Model Collapse in plain English.

Model collapse is a degradation effect that can occur when AI models are trained repeatedly on data that was itself generated by earlier AI models, rather than real, human-created data — errors and quirks from one generation get amplified and passed to the next, and rare or unusual patterns in the real world gradually get smoothed away. Researchers demonstrated the effect clearly in controlled experiments: a model trained on its own outputs across several generations degraded significantly in diversity and accuracy. It matters increasingly in practice because so much of the modern internet now contains AI-generated text and images, meaning newly scraped training data unavoidably contains some AI output mixed in with human-created content. Labs manage the risk by carefully filtering and tracking data provenance, weighting verified human-created sources more heavily, and using synthetic data deliberately and selectively rather than by accident. It remains a genuine, actively studied risk rather than a solved problem.