Abstract
The practice of Retrieval-Augmented Generation (RAG), which integrates Large Language Models (LLMs) with retrieval systems, has become increasingly prevalent. However, the repercussions of LLM-derived content infiltrating the web and influencing the retrieval-generation feedback loop are largely uncharted territories. In this study, we construct and iteratively run a simulation pipeline to deeply investigate the short-term and long-term effects of LLM text on RAG systems. Taking the trending Open Domain Question Answering (ODQA) task as a point of entry, our findings reveal a potential digital "Spiral of Silence" effect, with LLM-generated text consistently outperforming human-authored content in search rankings, thereby diminishing the presence and impact of human contributions online. This trend risks creating an imbalanced information ecosystem, where the unchecked proliferation of erroneous LLM-generated content may result in the marginalization of accurate information. We urge the academic community to take heed of this potential issue, ensuring a diverse and authentic digital information landscape.
Introduction
As Large Language Models (LLMs) become ubiquitous, the internet is being flooded with synthetic text. Most modern AI applications use Retrieval-Augmented Generation (RAG), which fetches documents from the web to help an LLM answer questions. This creates a dangerous feedback loop: LLMs generate content, which is then indexed by search engines, and subsequently retrieved by RAG systems to generate new text. This paper explores whether this cycle creates a "Spiral of Silence"—a term borrowed from communication science—where human-authored, diverse, and potentially more accurate information is pushed to the periphery by a "louder," more SEO-friendly mass of AI-generated text.
Methodology: The Iterative Simulation Pipeline
To model the long-term evolution of RAG systems, we constructed an iterative simulation pipeline. The process begins with a human-authored baseline (e.g., Wikipedia). We then introduce "zero-shot" LLM-generated answers into the search index. In each cycle, a RAG system answers questions based on the current index. These new LLM-generated answers are cleaned of "AI-like" phrases and added back into the index. This process is repeated for 10 cycles to observe environmental changes.
We tested various retrieval methods (BM25, Contriever, BGE) and re-rankers (MonoT5, UPR) across four standard Open Domain Question Answering (ODQA) benchmarks: Natural Questions (NQ), WebQuestions (WebQ), TriviaQA, and PopQA.
Key Insights
AI Text as "SEO on Steroids"
In the short term, LLM-generated text actually improves retrieval accuracy. LLMs generate text that is highly relevant and structured specifically to answer queries, making it easier for retrieval algorithms to find compared to "messy" human prose. However, this initial boost masks long-term degradation.
The Vanishing Human and Homogenization
After 10 iterations, the percentage of human-generated text in the top 50 search results plummeted to below 10% across all datasets. Human voices are effectively silenced by the sheer volume and "retrievability" of AI text. Furthermore, using the Self-BLEU metric, we found that the diversity of information in search results decreased significantly over time, as the system converged on a narrow set of viewpoints and phrasing.
The Misinformation Trap
When LLM-generated misinformation was deliberately injected, RAG systems struggled to self-correct. The "Spiral of Silence" caused misinformation to be prioritized and repeated, even when correct human information remained in the index.
Experimental Results
Our results show a significant Retrieval Decay. While initial iterations saw a boost, long-term retrieval effectiveness (Acc@5) dropped by an average of 21.4% for NQ and 19.4% for PopQA by the tenth iteration.
Interestingly, the "Exact Match" (EM) scores for final answers remained relatively stable. This suggests that LLMs are "lazy"—as long as at least one correct document remains in the top results, they can answer correctly, even as the overall quality of the search results degrades.
Conclusion
The proliferation of LLM-generated content risks creating an imbalanced information ecosystem. While source filtering (removing AI text) can preserve human content, it often hurts system performance because the models have become reliant on the highly relevant synthetic documents. We urge the community to develop retrieval models that balance relevance with information diversity and authenticity, and to create more robust AI detection methods that do not compromise system utility.
Summary updated