Source-linked AI summary
RAG Collapse: LLM Responses Collapse When Retrieved Documents Are Self-Authored
Gregory Druck, Ethan Smith
TL;DR
The paper asks what happens when AI systems retrieve content they authored amid increasingly prevalent AI-generated online text. Through simulations of self-authored-reference retrieval, it finds that RAG responses collapse in most simulations and that self-authored content is disproportionately cited. The authors conclude that this threatens the diversity of responses, while noting that the simulations do not definitively establish that RAG collapse is already occurring in practice.
Problem
AI-generated content is prevalent online, but the effects of search-enabled LLMs retrieving their own generated references remain an open question.
Method
The paper simulates search-enabled LLMs repeatedly retrieving references generated from their own responses, including recursively replacing original references with self-authored ones.
Results
79.6% (1,216/1,528 simulations) ended in collapse, and models disproportionately cited self-authored references even after controlling for reference quality.
Takeaways & Limitations
Self-authored references can cause responses to converge on a single perspective rather than reflect the diversity of human thought.
Takeaways & Limitations
The simulations suggest RAG collapse but do not definitively prove that it is already occurring in practice.
Abstract
from arXiv · showhide
LLM responses are based on the internet (via training or RAG), and AI is now used to generate a significant amount of content online (Paredes et al., 2026), creating the potential for a self-reinforcing feedback loop. Prior work has shown that when LLMs are recursively trained on their own output, they experience model collapse (Shumailov et al., 2024): responses become less diverse, and eventually no longer resemble the original training data. In this paper, we show that a similar collapse occurs if LLM-based AI systems retrieve references they authored using a search tool. We call this RAG collapse. We conduct extensive experiments with three types of simulations of AI systems retrieving references they generated, using three model families, and 1,019 information-seeking prompts, totaling 1,528 simulations and over one million LLM API calls, and find that 79.6% (1,216/1,528) of simulations end in collapse. Surprisingly, even a single self-authored reference can trigger collapse because the LLM disproportionately cites its own content. This self-bias persists even after controlling for reference quality.
1 Introduction
AI-generated content is prevalent online, and search-enabled LLMs can retrieve self-authored references, creating a potential feedback loop. The paper finds that this retrieval pattern can collapse response diversity and disproportionately favor the model’s own content.
- Motivation: AI-generated articles already constitute a substantial share of online references, creating the potential for a self-reinforcing feedback loop.The paper reports that AI-generated and human-written articles are now published in comparable numbers.
- Findings: 79.6% (1,216/1,528 simulations) ended in collapse when search-enabled LLMs retrieved self-authored references.Collapse means the system produces essentially the same response every time.
- Findings: Self-authored references can trigger collapse almost immediately, even when only one such reference is present.The results suggest these references have a disproportionate influence on responses.
- Findings: Self-authored references are cited more often than original references, including references that are themselves AI-generated.This pattern remains after controlling for reference quality and is described as self-bias toward the model’s own writing style.
- Example: Across five rounds of recursively replacing references, streamer visibility collapsed to either 0% or 100%, so every answer mentioned the same streamers.Some initially frequent streamers disappeared, while others became ubiquitous.
4 Simulations
The paper simulates AI systems retrieving content they previously generated, using three variants that differ in how self-authored references enter or replace retrieved context.
- Scope and Assumptions: The study focuses on RAG retrieval rather than retraining the base model and limits its scope to information-seeking question prompts.The variants therefore test collapse from retrieved self-authored content under stated simulation assumptions.
- Simulation Framework: The simulations generate responses from internet context, convert selected responses into online articles, and then replace or add those articles as references.This setup models publication of self-authored articles followed by their retrieval during RAG.
- Replace All: Replace All substitutes every reference with a self-authored reference generated from a previous-round response.The actual simulations use ten responses and at least five references, despite simplified figures showing three of each.
- Replace One: Replace One replaces one original reference with a randomly selected self-authored reference in each round.Unlike Replace All, original references are replaced more slowly, but the design still assumes self-authored references are retrieved.
- Search: The Search simulation adds self-authored references one at a time to a pool and retrieves the top-k most relevant chunks alongside original references.Self-authored and original references compete for retrieval from the pool.
5 Simulation Implementation Details
The implementation evaluates repeated response generation under controlled contexts, using commercial LLM APIs, sampled prompts, reference processing, and metrics designed to track convergence and self-authored retrieval.
- API-Based Simulation: The simulations use OpenAI, Gemini, and Anthropic APIs to control the context provided to each model.Each simulation requires hundreds of API calls across response generation, article generation, evaluation, entity extraction, and relevant-content extraction.
