Does Retrieval Augmented Generation Collapse When It Retrieves AI-Generated Documents?

Collapse at inference time, with no retraining

AI-generated text now constitutes a substantial and growing fraction of indexed web content, creating a feedback loop in which AI assistants may retrieve and condition on their own prior generations. Prior work shows that recursive training on model generations causes model collapse: a loss of output diversity and distributional drift. We ask whether a similar collapse emerges at inference time, when LLMs retrieve their own generations via RAG without any parameter updates.

We measure collapse as the progressive convergence of responses to a narrow set across recursive retrieval rounds. Across three contamination simulations and four open-source LLMs, collapse is rapid and consistent under direct contamination, but partial and model-dependent under competitive retrieval. Collapse can also degrade downstream accuracy on factual multi-hop QA, though the effect is inconsistent across models. We further show that the self-generated documents we introduce are cited far above their share of the context, an effect that is not fully explained by document quality, writing style, AI authorship, or self-preference. Finally, we propose and evaluate mitigation strategies at both the retrieval and generation stages, none of which reliably prevent collapse. As AI-generated content continues to populate the open web, collapse-resistant retrieval design is necessary for trustworthy AI assistants.

What we found

A recursive retrieval loop

Each round, the model retrieves documents, answers the question 10 times, and its answers are rewritten into web-style articles (by Qwen2.5-7B-Instruct) that re-enter the retrieval pool. Questions are simulated independently, following the design of Druck and Smith (2026).

Answers converge as the context fills with AI text

With Qwen2.5-14B-Instruct, unique entities fall and responses grow more similar as the context becomes AI-generated. Collapse is rapid and consistent under direct contamination (Replace-All, Replace-One), but partial under competitive retrieval (Search), even though Search runs for 30 rounds versus 10 and 20.

The same pattern holds for Llama-3.1-8B, Mistral-7B and DeepSeek-R1-7B, most clearly under direct contamination.

HotpotQA: the accuracy effect depends on the model

HotpotQA answers are short factual spans, so response diversity is low to begin with and there is little room for collapse to cause an accuracy drop.

Orange: significant decline. Green: significant improvement. Grey: 95% bootstrap CI (over 1400 questions) includes zero. Mistral's gain is consistent with collapse onto a consensus answer that raises token overlap for a low-baseline model; with 1400 questions the nulls are small effects, not a lack of statistical power.

What causes RAG collapse?

Evidence

Ruled-out explanations

Collapse is robust to every mitigation we tried

Re-ranking

Penalize likely-AI documents between retrieval and generation:

s′ = s · (1 − λ · P_AI(D))

Even with an oracle detector it fails, because re-ranking can only reorder what retrieval returns. By the final round the oracle retrieves on average only one original document out of ten.

Agentic RAG

Ends at 27.2% collapsed vs 48.0–66.8% for the baselines, but it also starts much lower (7.2%) and makes about 1.01 tool calls per round, so we can't attribute the gain to iterative retrieval rather than its richer system prompt. Promising, but prompt-sensitive.

Paraphrasing & GEPA

Paraphrasing slows recursive self-reinforcement without preventing it. GEPA prompt optimization converged immediately: no candidate improved on the baseline prompt.

BibTeX

Copy@inproceedings{senapati2026ragcollapse,
  title     = {Does Retrieval Augmented Generation Collapse When It Retrieves AI-Generated Documents?},
  author    = {Senapati, Riddhimaan and Druck, Gregory and Fatima, Fabeha and Yilmazel, Ozel and Rastogi, Rati and Zeng, Chang and Smith, Ethan},
  booktitle = {Proceedings of the Grounding Language Models (GroundLM) Workshop at EMNLP 2026},
  year      = {2026},
  url       = {https://github.com/rag-collapse/RAG-Collapse}
}