AI hallucination

Language models generate text by predicting the statistically most likely continuation – they do not check whether a statement is true. When a model lacks knowledge about a question, it fills the gap with plausible but invented content: fabricated citations, wrong figures, blended facts, or details that never existed. What makes this especially treacherous is the confident tone – hallucinated answers sound just as convincing as correct ones. Hallucinations are not a flaw of any single vendor but an inherent property of generative AI: they can be reduced, but never fully eliminated.

In everyday business, this creates concrete risks: a customer reply with incorrect product details, a contract summary citing a clause that isn't in the document, or an internal answer about a policy that never existed. Reusing such output unchecked risks bad decisions and lost customer trust. A practical rule of thumb: AI answers without a traceable source should always be reviewed by a knowledgeable person before they are used – the higher the stakes, the more thorough the review.

Chifty counters hallucinations with source grounding: answers in Chifty Chat and reply drafts in Chifty Office draw on approved content from the shared knowledge base and display source references, so every statement can be verified. The guiding principle is that the AI must not bypass permissions – it only answers from sources the asking person is allowed to see. If no reliable source exists, the system prefers returning nothing over inventing an answer. Email drafts are never sent automatically: a human reviews and sends every message.