- Reference Processing: Original references are collected from ChatGPT, Google AI Overviews, and Claude-related search outputs, then scraped and reduced to main content.Replace All and Replace One use question-relevant excerpts, while Search uses chunks from full main content.
- Round Structure: Each round generates ten responses, with maximum lengths of 10 rounds for Replace All, 20 for Replace One, and 30 for Search.Replace All fully replaces original references after round 1, Replace One by round 10, while Search may retain originals depending on retrieval.
- Control: As a control, simulations without self-authored references show only variation expected from sampling LLM responses.This isolates self-authored retrieval as the condition associated with collapse in the simulations.
6 Datasets
The datasets comprise information-seeking prompts assembled from search-related sources and manually written questions, with separate entity-comparison and editorial categories and references collected in 2026.
- Entity Questions: Entity questions compare named entities such as products, brands, businesses, and locations.They are collected from prompt tracking, a prior categorized-keywords dataset, and manually written questions.
- Editorial Questions: Editorial questions cover general information-seeking prompts that are neither entity comparisons nor factual questions with one correct answer.An example is “How can I improve my personal branding?”
- Dataset Construction: The study uses 1,019 unique questions after combining overlapping but nonidentical ChatGPT and Google AI Overview question sets.Questions were filtered based on whether ChatGPT triggered web search or Google produced an AI Overview.
- Reference Collection: The original reference datasets were collected in January 2026, with additional ChatGPT references collected in early June 2026.AI-generated references were identified using GPTZero labels of AI or Mixed.
- Reference Composition: 38.9% of ChatGPT references in January 2026 and 42.7% in June 2026 were classified as AI-generated.The percentage increased from January to June for both entity and editorial prompts, with more AI-generated references among entity prompts.
7 Evaluation Metrics
The evaluation tracks response similarity, lexical and entity diversity, self-authored retrieval, citation bias, and paraphrase convergence across simulation rounds.
- Response Metrics: Semantic similarity measures average cosine similarity between pairs of responses generated in the same round.Higher values indicate that responses are more similar to one another.
- Response Metrics: Unique words counts distinct words across ten responses, while same-answer percentage measures the proportion of response pairs that are paraphrases.GPT-5.2 evaluates paraphrases using a sample of 10 out of 45 possible response pairs.
- Retrieval and Citation Metrics: Self-authored citation percentage measures the share of citations pointing to self-authored references, and self-authored retrievals measures the share of retrieved Search chunks that are self-authored.Citation-over-expected compares observed self-authored citation rates with rates expected from their context share.
- Entity Question Metrics: Entity metrics include unique entities and entity-ranking similarity, which uses Kendall’s tau on the order of first entity mentions.A ranking similarity of 1.0 means two responses mention entities in the same order.
- Aggregation and Entity Processing: The reported metric values are means computed for each round across all questions in the dataset.For entity questions, entities are extracted, canonicalized across rounds, and supplemented with exact-match searches.
8 Defining Collapse
The paper defines collapse as responses converging in content or entities, then finds declining diversity, increasing similarity, and frequent collapse across simulation designs and models. Self-authored references are disproportionately cited, with collapse often beginning after only a small amount of self-authored content enters the reference pool.
- Defining collapse: Collapse means responses essentially convey the same content, even when they are not word-for-word identical.For entity questions, collapse requires identical entities across responses; for editorial questions, it requires a 100% same-answer percentage.
- Cross-experiment patterns: Most simulations end in collapse, with unique words and entities decreasing while semantic and entity-ranking similarity increase over rounds.These patterns are reported consistently across the experiments.
- Replace All: 88.1% of entity questions and 91.2% of editorial questions collapse in Replace All after ten rounds.Replace All replaces every reference with a self-authored reference after each round.
- Replace One: Replace One collapses 88.1% of entity questions and 94.7% of editorial questions, despite introducing only one self-authored reference per round.Collapse begins almost immediately, rather than being slower and less frequent than Replace All.
- Search: The Search simulation collapses 77.2% of entity questions and 75.4% of editorial questions, even though it retrieves documents from a growing content pool.Its rates are lower than the replacement simulations but remain high.
- Early collapse: After one self-authored reference enters the pool, 22.8% of Replace One entity questions collapse by round 2 versus 28.7% for Replace All.The corresponding editorial rates are 50.9% versus 64.9%, indicating that a small share of self-authored references produces most of the collapse observed with full replacement.
10 Why Do Self-Authored References Have Disproportionate Influence?
Self-authored references exert disproportionate influence on citation and retrieval, and this advantage persists after accounting for reference quality. The evidence suggests self-bias rather than AI generation alone, although the mechanism remains unresolved.
- Self-authored references trigger collapse even when they comprise only a small fraction of the reference set.The effect appears with even a single self-authored reference, or 10–20% of total references.
