A new problem is surfacing: researchers found that fake statistics can slip into…

The internet is increasingly shaped by machine learning systems that absorb vast quantities of text, then produce summaries, answers, and citations based on patterns in that training data. As these systems grow more powerful and more widely trusted, a structural vulnerability has become visible: bad information can enter the training data, get reinforced by algorithmic patterns, and exit as authoritative-sounding claims. Understanding how this happens requires looking at how large language models work, why citation matters, and what happens when scale meets unverified information. Training data is a mirror and an amplifier Large language models are trained on enormous collections of text scraped from the internet, books, articles, and other sources. They do not fact-check or verify what they absorb. A researcher's blog post, a Wikipedia entry, a news article, a social media thread, and a made-up statistic posted in a forum all end up in the same training bath. The models learn statistical patterns from all of it equally. If a particular phrase, number, or claim appears frequently enough in the training set, the model will learn to reproduce it. The more often something appears in the source material, the more likely the model is to treat it as a normal pattern and regenerate it. This creates an odd feedback loop. Fake statistics do not need to be true to be common. If a fabricated number gets repeated across dozens of websites, academic papers that cite it uncritically, blog posts that copy it, and social media posts that spread it, the model sees a pattern. The pattern looks like consensus. The model learns to output it. Why citations feel like verification When a language model outputs text with a citation, it creates a powerful psychological effect. A citation signals rigor, evidence, and accountability. Readers tend to trust cited claims more than uncited ones, because citations are supposed to point to verifiable sources. This intuition is sound when citations are made by careful humans. But language models do not reason about whether a source actually supports a claim. They learn to generate citations as patterns in text. A model might produce a convincing-looking citation to a source that either does not exist, or does exist but does not actually say what the model attributed to it. The citation is a hallucination, but it looks legitimate. Scale creates surface plausibility Before large language models, a fake statistic had limited reach. It lived in a blog post, maybe got quoted in a few places, and faded. The number of people who would encounter it and believe it was bounded by the visibility of those sources. Large language models changed the equation. A statistic that appears in a few dozen places online can now be absorbed into a model used by millions of people. Each time someone asks the model a question, the model might reproduce that statistic, with a plausible-sounding citation. The fake number now reaches audiences of a s