Research and technical references
Shumailov et al. — Model collapse (Nature, 24 July 2024)
Claim
Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson and Yarin Gal, “AI models collapse when trained on recursively generated data,” Nature 631, no. 8022 (24 July 2024): 755–759, DOI 10.1038/s41586-024-07566-y. Shows that carelessly training models on content earlier models generated causes “irreversible defects” in later versions: the rare, unusual cases gradually vanish. The effect showed up across several kinds of model (large language models, variational autoencoders, and Gaussian mixture models). The authors issued a correction in Nature 640, E6 (2025), DOI 10.1038/s41586-025-08905-3.
Sources
- Nature 631 (2024): 755–759 ↗
Referenced in
- Framework §2
- AI-SAF-N Data pillar
- Refining §3.1