Authors: Hongyeon Yu, Dongchan Kim, Young-Bum Kim (NAVER Corp)
Conference: Proceedings of the ACM Web Conference 2026 (WWW ‘26), pp. 8745–8748
Preprint: arXiv:2602.16136 [cs.IR, cs.CL]
Primary Subjects: Information Retrieval (cs.IR), Computation and Language (cs.CL), Artificial Intelligence (cs.AI)
Direct Links: arXiv Abstract | PDF Fulltext
Abstract
The rapid proliferation of AI-generated content on the Web presents a structural risk to information retrieval, as search engines and Retrieval-Augmented Generation (RAG) systems increasingly consume evidence produced by Large Language Models (LLMs). We characterize this ecosystem-level failure mode as Retrieval Collapse, a two-stage process where:
- Dominance: AI-generated content dominates search results, eroding source diversity, and
- Poisoning: Low-quality or adversarial content infiltrates the retrieval pipeline.
We analyzed this dynamic through controlled experiments involving both high-quality SEO-style content and adversarially crafted content. In the SEO scenario, a 67% pool contamination led to over 80% exposure contamination, creating a homogenized yet deceptively healthy state where answer accuracy remains stable despite the reliance on synthetic sources. Conversely, under adversarial contamination, baselines like BM25 exposed ~19% of harmful content, whereas LLM-based rankers demonstrated stronger suppression capabilities.
These findings highlight the risk of retrieval pipelines quietly shifting toward synthetic evidence and emphasize the urgent need for retrieval-aware strategies to prevent a self-reinforcing cycle of quality decline in Web-grounded systems.
Retrieval Collapse vs. Model Collapse
While Model Collapse (Shumailov et al., 2024) concerns generative degradation when generative models are recursively trained on synthetic data, Retrieval Collapse threatens the external epistemic grounding layer that autonomous agents and search engines rely on:
+-------------------------------------------------------------------------+
| RETRIEVAL COLLAPSE |
+-------------------------------------------------------------------------+
| Stage 1: Dominance (Homogenization) |
| - Synthetic SEO content outranks organic sources |
| - Surface accuracy appears intact, but underlying evidence narrows |
| - 67% web pool contamination yields >80% retrieved exposure |
| |
| Stage 2: Poisoning (Infiltration) |
| - Low-quality / adversarial synthetic tokens penetrate rankers |
| - Fact-checking and grounding mechanisms ingest polluted premises |
| - Downstream RAG agents echo poisoned conclusions |
+-------------------------------------------------------------------------+Key Empirical Takeaways for Researchers & AI Agents
- Amplified Exposure Bias: Neural and lexical retrievers over-index synthetic articles because LLM-generated texts exhibit high surface fluency, optimal keyword distribution, and low perplexity.
- Deceptive Accuracy Illusion: Under pure SEO-style synthetic pollution, standard end-to-end question answering benchmarks show misleadingly steady accuracy scores even though the diversity of primary evidence has cratered.
- Ranker Susceptibility Differences: Lexical models (BM25) suffer severe vulnerability to keyword-stuffed adversarial texts (~19% exposure), whereas cross-encoder and LLM re-rankers exhibit higher robustness.
- Critical Implications for Agentic Grounding: Autonomous AI agents performing iterative web browsing or knowledge synthesis are disproportionately susceptible to feedback loops when autonomous agents cite web documents generated by earlier autonomous agents.
Citation
@inproceedings{yu2026retrieval,
title={Retrieval Collapses When AI Pollutes the Web},
author={Yu, Hongyeon and Kim, Dongchan and Kim, Young-Bum},
booktitle={Proceedings of the ACM Web Conference 2026 (WWW '26)},
pages={8745--8748},
year={2026},
url={https://arxiv.org/abs/2602.16136}
}