An earlier answer returns as a reference
An AI answer becomes an article. Later, the same model retrieves that article to answer the question again. The page looks like an outside source, but its substance came from the model’s earlier answer. Adding a reference has not necessarily added an independent reason to believe it.
That loop is the subject of Gregory Druck and Ethan Smith’s paper, “RAG Collapse.” Their definition of “self-authored” is narrow: a page generated by the same model from its own answer. This is not a study of AI-written content in general.
The important change is that an earlier selection becomes a later input. An answer leaves things out; it is not a complete record of everything its sources could support. When that answer returns as reference material, the next response can reinforce the selection rather than revisit the full range of possibilities.
The answers became less varied
The researchers simulated this loop by repeatedly turning answers into articles and making those articles available as references for later answers. They did not retrain the models. What changed was the material supplied when the models answered.
Across 1,528 controlled simulations covering 1,019 information-seeking questions and three model families, 79.6% ended in what the authors call “collapse.” For questions comparing named entities, that meant every response in a round mentioned the same set of entities—not necessarily in the same order or wording. For open-ended advice questions, it meant all sampled response pairs were judged by a model to be paraphrases.
This measures agreement across responses, not factual correctness. A consistent answer can be right. A varied set of answers can contain mistakes. The finding is that repeated retrieval of answer-derived pages can narrow the range of responses; it does not establish that the resulting answers are false.
Consistency and corroboration are different questions
For a marketer reading an AI recommendation, the useful distinction is between seeing the same answer repeatedly and having independent support for it. Several pages descended from an earlier answer may repeat one selection, rather than provide several separate grounds for trusting it. Their agreement can be real without being independent.
The study does not establish that commercial AI products are already undergoing this collapse. Its simulations controlled the available references rather than reproducing a complete live search system, and commercial safeguards are unknown.
So a stable recommendation is not, by itself, evidence of either a failure or a well-supported consensus. The question that separates repetition from corroboration is where the supporting claims came from—not simply how many pages repeat them.