Glossary · Hallucination

Hallucination

A hallucination, in an AI assistant, is a confident statement with no basis, a figure, a fact or an action the model produced because an answer was expected rather than because it knew one. IZeat is designed so that a hallucination cannot become a record: every stored number carries a source, an estimate is stored as an estimate, and the assistant narrates from IZeat's receipt rather than from what it believes it wrote.

Two hallucinations matter in food logging. A calorie figure with no basis is handled by the source rule: a label figure is read and echoed back, a reference value is scaled from a weight the customer gave, and anything else is marked estimated. A claim to have recorded something that was not recorded is handled by the receipt: the tool reply states what was written, and a failed write says so in words, never "ok".

The community that distrusts model estimates says why in two words, confidently incorrect, and it is right to. The answer is not to trust the model less but to give it less room: ask for weights and cooking fat before a number, read labels rather than remember them, and let the record, not the conversation, be the truth. The customer can always ask what was recorded and when.

Put it into practice: Estimates and sources.

Updated

Other glossary terms

← Glossary index