Hallucination

A fluent, confident answer from a language model that is false: a figure that appears in no filing, a quotation nobody said, a citation to a report that does not exist. It happens because the model generates plausible text, and a plausible number is not the same as the right one.

Ask a model for a company's revenue in a given year and it may produce a figure that is off by 10% or 20%, with the right units and a confident tone, and nothing in the answer marks it as invented. Errors cluster where the training data was thin: small companies, recent years, specific line items and anything that needs arithmetic across several documents.

Defend against it as you'd against any secondary source. Give the model the primary document and ask it to quote the line it relied on, check every number that reaches your analysis against the original, and treat a citation you haven't opened as unverified.

Fluency isn't evidence.