# AI models collapse when trained on recursively generated data (Nature, 2024) Back to [[Users/@thesupermannx/thesupermannx-x-profile|@thesupermannx profile]] - **Journal:** Nature 631:755–759, published **Jul 24, 2024** - **DOI:** [10.1038/s41586-024-07566-y](https://doi.org/10.1038/s41586-024-07566-y) · PMID 39048682 - **Authors (per Crossref):** Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, Yarin Gal - **Affiliations:** not returned by Crossref ⚠️. Commonly reported as Oxford, Cambridge and others; not re-checked this run ## What it finds - **The defect:** training indiscriminately on model-generated data causes irreversible defects, and the tails of the original distribution disappear. - **How general it is:** shown for LLMs, variational autoencoders (VAEs) and Gaussian mixture models (GMMs). ## Posts about it - [[Users/@thesupermannx/Post Analyses/Post Analysis - Model Collapse (2026-09-26)|Sep 26, 2026 post]]: 1.32M impressions