A false claim rarely appears fully formed inside an AI assistant's answer. It usually starts somewhere quieter — a forum post, a translated press release, an unverified aggregator — and travels through several layers of repetition before a model repeats it as settled fact. Understanding that journey is the first step to interrupting it.
What "narrative laundering" means
The term borrows from financial laundering for a reason: a claim with no verifiable origin passes through enough intermediate sources that, by the time it reaches a mainstream outlet or a model's training and retrieval corpus, its origin is no longer visible. What started as an unsourced assertion arrives looking like an established fact, because it has accumulated the surface markers of legitimacy — multiple mentions, consistent phrasing, an apparent trail of citations — without ever passing through a single verification step.
This is not a new phenomenon; rumours have always spread this way. What has changed is the machinery downstream. A claim no longer needs to convince a human editor to become treated as fact — it only needs to appear often enough, in similar enough wording, across enough indexed pages, for a retrieval system to treat it as corroborated.
For a brand, the fastest-moving version of this pattern is not always driven by malice at all: a stale news aggregator that never updates a headline, or a translation service that garbles a quarterly figure, can start the same chain without anyone intending deception. Malicious laundering and accidental laundering look identical once they reach a model's answer — which is why correction has to work at the level of the answer, not at the level of intent.
The five-stage journey from rumour to "fact"
Every laundered claim we have traced follows a recognisable pattern, whether the origin is a disgruntled former employee, a competitor, or a simple misreading of a public filing.
1. Origin
An unverified claim appears somewhere with low editorial friction: a forum, a social post, an anonymous blog, a press release from an entity with no track record.
2. Amplification
Low-quality aggregator sites and content farms — some run by scrapers, some by SEO operators chasing the same search terms — republish the claim, often word for word, to fill thin pages.
3. Pseudo-citation
A secondary source cites the aggregator rather than the origin, lending the claim a first layer of apparent sourcing. Each further mention treats the previous mention as its source, and the original disappears from view.
4. Indexation
Search engines and retrieval systems index the now-repeated claim across dozens of pages. Volume, not veracity, becomes the dominant signal a ranking or retrieval layer can measure.
5. Generation
A model, asked a related question, retrieves the now-widespread version of the claim and states it with the same confidence it would use for an audited fact — because, mechanically, both look like well-supported content.
Why a model can't easily tell the difference
Systems built on retrieval-plus-generation pipelines have no built-in hierarchy of trust between a regulatory filing and a content farm, unless the underlying index or ranking layer has been explicitly built to weigh source quality. What they have instead is a strong prior toward whatever pattern recurs most consistently in the material retrieved. A claim repeated across ten low-quality sources can statistically outweigh a single high-quality source that contradicts it, especially when that single source is crawled less often.
This is a mechanical property of how these systems work, not a deliberate editorial choice — which is exactly why it is exploitable, and why it rarely resolves itself over time without active correction. The effect compounds on niche topics, where only a handful of sources discuss the subject at all and there is little competing signal to dilute a repeated claim.
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.
Repetition does the work persuasion used to do
Cognitive research has long documented an "illusory truth" effect in humans: statements feel more true the more often they are encountered, independent of accuracy. Retrieval-based generation reproduces a version of the same dynamic at infrastructure scale. A model does not need to "believe" a claim in any human sense — it only needs the claim to be statistically dominant in what it retrieves. Actors who understand this can manufacture apparent consensus cheaply, by seeding a claim across many low-cost surfaces rather than convincing a single credible outlet.
This produces a counterintuitive failure mode: coverage written specifically to debunk a claim can still reinforce the pattern a retrieval system measures, because the debunking article still contains the claim itself, indexed alongside every other mention of it. The answer is not to avoid mentioning a false claim when correcting it — that makes the correction unreadable — but to make sure the correcting source is specific, dated and structured well enough for a retrieval layer to weigh it differently from the pile of repetition it is meant to displace.
Anatomy of a laundered claim
The pattern is consistent enough that it can be diagnosed after the fact. A claim about a leadership change, a product recall, or a lawsuit typically shows a burst of near-identical mentions across low-authority domains within a short window, followed by a slower trickle onto higher-authority sites that cite the burst without adding new sourcing. By the time it reaches a model's answer, the chain of custody has been erased — the assistant states it as a fact drawn from "public reporting", with no single reporter behind it.
A concrete illustration, stripped of any identifying detail: a mid-sized firm's founder steps back from day-to-day operations for personal reasons. A single low-traffic blog frames the move as a forced departure tied to unspecified "internal issues." Three content-farm domains republish the framing within a week, each adding a slightly different invented detail. By the following month, a model asked about the founder's status states, as fact, that they "left following internal issues" — a claim with no primary source, built entirely from the accumulated weight of its own copies.
Warning signs to watch for
- No traceable primary source — every citation points to another aggregator, never to a filing, a transcript, or a dated original statement.
- Suspiciously uniform phrasing — near-identical sentences across supposedly independent outlets, a signature of copy-paste amplification.
- A sudden appearance with no prior history — a claim about a long-established entity with no footprint before a specific date.
- Confidence disproportionate to sourcing — an assistant states a specific figure or event with no hedge, yet no single authoritative document supports it.
- The claim dissolves at the primary source — checking the entity's own filings, transcripts or statements directly turns up nothing resembling the specific detail in question.
Breaking the chain: correction over deletion
Because a laundered claim lives across many indexed copies, no single takedown request resolves it. The protocol we recommend treats correction as a competing-signal problem: publish an authoritative, dated, well-sourced counter-statement where models and their retrieval layers are likely to find it, then track — over weeks, not days — whether it displaces the false version across the models being monitored. Our audits typically find that a claim which took months to accumulate enough repetition to be treated as fact takes a comparable, sustained effort to correct; there is rarely a single fix.
Persistence tracking matters as much as the initial correction. A model that stops repeating an error in one session can resurface it in another, particularly if the underlying laundered content is still indexed and gets re-crawled. In practice this means treating the correction the way the original claim spread — not as a single article, but as a sourced reference deliberately placed where retrieval systems are likely to encounter it, then checked against the same panel of models on a fixed cadence until the false version stops appearing. See our public methodology for how we structure that tracking.