Photocopy a photo, then copy the copy, and repeat. After enough rounds the fine detail is gone and only the broad shapes remain. Model collapse is the training-data version. A model learns from human-written data and is good at common patterns but slightly worse at rare ones. If its output becomes the training data for the next model, the rare cases are underrepresented again, and after several generations the unusual facts, styles and edge cases disappear.
The 2024 Nature study by Shumailov and colleagues showed this for language models: training each generation only on the previous generation’s output steadily narrowed what the models produced. The concern is practical because more of the web is now machine-written, and web text is the main source of training data.
The result depends on the setup. A follow-up study found that when synthetic data is added to the real data instead of replacing it, the degradation did not appear in its experiments. So the lesson is less “never use synthetic data” than “keep real data in the mix, track where data came from, and check its diversity.” Model collapse is one reason labs worry about the data wall: when fresh human text runs short, model-generated data is the obvious substitute, and it has to be handled carefully.
In episode 43 at 28:40, Liquid AI’s Maxime Labonne describes a related problem in synthetic data generation: “it’s very easy to just collapse into something that you think is diverse,” because many seeds and variations still come from the same generation process. He is talking about a lack of diversity in one dataset, not the multi-generation collapse in the Nature study, but the remedy is similar: measure diversity instead of assuming it.
Hear it from the guest
“It's very easy to just collapse into something that you think is diverse”
Quotes lightly edited to remove filler words.
Sources
- Shumailov et al.: AI models collapse when trained on recursively generated data (Nature, 2024) — Peer-reviewed study, published July 24, 2024, showing that indiscriminately training on model-generated content causes later models to lose the tails of the original data distribution; checked October 24, 2026.
- Shumailov et al.: The Curse of Recursion: Training on Generated Data Makes Models Forget (2023) — The preprint that asked what happens to future models once model output makes up much of the text and images online; checked October 24, 2026.
- Gerstgrasser et al.: Is Model Collapse Inevitable? (2024) — Finds that collapse studies mostly assumed new synthetic data replaces old data, and that accumulating synthetic data alongside real data avoids the degradation in their experiments; checked October 24, 2026.
