RAG Collapse: LLM Responses Collapse When Retrieved Documents Are Self-Authored
Gregory Druck, Ethan Smith
cs.CL, cs.IR
2026-08-23
In 1,528 simulations, 79.6% collapse after retrieving self-authored articles. One such page often suffices, cited several times more than original sources.
Model collapse is the training-loop story: retrain on your own outputs and the distribution sharpens until it no longer looks like the original data (Shumailov et al., 2024). This paper asks the retrieval version. AI-written pages are already as common as human ones, and search-enabled LLMs read those pages into context. Graphite calls the result RAG collapse: answers stop tracking the spread of opinion online, and ten samples converge on one list.
They also measured the fuel. Among pages ChatGPT actually cited, GPTZero labeled 38.9% AI or mixed in January 2026 and 42.7% in June. Entity prompts sit higher (41.1% to 45.8%). The training loop is old news. The retrieval loop is already turning.
The base model is never retrained. Only the RAG context changes. Each question starts from real citations collected from ChatGPT or Google AI Overviews. Each round writes ten answers, expands some of them into article-shaped pages, and feeds those pages back. Three schedules:
Caps are 10 / 20 / 30 rounds. Prompts are entity comparisons (“best Twitch streamers right now”) or editorial how-to questions. There are 1,019 unique questions. Temperature stays at the API default of 1.0. A control with no self-authored pages shows sampling noise only, not collapse.
Collapse is exact entity-set identity across the ten answers for entity questions, and a 100% same-answer rate for editorial questions (GPT-5.2 judging 10 random pairs per round).
Of 1,528 simulations, 1,216 (79.6%) end collapsed. Replace All and Replace One finish almost tied: 88.12% of GPT-5.2 Chat entity questions in both, 91.23% vs 94.74% on editorial. Search is the most realistic design and still hits 77.23% and 75.44%. A larger Search run with GPT-5.2 on 742 entity and 102 editorial questions ends at 73.85% and 83.33%.
Collapse is fast. After one self-authored page (about 10–20% of the pool), Replace One has already collapsed 22.8% of entity questions, against 28.7% for Replace All. One page does most of the damage of full replacement. Pairwise paraphrase rates go from under 30% in round 1 to over 90% later. Entity visibility flips from a spread to all-or-nothing 0/100, and names that started at 100% visibility can still vanish.
Gemini 3 Pro and Claude Sonnet 4.5 collapse under Replace One as well. Gemini is slower; most confidence intervals overlap.
The driver is self-authorship, not “AI style.” In Replace One round 2, 39.7% of original pages are already AI-generated, yet only about 3% of entity questions are collapsed at the start. Citation rates: 38.9% self-authored, 9.4% AI originals, 7.4% human originals. Restrict to pages scored 5/5 on directness, organization, and relevance (75 self-authored vs 118 originals) and the gap is 38.2% vs 13.3%, about 2.9×. A regression with eight quality scores still attributes about +26 percentage points to self-authorship; “original and AI-generated” is not significant.
| Setting | Collapsed at start | Collapsed at end |
| Replace One / GPT-5.2 Chat / entity | 2.97% | 88.12% |
| Search / GPT-5.2 Chat / entity | 1.98% | 77.23% |
| Replace One / Claude 4.5 / entity | 1.67% | 91.67% |
| Search / GPT-5.2 / 742 entity | 7.01% | 73.85% |
Retrieval does not automatically mirror the web. Models prefer their own prose and the answers they would have given, so a small share of self-authored pages can rewrite who gets named. AEO/GEO teams will care which entities survive. RAG product teams should treat old model answers in the corpus as a diversity risk. The paper does not ship a mitigation; it makes the failure mode repeatable.
Seed citations come from commercial systems in January 2026, 38.9% already AI-labeled, with no way to mark which were self-authored by the same model. Any such pages in the seed would shrink the contrast, so the authors read the results as conservative. They never iterate on AI pages written by a different model.
These are controlled simulations, not an audit of ChatGPT in production. The paper is explicit: this does not prove RAG collapse is already happening, and live retrieval stacks may already damp it. Coverage skews to information-seeking and entity comparison. Factual questions with one right answer, and open-ended writing, are untested. Cost forced one run per question; nine repeats of the Twitch prompt show different survivors, so point estimates carry variance. In Search, larger chunks collapse faster, so retriever hyperparameters move the speed.