AI disinformation

AI-generated content and AI citation: the recursive risk Deep Narrative Intelligence addresses

October 24, 2026 6 min read
Kamel Malek Director of SetRecord.ai

A new risk has emerged in the AI information ecosystem, and it is fundamentally recursive: AI generates content, that content is published on the web, another AI retrieves and cites it as a source, and the cycle repeats. Each iteration can compound the errors of the previous one, creating a feedback loop where distortions amplify rather than decay. Deep Narrative Intelligence is the discipline uniquely positioned to detect and break this loop, because it is the only discipline that measures the model layer where the recursion occurs.

The recursive citation loop

The loop works as follows. An AI model generates a response to a query — say, a factual claim about a company that is subtly wrong. That response is published somewhere on the web: a blog post that used ChatGPT to draft an article, a Q&A site where a Gemini answer was copied as a response, a wiki page edited with Claude-generated content. A second AI model, in a later training cycle or during web-search retrieval, ingests that published content as a source. The second model now treats the first model's output as authoritative, because it appears on a web page. When the second model is asked the same query, it reproduces the distortion — possibly with additional drift — and publishes its output, which a third model may later ingest. The loop is recursive because each model's output becomes another model's input.

How errors compound in the loop

The compounding effect is the dangerous part. In each iteration, the distortion can gain specificity it did not have before. A model that initially said "Company X may have had revenue issues" might, after citing a source that was itself generated by a model that elaborated on that vague claim, produce "Company X had revenue issues in 2024, citing reports" — a more specific, more confident, and more false version of the original vague distortion. The original uncertainty ("may have had") has been laundered into certainty through the citation chain, because each model in the chain treats its source as authoritative without knowing that the source was itself model-generated. This is how a hallucination becomes a fact in the model layer: not through deliberate deception, but through recursive citation that strips away the uncertainty markers at each step.

Don't know what AI says about you today?

A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.

Request an audit

Why traditional tools miss it

Fact-checking misses the recursive loop because it operates on individual claims, not on the chain of citations that produced them. Media monitoring misses it because it tracks published content, not the model-generated content that was the original source. Platform moderation misses it because the content is not necessarily policy-violating — it is a plausible-looking claim on a legitimate-looking page. The recursive loop is invisible to every layer of the existing anti-disinformation stack because it is a model-to-model phenomenon, and no existing tool monitors the model layer as its primary object. Only DNI — which systematically interrogates models and traces provenance through the citation chain — can detect that a model's claim traces back to another model's output, and identify where the recursion entered the chain.

How DNI breaks the loop

Breaking the recursive loop requires intervention at two points. The first is source identification: tracing the model's claim back through the citation chain to identify where a model-generated source entered the loop. This requires provenance methods that can distinguish between human-authored and model-generated web content — using creation timestamps, content patterns, and source-history analysis. The second is authoritative replacement: publishing a human-authored, evidence-based source that contradicts the compounded distortion, structured so that the model's retrieval system prefers it over the model-generated source. The replacement must be precise enough to override the compounded specificity of the recursive claim. SetRecord is developing provenance methods specifically designed to detect recursive citation patterns, because this risk will only grow as more web content is model-generated and more models cite it.

The outlook

The recursive citation risk is not a future hypothetical — it is happening now, and it will accelerate. As more content on the web is AI-generated (some estimates suggest over 50% of new web content will be AI-assisted by 2027), the probability that any given model is citing another model's output grows. The compounding effect means that distortions that would naturally decay in a human-authored ecosystem may instead amplify in a model-authored one. DNI is the early-warning system for this risk: it detects the loop where it occurs — in the model layer — and intervenes at the source where it can be broken. Organisations that do not monitor the recursive citation risk will discover, eventually, that the distortions about them in the model layer are not sourced from human errors that can be corrected by publishing accurate human content, but from machine errors that require a fundamentally different correction strategy.

See what AI says about you

A SetRecord audit checks what supported AI models say against your verifiable facts — within five business days.