Large Language Model Autophagy: Quantifying Epistemic Decay and Governance Intervention in Recursive AI Training Ecosystems

This paper presents the first empirically validated, quantitative framework for understanding and governing epistemic decay in recursive AI training ecosystems. Model autophagy occurs when large language models recursively ingest their own synthetic outputs during iterative retraining cycles, producing progressive degradation of corpus integrity, output diversity, and factual reliability. The research constructs a four-layer causal architecture spanning session-level confirmation bias (seconds to minutes), corpus-level data contamination (weeks to months), generation-level parameter drift (quarters), and civilizational-scale epistemic consequences (years to decades). This architecture is formalized through 21 discrete-time recurrence equations governing eight measurable state variables: corpus integrity I(t), bias B(t), misinformation M(t), error propagation E(t), homogeneity H(t), diversity D(t), quality Q(t), and provenance integrity P(t). The framework is validated empirically through controlled GPT-2 124M recursive retraining experiments across 10 generations under progressive synthetic contamination (S(t) from 0.10 to 0.80). Phase 3a baseline experiments (5 replicate tracks, 55 observations) confirm exponential integrity decay with rate alpha = 1.93 (R-squared = 0.98) and an asymptotic floor of I = 0.468. Phase 3b governed experiments (5 replicate tracks, 55 observations) demonstrate that combined SPC-based monitoring (G3) and provenance filtering (G2) not only prevents collapse but partially reverses decay, maintaining I(10) = 0.894 versus the baseline I(10) = 0.489 (Mann-Whitney U = 0.0, p = 0.004). The calibrated Bias Reduction Factor BRF = 0.115 yields FIF * BRF = 0.179, satisfying the theoretical stability condition. Accompanying artifacts include: an interactive Anti-Autophagy Monitor simulator deployed at https://darutherford.github.io/model-autophagy/, a Streamlit dashboard for governance scenario exploration, validation scripts, and the complete empirical dataset (110 observations across 10 retraining generations). The governance framework maps to ISO/IEC 42001 and NIST AI RMF compliance pathways. Resource type: Journal article License: Creative Commons Attribution 4.0 International (CC BY 4.0) Related identifiers: Interactive simulator: https://darutherford.github.io/model-autophagy/ (isSupplementTo)