The Number That Wrote Itself

A statistic is loose in the industry. Several statistics, actually. They compound across the citation chain because the citation chain is a closed loop. The phenomenon has a name now: citation laundry.

For three weeks, a small experiment ran inside a single industry research feed. The feed pulled in daily intel digests from AI content vendors covering software delivery, AI coding tools, and merge-gate governance. Every figure that arrived attached to a specific vendor or attribution was logged. Every figure was then traced back, where possible, to its primary source. The exercise was meant to populate a stats library for an essay series and a product positioning document. It produced something else entirely.

Roughly one in three of the figures could not be traced back to a primary source. Roughly one in five was attributed to a real company that does not, in fact, ship the product the figure was describing. Two of the most-cited research papers in the feed did not exist. Their arXiv identifiers returned no matching results. The identifiers had been generated by something, and the something was confident.

The phenomenon

An AI content marketing shop produces a blog post on the state of AI software development. The post needs a striking statistic to anchor the lead. The AI is asked for one. The AI confidently produces a number, attributing it to a plausible-sounding source. The blog post ships. A second AI content shop, building its own piece on the same topic, reads the first post as part of its context. The second AI cites the statistic, sometimes reattributing it to a different plausible source. A third shop reads both posts. The figure now appears in the analytical universe as a settled fact, supported by multiple independent citations, none of which trace back to anything.

Nobody is fact-checking. Each shop in the chain assumes someone earlier already did the work. The chain is recursive, automated, and increasingly fast. The original statistic, if there ever was one, drowns in the citations of the statistics that wrote it.

Three weeks of evidence

The pattern is not hypothetical. Across three weeks of daily digests, the same figures appeared and reappeared with shifting attributions. A specific percentage stat about useEffect lifecycle errors appeared three times with three different source companies. A real and verified telemetry figure about incidents per pull request, drawn legitimately from a study of twenty-two thousand developers, was rewrapped in at least three subsequent digests with the attribution changed to a fake-sounding consulting firm that produced no such study. A real company in the secrets-management space was named three separate times as the source of merge-gate governance research it does not conduct.

An arXiv identifier circulated across four separate digests over six days with the attribution shifting between ResearchGate, OpenReview, the Mining Software Repositories conference, and a Medium article. The identifier was checked on arxiv.org. No matching paper exists. The real paper covering the underlying topic uses a different identifier and a different author list, neither of which appears in the digest stream. The fake identifier kept appearing because the digest stream is increasingly trained on its own outputs, and the fake identifier already has citations.

A second fake arXiv identifier appeared weeks later, accompanied by a fabricated technical specification (a "Masking Index" of 1.8, an "85% Corrector Capacity Threshold") that read like the contents of a real systems paper. The identifier was checked. No matching paper exists. The specification has no source. The mathematics was generated.

Across the three-week stretch, the case count of recycled-attribution and fake-citation samples reached twelve. Twelve documented instances inside one feed inside one industry vertical. The phenomenon is not anecdotal anymore.

Why this matters more than vendor noise

The industry has always tolerated marketing copy with shaky statistics. Vendors have always padded their claims, surveys have always been sponsored by the people whose products they evaluate, analyst reports have always had financial ties to the companies they cover. The phenomenon described here is different in one specific way: the citation chain has closed. The marketing copy is no longer being fact-checked by humans against primary sources. It is being generated by AI, fed as context to other AI, cited by that AI in turn, and circulated as analytical signal.

The closure of the loop is the new condition. In the older world, a fabricated number entered the discourse, got challenged by a reporter or an analyst or a competing vendor, and either died or was sourced. The challenge step is what kept the discourse honest. It was always sparse, but it was real, and a fabricated number had to survive contact with someone whose job was to ask where it came from.

That step has been removed at scale. Each AI in the chain treats the prior AI's output as evidence. None has access to the primary world the way a human reporter would. None has the institutional incentive to verify. The challenge step is gone, and the figures are compounding without the friction that used to kill the worst of them.

The principle: citation is not verification

The category-defining mistake is to treat a multi-source citation as a verification signal. A figure that appears in four different blog posts feels more credible than a figure that appears in one. A statistic with three different vendor attributions feels independent in a way that a statistic from a single vendor does not. The mind weighs replication as evidence, because for most of recorded history replication required independent observation by independent observers.

That heuristic does not hold inside the AI content chain. Inside the chain, replication is the default state. Repetition is cheap. A figure appearing in four blog posts costs the same as a figure appearing in one. The shape that used to indicate independent verification now indicates only that the chain has had time to copy the figure forward. Multiple citations of the same number in the AI content marketing universe is not evidence the number is true. It is evidence that the number is old enough to have been re-prompted into the next round of context.

The reader who treats a multi-cited statistic as credible because it is multi-cited is using a pre-AI heuristic in a post-AI environment. The heuristic is broken. The signal it carried is no longer present in the signal.

What this requires

For readers, the discipline is to chase every cited statistic to a primary source before using it. Not the source the AI summary attributes it to. The actual document, with a URL or DOI, written by a named human or institution, dated, with a methodology disclosed. If the primary source cannot be found in two minutes of searching, the statistic does not exist as evidence. It exists as text.

For writers, the discipline is to treat the absence of a primary source as a hard stop. If a figure passes through a draft without a verifiable origin, it does not enter the final document. The cost of writing without an unverifiable statistic is a slightly weaker sentence. The cost of publishing one is the slow corrosion of the writer's credibility every time a careful reader checks the source and finds nothing there.

For analysts and the institutions that pay them, the discipline is to recognize that the analytical input pile is now substantially contaminated. The intelligence feeds that summarize the industry for executives are increasingly fed by AI content shops citing each other. The cleaner the summary reads, the less it can be trusted as a representation of underlying reality. A serious analytical practice now has to budget time to verify what was previously taken as given.

For regulators, the direction is already visible in adjacent domains. Synthetic media has prompted requirements around provenance. Synthetic statistics will eventually require the same. The audit answer to "where did this figure come from" will stop being acceptable as "from our market intelligence vendor." It will have to terminate in a primary source with a verifiable origin.

The closing observation

The phenomenon described here is not, fundamentally, about AI being good or bad at writing. It is about what happens when a closed-loop generative system is used as if it were an open-loop investigative system. The chain treats its outputs as inputs. The inputs degrade across each pass. The degradation is invisible from inside the chain because the chain has no contact with the world the figures purport to describe.

The cure is not better AI. The cure is humans reintroduced at the verification step. A reporter checking a primary source. An analyst calling the company that supposedly conducted the study. A writer refusing to publish a statistic that does not terminate in a verifiable origin. None of these are scalable in the way the chain is scalable. That is the point. The thing that keeps the discourse honest is necessarily slower than the thing that erodes it. The asymmetry has to be paid for somewhere.

I could write an article stating 99 percent, and half of the industry would cite it before lunch.

Cross-link: this piece sits adjacent to the publication's existing arguments about what verification means under AI velocity. Companion in spirit to N° 017 (Phantom Coverage) on what dashboards measure when the measurement itself has detached from the thing measured, and N° 018 (The Question No One Signed) on what happens to accountability when the speed of decision exceeds the speed of human verification. The pattern is the same shape in different surfaces: an intermediary layer (a dashboard, a signature, a citation) loses its connection to the thing it was supposed to represent, and the loss is invisible from inside the system because the system no longer touches the ground.