The Mirage of the Independent Citation

Imagine a language model answers a query, and that output is subsequently packaged and published as a standalone web article. Later, when faced with the same query, the same model pulls that newly minted article into its retrieval context to formulate a fresh response. On the surface, the model appears to be grounding its claims in an external source. Beneath the hood, however, it is merely looking in a mirror. Adding another URL to the citation list has not actually introduced an independent line of evidence.

This specific feedback loop is the focus of a paper by Gregory Druck and Ethan Smith titled “RAG Collapse.” Rather than making sweeping claims about all machine-generated text on the internet, the authors isolate a very specific phenomenon: what happens when a model ingests content synthesized directly from its own prior outputs.

The core problem here is one of information bottlenecking. Whenever a model generates an answer, it inevitably filters, compresses, and omits details that were present in the broader pool of original source material. When that distilled answer is recycled back into the retrieval pipeline as a primary reference, the system can reinforce its earlier editorial choices rather than re-evaluate the full landscape of possibilities.

Watching Diversity Evaporate

To test how this dynamic plays out over time, Druck and Smith built a simulation that repeatedly converted model answers into reference articles and fed them back into the retrieval pool for subsequent rounds of questioning. Crucially, the underlying model weights were never altered or retrained; the only moving variable was the reference material supplied at inference time.

The results were stark. Across 1,528 controlled simulations spanning 1,019 information-seeking prompts and three distinct model families, 79.6% of the runs terminated in what the researchers define as “collapse.”

How collapse looked depended on the task:

  • Entity Comparisons: For prompts asking to compare named entities, collapse meant that every single generated response in a given round converged on the exact same roster of entities—even if the prose or ordering varied.
  • Open-Ended Advice: For broader advisory prompts, collapse meant that every sampled pair of responses was judged by an evaluating model to be a semantic paraphrase of one another.

It is vital to separate this loss of variance from a loss of factual accuracy. The study measures convergence, not truth. A locked-in, unvarying answer might be entirely accurate, just as a highly diverse set of outputs can be riddled with hallucinations. What the experiment demonstrates is that feeding answer-derived pages back into the retrieval context systematically narrows the aperture of future responses—not that those narrowed responses are inherently wrong.

Repetition Is Not Corroboration

For marketers and analysts auditing how AI systems recommend brands or products, this dynamic highlights a critical blind spot: mistaking repetition for consensus. If five different web pages all echo the same recommendation, it is tempting to view that convergence as robust, multi-source validation. Yet if those pages were all downstream derivatives of a single earlier AI response, their agreement is an artifact of lineage, not independent verification.

Of course, a controlled experiment is not a live production environment. This research does not prove that commercial answer engines are currently spiraling into retrieval collapse. The authors deliberately constrained the reference pools in their simulations rather than replicating the chaotic, constantly updating reality of a live web index, and we do not know what proprietary deduplication or source-weighting guardrails commercial platforms deploy behind the scenes.

Still, the takeaway remains sharp. Seeing a model return the same stable recommendation time after time tells you nothing, by itself, about whether the system is failing or succeeding. To know whether you are looking at genuine corroboration or a closed loop of self-citation, counting the number of supporting pages isn't enough—you have to trace the genealogy of the claims inside them.

  • Source Note: Research by Gregory Druck and Ethan Smith (Published 2026-08-22).
  • Methodological Context: The simulations centered on information-seeking queries, heavily weighting comparisons between named entities. The replacement experiments assumed that newly generated articles successfully entered the model's context window, while the search simulations operated within a bounded document pool without modeling other content continuing to be published alongside the generated articles.
  • Disclosure: Both researchers are affiliated with Graphite Growth, and a portion of the prompts evaluated in the study were sourced from Graphite's prompt-tracking platform.