Fake statistics cited in LLMs

Astronomers logged a blog post, published to the Hacker News community and picked up for discussion there on October 20, titled “Publish fake stats, get cited in LLMs.” The post, hosted at peec.ai, describes a tactic in which fabricated statistics are published online with the aim of having large language models absorb and repeat them as fact.

The piece argues that LLMs, when generating answers, frequently cite figures found on the open web without independently verifying their origin or accuracy. According to the account, this creates an opening for anyone willing to publish invented numbers, since a model may later surface those numbers in its own output, lending them an unearned appearance of authority.

The record does not specify which statistics were fabricated, how widely they were cited by models, or which LLMs were tested. No methodology, dataset, or named model is detailed in the material available to this desk. The post’s claims rest on the author’s own account of the experiment, as relayed through the Hacker News thread.

The observatory notes the discussion as reported, without independent verification of the underlying test. The broader implication raised in the community discussion concerns the reliability of AI-generated answers that draw on unvetted web sources, though the specific mechanics of the exploit remain outside the scope of what this record can confirm.