Citation hallucination is a specific, checkable form of AI hallucination: a reference that has the familiar shape of a scholarly or factual source but fails when you try to resolve it. It comes in two forms. Fabrication, where no work matching the cited title and authors exists at all, and author-identity corruption, where a real work is credited to a substantially wrong author list. Ordinary drift, like a changed year or venue, is not counted.
It matters because references are the currency of trust. LLM tools make it trivial to generate polished bibliographies and confident-looking sources that were never verified, and a 2026 Microsoft audit found such references surviving peer review into published papers. The defense is the same mechanism AI search uses to decide what to cite: resolve each claim against multiple independent authorities and distrust what can't be corroborated. For anyone publishing AI-assisted content, the lesson is direct. Verify before you publish, because unverifiable claims are exactly what retrieval systems are built to discount.