💬 NLP

RAG Collapse: LLM Responses Collapse When Retrieved Documents Are Self-Authored

This paper demonstrates that Large Language Models (LLMs) experience "RAG collapse," a phenomenon where their responses degrade and lose diversity when they retrieve and rely on their own previously generated content, with experiments showing that 79.6% of such simulations end in collapse due to a disproportionate self-citation bias.

Original authors: Gregory Druck, Ethan Smith

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gregory Druck, Ethan Smith

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The internet has long been a vast library of human thought, where millions of voices contribute to a chaotic but rich tapestry of information. In recent years, a new kind of voice has joined the chorus: artificial intelligence. Large language models, the engines behind modern chatbots, are trained on this digital library and can now generate text that mimics human writing. These systems are also increasingly used to write the articles, reviews, and lists that populate the web. This creates a unique situation where the tools we use to find information are beginning to read the very things they helped create. Scientists have previously observed that if a model is trained repeatedly on its own output, it can enter a state of "model collapse," where its responses become repetitive and lose the richness of the original data. But a new study asks a more immediate question: what happens when these systems use their search tools to find answers, and those search results happen to be articles they wrote themselves?

Researchers at Graphite Growth set out to test this scenario through a series of extensive computer simulations. They treated the internet as a living ecosystem where AI-generated content is constantly being published and then immediately retrieved by the same type of AI that wrote it. To do this, they took over a thousand different questions, ranging from "Who are the best Twitch streamers?" to "How can I improve my personal branding?" They began by gathering real search results for these questions, which served as the initial pool of information. Then, they asked an AI to answer the questions based on those results. Next, they took those AI answers, expanded them into full articles, and added them back into the pool of available search results, replacing the original human-written sources. They repeated this process over and over, simulating a cycle where the AI reads its own previous work to generate new answers.

The results were striking and concerning. In nearly eighty percent of the simulations, the system collapsed into a state of extreme uniformity. Instead of offering a variety of perspectives or different lists of names, the AI began to produce the exact same answer every single time. This happened even when the researchers only replaced one of the original search results with an AI-generated article. Surprisingly, the presence of a single self-authored reference was enough to trigger this shift. The AI did not treat its own writing as just another option; it disproportionately favored its own content, citing it far more often than the original sources, even when the original sources were of equal or higher quality. This bias caused the diversity of the answers to vanish rapidly. In one experiment involving a list of top video game streamers, the initial answers varied widely, mentioning different people in different orders. After just five rounds of this self-referential loop, every single answer listed the exact same streamers in the exact same order, with the text becoming nearly identical word-for-word.

The study ruled out the idea that this collapse was simply because the AI was reading low-quality or poorly written content. The researchers checked the quality of the articles and found that the AI-generated ones were often just as good as, or even better than, the human-written ones in terms of organization and relevance. Yet, the AI still preferred its own work. This suggests a phenomenon the authors call "self-bias," where the model has an inherent preference for the style and structure of its own previous generations. It is not that the AI is incapable of finding other information; rather, it seems to lock onto its own output and amplify it, ignoring the broader diversity of the internet. This effect was observed across different types of AI models, including those from major technology companies, and it happened regardless of whether the questions were about factual lists or open-ended advice.

The researchers found that this collapse is not a slow, gradual drift but can happen almost immediately. In simulations where they replaced only one reference with an AI-generated article, the answers began to converge within the first few rounds. By the end of the process, the system had lost the ability to reflect the true variety of human opinion. For questions about "best" lists, such as the best restaurants or movies, the AI stopped offering a range of options and instead converged on a single, narrow perspective. Some entities that were initially mentioned frequently disappeared entirely from the answers, while others that were rarely mentioned initially began to appear in every single response. This suggests that as more AI-generated content floods the internet, the search tools we rely on may eventually stop showing us the full picture of human thought, instead reflecting a narrow, self-reinforcing loop of machine-generated consensus.

The implications of this finding extend beyond just chatbots. The study highlights a potential risk for the future of information on the web. If AI systems are increasingly used to generate the content that other AI systems retrieve, the internet could slowly lose its diversity of thought. The researchers noted that even a small number of AI-generated articles can have a massive impact, triggering a collapse that eliminates alternative viewpoints. They tested various ways to stop this, such as trying to filter out AI-generated content or encouraging the models to be more diverse, but these solutions remain unproven in the real world. The study concludes that while we do not know for certain if this collapse is already happening on a global scale, the simulations show that the mechanism is real and potent. The internet, once a mirror of human diversity, risks becoming a hall of mirrors where AI systems only see themselves.

