov et al. (2024), AI models collapse when trained on recursively generated data — the Nature paper; collapse under the replace regime.
- Gerstgrasser et al. (2024), Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data — the same question under accumulation; test error plateaus rather than diverging.
- Shumailov et al. (2023), The Curse of Recursion: Training on Generated Data Makes Models Forget — the earlier arXiv statement of the mechanism.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Citing the Nature result without the regime. Collapse follows from replacing data each generation; under accumulation, the same experiment plateaus.
- Reading it as an argument for avoiding synthetic data. The finding argues for keeping your real data, which is a different instruction entirely.
- Assuming it's visible by reading outputs. The degradation is distributional — narrowed tails, converged lengths — and looks like fluent, confident prose right up until it matters.
At a glance
FieldMachine Learning
Mechanismtails truncate, errors compound
Requiresreplacing real data with synthetic
Under accumulationlargely doesn't happen
DifficultyIntermediate