Model collapse is the gradual damage that happens when AI models are trained on content made by earlier AI models instead of by people.
In plain words
As more of the web is written by AI, new models risk learning from that output. Each generation copies the previous one's picture of the world a bit imperfectly, and the rare, unusual cases are the first to disappear. After several rounds the model's output narrows and drifts. This makes genuine human-made data more valuable.
A quick example
Picture training a model on text, then training a second model only on the first model's writing, then a third on the second's. Unusual words and rare facts fade with each round, until the latest model repeats a narrow, bland version of the original content.
Where you'll meet it
- Planning training data as AI text spreads online.
- Research on when synthetic data helps and when it harms.
Related
Sources
- PaperAI models collapse when trained on recursively generated data, Shumailov et al. (Nature) · read 29 Sept 2026
- PaperThe Curse of Recursion: Training on Generated Data Makes Models Forget, Shumailov et al. (arXiv) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.