Epidemiology of Model Collapse: Modeling Synthetic Data Contamination via Bilayer SIR Dynamics

In three linesEpidemiological study of model collapse from synthetic data training. Bilayer SIR/SIRS framework models cross-contamination between data corpora and AI models. GPT-2 experiments on WikiText and Shakespeare (192 runs) confirm dose-response degradation; R₀ > 1 indicates supercritical dynamics. Synthetic-text detection and filtering identified as highest-leverage interventions.

Summary generated by Claude — human-verified