As AI-generated text saturates the web, a feedback loop has emerged: models train on corpora contaminated by prior model outputs, producing outputs that in turn contaminate future corpora. The phenomenon of model collapse, where recursive training on synthetic data degrades output quality and diversity, was formally characterised by Shumailov et al. and has since been studied in language models, image generators, and mixed real-synthetic regimes. Every prior formal analysis, however, treats collapse as a single-chain process: one model lineage training on its own outputs. The real ecosystem is a network, with thousands of models sharing and polluting common data pools. This paper, arXiv:2606.05168, is the first to model that ecosystem-level cross-contamination mathematically.
Key Contributions
The central contribution is a bilayer coupled SIR/SIRS framework treating data corpora and AI models as two interacting populations. Each layer has susceptible, infected, and recovered compartments, linked by cross-layer transmission: contaminated data infects models during training, and infected models transmit synthetic artifacts back into data corpora. The SIRS variant, which the authors recommend as primary, incorporates immunity waning to reflect that filtered corpora and retrained models remain susceptible to re-contamination over time.
- A closed-form basic reproduction number R0 is derived via the Next Generation Matrix method: R0 = sqrt(beta_D * beta_M / [(gamma_D + mu_D)(gamma_M + mu_M)]). This places the bilayer contamination problem squarely within established epidemic threshold theory.
- Scenario-based calibration using public AI text prevalence data yields supercritical dynamics (R0 > 1) across three scenarios, ranging from an optimistic R0 = 1.10 to a pessimistic R0 = 6.63.
- Sobol variance-based global sensitivity analysis identifies synthetic-text detection (gamma_D, the data recovery rate) as the single highest-leverage parameter, followed by data contamination rate and model recovery rate. Turnover rates contribute negligibly.
- A bipartite agent-based model (ABM) confirms mean-field consistency (R2 > 0.96) for dense networks, with degradation under heterogeneity.
- GPT-2 contamination chain experiments and a matched-budget source-diversity ablation provide empirical grounding for the theoretical framework.
Methodology
The theoretical component is well-grounded in established epidemic modelling literature, drawing on the Next Generation Matrix formalism of Diekmann et al. and van den Driessche and Watmough. The bilayer structure is analogous to two-population epidemic models used in, for example, vector-borne disease modelling, and the authors apply standard bifurcation and threshold results carefully.
Calibration is explicitly described as scenario-based and illustrative rather than empirically measured. Three parameter sets are constructed from public data on AI text prevalence and model development practices. The authors are appropriately cautious here: these are plausibility arguments, not fitted ecosystem measurements. Sobol indices are computed over all six ODE parameters with N=512 base samples using Saltelli's sampling scheme, and interactions are modest (S_T - S_1 < 0.06), indicating R0 is dominated by main effects.
On the empirical side, GPT-2 (124M parameters) is trained across contamination chains on WikiText-103 and Tiny Shakespeare, with contamination fraction alpha varying from 0 to 1 across 192 single-chain runs. The matched-budget source-diversity experiment (1,088 additional runs) fixes total pool size while varying the number of source models K in {1, 3, 5}, isolating source diversity as a bilayer-specific variable distinct from what single-chain models predict.
Results
The empirical results are clear on the dose-response relationship. Perplexity degrades monotonically with contamination fraction: at alpha=1.0, WikiText perplexity rises from 33.52 at generation 0 to 126.92 at generation 7, while the alpha=0 control remains flat. Bigram diversity (Distinct-2) drops from 0.68 to 0.38 under full contamination, confirming support shrinkage as an independent degradation signal. AIC comparisons favour a growth model over a plateau for all contaminated chains, and overwhelmingly so at alpha=1.0 (delta AIC = -60.2).
The source-diversity experiment yields more nuanced results. Multi-source mixing (K=3 or K=5) modestly attenuates collapse relative to pure self-training (K=1) at alpha=1.0, with approximately 2 PPL reduction and Cohen's d of approximately 0.8. The effect does not appear at alpha=0.5, confirming contamination fraction as the dominant driver rather than source diversity. Statistical significance is borderline: one-sided exact permutation p=0.047 for K=1 vs K=5, with two-sided paired t-test p=0.068. The authors appropriately describe this as suggestive rather than confirmatory evidence.
The ABM threshold verification correctly classifies 18 of 20 parameter configurations as sub- or supercritical, with the two errors occurring near R0 = 1, exactly where mean-field approximations are expected to be least reliable.
Limitations and Implications
The authors are admirably transparent about the limitations, and they are significant. The mean-field ODE degrades under network heterogeneity, which is the realistic case for the actual AI ecosystem. The R0 estimates and intervention projections are conditional on the model structure and assumed parameter ranges, not empirical measurements of the ecosystem. GPT-2 at 124M parameters is a long way from production-scale LLMs, and whether collapse dynamics at this scale are representative of 7B+ parameter continual pretraining is an open question. The SIR mapping from the chain experiments is explicitly phenomenological: alpha is a mixture fraction, not a transmission rate, so the correspondence is structural rather than quantitative.
The source-diversity finding rests on a borderline one-sided p-value from 8 seeds, and the non-monotonicity (K=3 approximately equals K=5) weakens the mechanistic interpretation. A more conservative reading is that contamination fraction dominates and source diversity is at best a second-order effect within this experimental regime.
That said, the conceptual contribution here is genuine. Framing ecosystem-level model collapse as an epidemic process gives the field a principled vocabulary for discussing thresholds, interventions, and equilibria that single-chain analyses cannot provide. The Sobol analysis result, that detection-based filtering of contaminated data is the highest-leverage intervention, is a model-conditional finding but one with clear practical implications: investment in synthetic text detection and corpus filtering is more valuable than accelerating model retirement or retraining cycles. The identification of a herd immunity analogue, where sufficient real-data preservation can push the system subcritical, is a useful conceptual bridge to existing work on data accumulation as a mitigation strategy.
Future work needs larger models, data-driven calibration from actual corpus provenance measurements, and direct measurement of compartment state variables to move this from a phenomenological framework to a quantitatively predictive one. The bilayer structure also opens natural extensions to heterogeneous network topologies and time-varying transmission rates as AI text prevalence continues to grow.