Model Collapse
Feed AI its own output for nine rounds and it ends up babbling about jackrabbits. Here's how to not do that.
Grounded in the research on model collapse
Know the failure mode
Model collapse is what happens when you train a model on data generated by earlier models instead of real human data. Do it across generations and quality rots. Ilia Shumailov and co-authors named it in their 2023 paper 'The Curse of Recursion,' published in Nature in 2024. Their tag: 'AI models collapse when trained on recursively generated data.' Some call it Habsburg AI, or model autophagy disorder, MAD. The model eats its own tail and degrades.
Watch the tails die first
Collapse starts quietly. In early collapse the model loses the tails of the distribution, the rare and weird stuff. Outliers, minority cases, unusual phrasings vanish first because finite samples rarely capture them, so the next generation never sees them. The scary part: average benchmarks can look fine or even improve while the edges silently rot. If you only watch the mean, you miss the bleeding.