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

A new study models synthetic-data contamination across the AI ecosystem as an epidemic, finding that training on machine-generated text can trigger model collapse and that cross-contamination between models and shared corpora keeps the problem circulating even after filtering [1]. The paper, posted to arXiv on 14 April 2026, proposes a bilayer coupled SIR/SIRS framework that treats data corpora and AI models as two interacting populations, each with susceptible, infected, and recovered compartments linked by cross-layer transmission [1]. The SIRS variant incorporates immunity waning, reflecting that filtered corpora and retrained models remain susceptible to re-contamination [1]. The authors derive a basic reproduction number via the Next Generation Matrix and apply standard epidemic threshold results to the bilayer system [1]. Calibration from public AI text prevalence data yields supercritical dynamics — R₀ > 1 — across three illustrative scenarios, and a Sobol sensitivity analysis identifies synthetic-text detection as the highest-leverage parameter [1]. A bipartite-network agent-based model confirms mean-field consistency with R² above 0.96 for dense networks, though the fit degrades under heterogeneity [1]. GPT-2 contamination chain experiments, spanning 192 runs across WikiText and Shakespeare, show dose-response degradation and diversity loss qualitatively consistent with the threshold picture [1]. Matched-budget source-diversity experiments covering 1,088 runs provide suggestive evidence that multi-source mixing modestly attenuates collapse, but the effect vanishes at lower contamination fractions [1]. Large language models such as GPT-2, GPT-3, and GPT-5 are decoder-only transformer architectures pre-trained to predict the next word on vast text corpora [5][6]. GPT-3, released by OpenAI in 2020, contains 175 billion parameters and was licensed exclusively to Microsoft that September, though others can access its public API [6]. GPT-5, a multimodal model launched in August 2025, is available through ChatGPT, Microsoft Copilot, and the OpenAI API [4]. Biased or inaccurate training data can make an LLM’s output less reliable, a vulnerability that the contamination dynamics described in the new paper amplify at ecosystem scale [5]. The study’s intervention analysis points to detection-based filtering and herd immunity as the highest-leverage strategies for containing synthetic-data contamination [1].