Abstract: This study aims to investigate the degradation risks of Generative Artificial Intelligence (GAI) models in self-training loops, with a focus on two core phenomena: content homogenization and the widening divergence between human and machine-generated texts. We select two representative generative models with distinct architectures and build an iterative self-training framework, using the proportion of human data in the training set (α) as the key hyperparameter. Under different initial values of α, we conduct controlled experiments combining two typical dynamic strategies—linear decay and exponential decay—and systematically evaluate the quality, diversity, and human-likeness of generated content using multidimensional performance metrics. The results show that, during self-training, GAI models exhibit a persistent decline in performance, a marked reduction in output diversity, and a gradual increase in the gap between human and machine-generated texts. The linear decay strategy can effectively slow down the decline of information entropy and help maintain content diversity, but it becomes increasingly vulnerable to the cumulative impact of model-generated data pollution in later stages. In contrast, although the exponential decay strategy leads to more pronounced performance fluctuations in the early phase, it achieves superior stability in the long run. Moreover, lightweight unidirectional language models (GPT2) are more prone to falling into a vicious cycle of noise amplification during self-training, whereas bidirectional encoder models (BART), endowed with stronger global modeling capacity, demonstrate greater robustness in the presence of synthetic data contamination. These findings provide important empirical support for optimizing dynamic data-mixing strategies in GAI self-training.

摘要： 本研究旨在探究生成式人工智能（Generative Artificial Intelligence, GAI）在自训练循环中的模型退化风险，重点聚焦内容同质化与人机文本差异两大核心现象。研究选取两种结构具有代表性的生成模型，构建自训练迭代实验框架，以人类数据在训练集中的占比α为核心超参数，在α不同取值下并结合线性递减、指数衰减两类典型动态策略开展对照实验，通过多维度性能指标系统评估生成内容的质量、多样性及与人类文本的差异程度。结果显示，GAI在自训练过程中性能呈持续下降趋势，生成内容多样性显著弱化，人机文本差异逐步扩大；线性递减策略可有效延缓信息熵下降、维持内容多样性，但后期易受模型生成数据污染的累积影响；指数衰减策略虽初期性能波动较明显，但其长期稳定性更优。此外，轻量级单向语言模型（GPT2）在自训练中更易陷入噪声放大的恶性循环，而具备更强全局建模能力的双向编码器模型（BART）在面对生成数据污染时，展现出更优异的鲁棒性。本研究为优化GAI自训练的动态数据配比策略提供了重要实证支撑。