AI systems are citing statistics that no one can actually trace back to an original source

Statistics become detached from their origins all the time. When numbers travel through citations, rewrites, and secondhand sources without anyone checking the trail back to raw data, they can develop lives of their own, spreading widely while their actual foundation grows increasingly unclear. This pattern has existed long before AI, but AI systems have made it newly visible and consequential. How citations create vulnerability Academic and journalistic writing depends on attribution. A researcher publishes a finding with methodology and raw numbers. Someone else cites that researcher. A third party cites the second person. Each step introduces opportunity for simplification, rounding, reframing, or honest error. In fields where primary sources are difficult to access or studies are conducted in languages readers don't know, the citation chain becomes harder to verify. The farther a number travels from its source, the less likely anyone will trace it backward. The problem predates language models Universities have documented citation errors for decades. Historians have found that famous quotes are often misattributed or fabricated, spreading through books and speeches for centuries. In medical literature, researchers have discovered that incorrect data from one influential paper get repeated in dozens of subsequent papers, all citing the wrong figure as if it were established fact. Meta analyses and literature reviews sometimes inadvertently canonize errors simply by being cited more often than the original work. Why AI amplifies the pattern Language models trained on large text collections don't distinguish between a carefully verified statistic and a widely repeated error. They have no mechanism to independently verify numbers. When a model encounters the same figure across multiple documents, it treats repetition as a signal of reliability. The model can generate plausible sounding prose around any number it has seen, making invented or garbled statistics sound credible and authoritative. Unlike a human writer who might feel uncertain and check a source, an AI system produces confident looking output regardless of whether the foundation exists. The traceability problem Digital archives have made some original research easier to find, yet most statistics cited in business articles, policy briefs, and newsletters are cited from secondary sources, not primaries. A reporter reads a claim in another article, not in the original study. A consultant includes a statistic from a report without accessing the underlying data. Each intermediary adds credibility through their own authority or publication venue, even if they never verified the number themselves. When AI systems are trained partly on this ecosystem of secondhand citations, they inherit an already compromised landscape. What this means for information consumers The existence of zombie statistics doesn't mean all widely cited numbers are false. It means that confidence in a number a