- Self-authored references are cited more often than both AI-generated and human-written original references.AI-generated originals have a 9.4% citation rate versus 7.4% for human-written originals, while self-authored references reach 38.9%.
- Self-authored references score higher on most judged quality dimensions, but the types are tied in accuracy.The largest gaps favor self-authored references in direct answer, relevance, thoroughness, organization, and overall quality.
- 38.2% versus 13.3%: among references scored 5/5 on direct answer, organization, and relevance, self-authored references retain a 2.9× citation advantage.The quality-controlled subset contains 75 self-authored and 118 original references.
- After controlling for all eight quality dimensions, self-authorship adds about 26 percentage points to citation rate, while AI-generated originality adds none.The self-authorship coefficient is 0.26 with a 95% bootstrap CI of [+0.23, +0.29], whereas AI-generated originality is −0.01 and not significant.
- The results suggest self-bias, but the experiments cannot distinguish preference for matching reasoning from preference for matching writing style.The authors also note that another uncontrolled attribute could explain the citation gap.
11 What Types of Prompts Collapse?
Collapse is more resistant for prompts whose initial responses are longer and contain more unique entities. The observed correlations are statistically significant but modest, indicating that other factors also contribute.
- Longer initial responses and more unique entities are the factors most correlated with eventual collapse in Search simulations.The analysis uses generation statistics from the initial round with only original references.
- Prompts with longer, more varied initial responses are more resistant to collapse.The correlations are statistically significant but modest.
12 Which Entities Drop Out?
Entities with low initial visibility are most likely to disappear during simulation, while visibility and early response position both predict retention. Even highly visible entities sometimes drop out.
- Entity dropout decreases as initial visibility increases, but even 100%-visibility entities drop out about 7–10% of the time.Low-visibility entities that initially appear in few responses drop out very frequently.
- Better-positioned entities tend to drop out less frequently, and high-visibility entities tend to appear near the beginning of responses.Visibility and position are correlated in the observed responses.
13 Does RAG Mirror the Input Entity Distribution?
RAG responses do not simply reproduce the distribution of entities in their references: response mentions can exceed reference mentions and favor less-mentioned entities. The experiments and proposed mitigations also have important scope boundaries.
- Response entity frequencies do not necessarily mirror reference frequencies.The analysis compares responses mentioning each entity with references mentioning it across questions with ten references.
- Across models, more responses generally mention an entity than references do, with saturation around five input references.Gemini 3 Pro includes low-mention-count entities in responses more often.
- The authors did not test prompts explicitly designed to prevent collapse.They report generally observing collapse across the prompts tried and identify prevention prompts as future work.
- The reported experiments use the default Search chunk size, although larger chunks increase collapse rate and probability.Smaller chunks lead to less collapse, supporting the hypothesis that complete self-authored answers exert disproportionate influence.
- Filtering AI-generated references, diversifying results, or encouraging response diversity are proposed mitigations, but their effectiveness remains unknown.Commercial systems may already use such mechanisms, but the authors do not know precisely how those systems work.
- Diversifying search results helps only when diverse, relevant information exists online, since low diversity can also reflect genuine consensus.Distinguishing consensus from insufficient diversity may be challenging.
- AI detection may miss human-written articles whose substance comes from model answers, even though they could contribute to collapse.This differs from the paper’s self-authored-reference definition and is described as a possible knowledge-collapse pathway.
16 Limitations
The paper’s evidence is limited to simulations and information-seeking prompts, while several broader settings and practical questions remain unresolved. The authors also note that existing or future mitigations may affect whether RAG collapse occurs in practice.
- 38.9% of seed references are already AI-generated, and the simulations cannot identify which are self-authored.Any self-authored content in the seed pool would reduce the measured contrast, making the reported collapse and disproportionate-influence results conservative.
- The study does not test iterative addition of other types of AI-generated content, which could also lead to collapse.
- The simulations suggest RAG collapse but do not definitively establish that it is already occurring in practice.The authors note that mechanisms intended to mitigate collapse may already exist, although their effectiveness is unclear.
- The experiments focus on information-seeking prompts, especially comparisons of named entities, so collapse may behave differently on other prompt types.
- Future work will examine agentic search, mitigation effectiveness, cross-model retrieval of authored references, and additional prompt types.
A Complete Replace One Results
This section provides complete Replace One results for the ChatGPT and AI Overview datasets, including questions outside their intersection, with separate figures for three model families.
- The complete results include questions that are not in the intersection of the ChatGPT and AI Overview datasets.
- Figure 29 presents Complete Replace One results for GPT-5.2 Chat.
- Figure 30 presents Complete Replace One results for Gemini 3 Pro.
- Figure 31 presents Complete Replace One results for Claude Sonnet 4.5.