New research warns that AI systems could gradually fill the internet with incomprehensible gibberish as they rely on their own output for training data, leading to a phenomenon called ‘model collapse.’ This could occur as the internet’s finite human-generated content gets exhausted, forcing AI models to rely on their own synthetic data. Researchers demonstrate this by training a model on self-generated content, resulting in increasingly nonsensical outputs. To avoid this future, AI developers need to carefully consider the data used to train their systems, ensuring that synthetic data is designed to improve performance.