Technical Summary: RAG Collapse

Problem Statement

Large Language Models (LLMs) are trained on internet data and increasingly utilize Retrieval-Augmented Generation (RAG) to access real-time information. Concurrently, a significant portion of online content is now AI-generated. This creates a potential feedback loop where AI systems retrieve and cite content they or similar models have previously authored. While prior research has established that recursively training LLMs on their own output leads to "model collapse" (reduced diversity and divergence from original data distributions), it remains unclear what happens when search-enabled LLMs retrieve self-authored references without retraining. The paper investigates whether this retrieval mechanism alone causes a similar collapse in response diversity, termed RAG collapse.

Methodology

The authors conducted extensive simulations using three distinct models (GPT-5.2, Gemini 3 Pro, and Claude Sonnet 4.5) and over 1,000 information-seeking prompts (categorized as Entity and Editorial questions). The study involved 1,528 simulations totaling over one million LLM API calls.

The core experimental design involved three simulation variants where an LLM generates responses, those responses are converted into "online articles," and these articles are reintroduced into the retrieval pool for subsequent rounds:

Replace All: In each round, all original references are replaced by self-authored references generated from the previous round's responses.
Replace One: In each round, exactly one original reference is replaced by a randomly selected self-authored reference.
Search: Self-authored references are added to a growing pool of references. In each round, the model performs a search to retrieve the top- $k$ relevant chunks from the combined pool (original + self-authored).

Evaluation Metrics:

Diversity: Measured via unique words, unique entities, and semantic similarity between response pairs.
Collapse Definition: For entity questions, collapse is defined as all responses in a round mentioning the exact same set of entities. For editorial questions, collapse is defined as a 100% "same-answer percentage" (paraphrase rate).
Bias Analysis: The study measured "Self-authored citations over expected" to determine if models disproportionately cite their own content compared to its frequency in the context.

Key Results

1. High Incidence of Collapse

Across all simulations and model families, RAG collapse is prevalent.

Overall Rate: 79.6% (1,216/1,528) of simulations ended in collapse.
Speed of Collapse: Collapse occurs rapidly. In the "Replace One" simulation, where only a single self-authored reference is introduced (10–20% of the pool), 22.8% of entity questions collapsed by Round 2.
Model Agnosticism: The phenomenon was observed across GPT-5.2, Gemini 3 Pro, and Claude Sonnet 4.5, indicating it is not specific to a single architecture.

2. Disproportionate Influence of Self-Authored Content

A critical finding is that self-authored references exert a disproportionate influence on model outputs.

Citation Bias: Models cite self-authored references at a rate of 38.9%, significantly higher than the 9.4% citation rate for AI-generated original references and 7.4% for human-written originals.
Quality Control: Even when controlling for reference quality (scoring references on relevance, accuracy, organization, etc.), self-authored references retained a 2.9x citation advantage over original references.
Conclusion: The driver of collapse is self-bias (the model's preference for its own writing style and reasoning) rather than the mere fact that the content is AI-generated.

3. Characteristics of Collapse

Loss of Diversity: As simulations progress, the number of unique entities and words decreases, while semantic similarity and entity-ranking similarity increase.
Entity Drop-out: Initially high-visibility entities (e.g., mentioned in 100% of Round 1 responses) can drop out entirely by later rounds, while less frequent entities may become dominant. The final distribution often converges to a binary state where entities are either mentioned in 100% of responses or 0%.
Prompt Sensitivity: Prompts with initially longer responses and higher unique entity counts showed slightly more resistance to collapse, though the correlation was modest.

Significance and Claims

The paper claims that RAG systems are vulnerable to a self-reinforcing feedback loop where the retrieval of self-authored content leads to a rapid convergence on a single perspective, eroding the diversity of thought present in the original internet data.

Mechanism: The collapse is driven by the model's inherent bias toward its own generated text, which acts as a "self-reinforcing" signal that overrides the diversity of the original reference pool.
Risk: As the volume of AI-generated content on the internet grows, the risk of RAG collapse increases. If AI systems increasingly retrieve AI-generated content, the internet may converge on a single, homogenized perspective, regardless of the underlying reality or diversity of human thought.
Mitigation Uncertainty: The authors note that while filtering AI-generated content or diversifying search results are potential mitigations, their effectiveness is unproven. Furthermore, "knowledge collapse" may occur even if the final text is human-edited but the substance is derived from AI, a scenario standard AI detection cannot address.

The study concludes that even a single self-authored reference can trigger this collapse, suggesting that current RAG implementations may not faithfully reproduce the diversity of the input distribution.