In the run-up to the 2003 Iraq invasion, one of the load-bearing claims about mobile bioweapons labs traced back to a single defector, code-named Curveball, debriefed by German intelligence. American analysts never spoke to him directly. What they had instead were multiple reports — filtered through liaison channels, re-transcribed, summarized by different desks — that all said roughly the same thing. Several documents, several distribution paths, several analysts nodding at each other's summaries. It read like corroboration. The postmortems later established that it was one man's testimony, laundered through enough intermediaries to look like a crowd.
This isn't a story about lying, mostly. Curveball's information degraded through ordinary channel noise as much as through fabrication. The interesting failure is structural, and it recurs anywhere information gets copied before it gets counted: a claim enters a system once, gets forwarded, paraphrased, filed under different headers, cross-referenced by people who never touch the original — and by the time someone tallies "how many independent things point to this," they're counting distribution, not origin.
Journalism has a cleaner-cut version of the same mechanism, sometimes called citogenesis: an unsourced claim lands on a low-traffic page, a journalist on deadline lifts it because it reads authoritatively, a second outlet cites the first outlet rather than digging for a primary source, a third cites the second. Within a week the claim has five outlets behind it and feels five times as solid, and every single one of them is downstream of the same unverified sentence. Nobody along the chain did anything unreasonable by the local standard of their job. The aggregate is still worthless as corroboration, because corroboration was never what happened — repetition was.
Law has the opposite instinct built into doctrine on purpose. The two-witness rule and its descendants exist specifically to block this failure: a second account only counts if it could plausibly have arrived independently. A co-conspirator's testimony needs outside corroboration precisely because conspirators talk to each other before they talk to investigators — their stories agreeing is what you'd expect whether or not the story is true. The rule isn't cynicism about liars. It's an acknowledgment that shared origin produces agreement all by itself, with no truth-tracking involved, and a legal system that counted agreeing stories as strength regardless of ancestry would reward whoever coordinates best, not whoever's most accurate.
Science has been re-deriving this the hard way for about a decade, under the name of the replication crisis. A meta-analysis that pools twelve studies looks more solid than any one study — unless several of those twelve pull from the same underlying dataset, the same lab's earlier cohort, the same set of research assistants trained by the same person to make the same judgment calls on ambiguous cases. Meta-analysis, done carelessly, is a machine for converting one measurement into an apparent twelve by counting publications instead of measurements. The statistics get more confident. Nothing about the world got more confirmed.
What all four cases share is the same move at the point of failure: counting instances of a claim rather than tracing the claim's ancestry. Instances are cheap to produce — copying, forwarding, re-filing, and citing all manufacture instances without manufacturing any new contact with the world. Independence is not a property of how many places a claim shows up; it's a property of whether the paths that produced those appearances actually diverged before they agreed. Two witnesses who talked to each other first are not two witnesses. Five outlets citing one another are not five sources. Twelve studies sharing a dataset are not twelve data points. Three intelligence reports laundering one defector through three channels are not three streams of evidence.
The correction isn't "trust nothing that appears more than once" — that overcorrects into a different failure, refusing real corroboration because it looks structurally identical to fake corroboration from the outside. The correction is asking a specific, answerable question before the count matters at all: if I traced each of these back, do they touch a common point before they touch the world? If yes, the number of times the claim appears is a fact about the distribution network, not about the claim. If the paths genuinely diverge — different methods, different witnesses with no contact, different datasets collected under different conditions — then agreement is actually informative, for the ordinary reason that independent processes don't usually agree by accident.
The check is cheap wherever the lineage is available: ask what each signal actually touched, not how many signals there are. It's only expensive in the cases where it matters most — where the paths are hard to trace, the intermediaries are motivated to smooth over gaps, and the appearance of a crowd is exactly what nobody along the way had any reason to correct.