Model collapse: what happens when AI learns from AI
Model collapse is the gradual degradation of generative models trained, generation after generation, on data produced by earlier models rather than by people. Rare cases and nuance fade away, and outputs become poorer and more uniform.
Data as of October 22, 2026
Key figures
| Level | Advanced |
|---|---|
| Origin | 2023 |
| Introduced by | Shumailov et al. |
A plain definition
Photocopy a photocopy, then do it again: each pass loses a little detail. Model collapse is the AI equivalent. A model learns a distribution of data, and what it produces is a slightly impoverished version of that distribution. If the next model learns mostly from those outputs, it inherits the loss and amplifies it.
The mechanism
A generative model mostly reproduces what is common in its training data. Rare events, the “tails” of the distribution, are under-represented in what it generates. The authors identify several sources of error that compound: sampling (a finite dataset never captures everything), the limits of what the model can express, and imperfections in the learning process. Generation after generation the tails vanish, and then variety itself shrinks. The effect hits large language models, but also simpler ones such as variational autoencoders and Gaussian mixture models.
Origins
The term was popularised by “The Curse of Recursion: Training on Generated Data Makes Models Forget” by Ilia Shumailov and co-authors, posted on arXiv in May 2023. A version appeared in Nature in July 2024 under the title “AI models collapse when trained on recursively generated data”.
A concrete example
Take a model trained on recipes from all over the world. Have it generate a large batch of recipes, train a new model only on those, and repeat. Little-known regional dishes show up less and less, then not at all; after a few rounds the model mostly offers variations on a handful of very common dishes, worded in an increasingly repetitive way.
Why it matters, and common myths
A growing share of the web is generated content, which can end up in pre-training corpora. Hence the interest in watermarking, detection and source traceability.
- “Collapse is inevitable.” Gerstgrasser et al. showed in 2024 that keeping the real data and adding each new generation on top, instead of replacing it, avoided collapse in their experiments.
- “All generated data is toxic.” Well-designed, filtered synthetic data is a standard tool, for example in distillation.
Related terms
- Synthetic data · Artificially created data, often generated by another AI model, used to train, fine-tune or test a model when real data is scarce.
- Training data · The examples (text, images, audio and more) an AI model learns from; they shape what the model knows and where it goes wrong.
- Pre-training · The first training phase of a language model: it learns to predict the next piece of text across a huge corpus, with no specific task.
- Knowledge distillation · A technique that trains a small student model to imitate the outputs of a large teacher model, so it becomes lighter and cheaper to run.
- AI watermarking · A mark, often invisible, embedded in an image, audio, video or text to signal that it was generated by AI.
Full glossary
Frequently asked questions
Are today's AI models already collapsing?
There is no evidence for that. The effect has been shown in controlled experiments where each generation replaces real data with generated data. AI labs filter their data and keep human-written sources, which limits that scenario; how large the real-world risk is remains debated.
Should synthetic data be avoided altogether?
No. Carefully chosen and filtered synthetic data is routinely used to train models, notably through distillation. The problem lies in uncontrolled, looped use where it replaces real data.
Why is human-generated data becoming more valuable?
Because it carries the diversity and rare cases that models tend to smooth over. The authors of the founding study expect data from genuine human interactions to become increasingly valuable.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.