As the internet runs out of fresh human-made text, images, and code, AI models are increasingly learning from each other instead. That shift to synthetic data could supercharge progress – or quietly poison the next generation of systems with subtle errors and sameness. Here is what you need to know before you let AI train on AI.