SIRS Model for Generative AI Data Contamination: MATLAB Implementation and Numerical Analysis

This project uses the SIRS model from epidemiology as a framework to innovatively study the phenomenon of generative AI data contamination, providing complete MATLAB code implementation and numerical datasets. The core goal is to understand the impact of AI-generated content on the data ecosystem through a mathematical modeling perspective, and to provide a scientific basis for AI governance. The project source is GitHub (author: kavindisubawickrama-blip, release date: August 11, 2026).

The rapid development of generative AI brings data contamination challenges: AI-generated content entering training sets leads to training data degradation, model collapse, amplified biases, and content homogenization. Studies show that by 2026, more than half of internet content may be AI-generated; if used to train the next generation of models, it will form a vicious cycle of "generation → contamination → performance decline".

Map the epidemiological SIRS model to the AI data contamination scenario: S (high-quality human data), I (contaminated data), R (filtered/marked data), and allow R→S transition. The mathematical model is a system of differential equations: dS/dt = -βSI + δR dI/dt = βSI - γI dR/dt = γI - δR Parameter meanings: β (infection rate, AI content penetration speed), γ (recovery rate, detection and filtering efficiency), δ (immunity loss rate, limitations of filtering mechanisms).

The project provides MATLAB code, including ODE solvers (supporting parameter scanning and initial condition changes), visualization tools (time series plots, phase diagrams, sensitivity analysis), and precomputed datasets. Typical simulation results include: stable equilibrium points (coexistence of three types of data), periodic oscillations (game between generation and detection), and critical points (contamination outbreak or purification success).

Key findings: Increasing the detection rate γ is the core of controlling contamination; data contamination has a threshold effect, requiring early intervention; filtering mechanisms need to balance precision and recall (to avoid deleting high-quality data by mistake). Policy recommendations: Establish AI content traceability mechanisms, continuously invest in detection technologies, protect the diversity of human-created data, and restrict recursive training of AI-generated content.

Model limitations: Uniform mixing assumption, simplification of data binary classification, ignoring technological progress, static parameters. Extension directions: Multi-category models (distinguishing content types/quality), spatial models (cross-platform/regional spread), game theory perspective (multi-party strategy interaction), combining machine learning theory (analyzing the impact on generalization ability).