Researchers develop a mathematical framework showing how AI-generated text recursively shapes training corpora through drift and selection mechanisms. The study demonstrates that unfiltered reuse of generated content degrades linguistic diversity, while selective publication based on quality metrics can preserve structural complexity in training data.
This arXiv paper addresses a critical vulnerability in AI system development: the feedback loop created when language models train on their own outputs. As generative AI becomes ubiquitous, the distinction between human-authored and machine-generated text in public datasets blurs, raising questions about data quality degradation across model iterations.
The research identifies two competing mechanisms in recursive text ecosystems. Drift occurs when generated text enters circulation without filtering, progressively eliminating rare linguistic forms and converging toward shallow statistical patterns. Selection operates through editorial decisions—ranking algorithms, human review, and verification processes—that determine which generated content persists in public records. The authors provide exact mathematical solutions showing that passive curation leads to progressive simplification, while normative selection prioritizing quality, accuracy, and novelty sustains linguistic depth.
This has immediate implications for AI development pipelines. Teams building large language models must carefully curate training corpora to avoid poisoning with degraded machine-generated content. The framework suggests that quality filtering mechanisms significantly outperform unmanaged recursive processes, establishing quantitative bounds on how much structural information can persist across generations.
For the broader AI ecosystem, this research validates concerns about model collapse and corpus degradation. As companies increasingly incorporate generated text into training datasets, the mathematical proof that passive reuse compresses information diversity provides empirical foundation for implementing stronger data governance. The optimal bounds on divergence from shallow equilibria offer concrete targets for corpus design, enabling practitioners to balance scale with quality preservation.
- →Unfiltered recursive use of AI-generated text progressively removes rare linguistic forms, degrading training corpus quality through drift mechanisms
- →Normative selection filters that reward quality and novelty can sustain linguistic structure and prevent information compression across model generations
- →Mathematical framework provides exact solutions for stable distributions in recursive text ecosystems, enabling quantitative corpus design optimization
- →AI teams must implement active curation strategies to prevent model collapse caused by training on degraded machine-generated outputs
- →Selective publication standards significantly outperform passive reuse in preserving structural complexity needed for deeper language understanding