Issue Synthetic data may outweigh real data by 2030 while risking model collapse

Summary

This issue weighs the promise of synthetic data against the risk that it degrades AI. As human-generated data becomes harder to obtain, more AI companies are using AI-generated data, and some experts think it could matter more than real data by 2030. For now it is mainly used for narrow tasks such as training autonomous vehicles. Synthetic data works best when derived from real data, but it can flatten real-world nuance, amplify biases in the source sample and expose personal information through reidentification attacks. If models are increasingly trained on lower-quality synthetic material, they could degrade in a self-reinforcing cycle known as model collapse, a risk that grows as synthetic content fills the internet.