Hallucination
A generative model's output that is fluent, confident, and wrong.
Last updated
In AI parlance, a hallucination is a generative model’s output that is fluent, confident, and wrong. Because LLMs are trained to predict plausible next tokens, they will happily produce plausible-sounding citations that do not exist, dates that are off by a century, and quotations nobody ever wrote.
Why it matters for your research. Hallucinations are the single biggest risk in using generative AI for humanities research. A hallucinated footnote can poison a citation chain for years. The defensive posture is to treat every generative claim as unsupported until verified against a primary source.
In Archēglyph. Architecturally impossible for the text the UI shows. Every passage, quotation, and neighbour is an extracted span from a real document in your bundle, annotated with the engine that produced it. If a model was not used, none is recorded. See Why Archēglyph cannot hallucinate.
Not to be confused with. OCR errors are not hallucinations — OCR reads actual ink, and its mistakes trace back to a real page region.