noun · AI failure
Hallucination
also called confabulationmaking things up
A confident answer that is simply wrong. The AI sounds sure because sounding right is what it is built to do.
Model output that is fluent and confident but false or unsupported, such as a made-up citation, function name or fact.
The model generates the most plausible continuation, not a verified fact. When the right information is missing or weak, it still produces fluent text in the same style, with no built-in marker in the text that separates a guess from a known answer.
Real-life example
You ask for a library function that parses dates. The assistant gives you a clean example with a function name that looks perfect. You run it and the function does not exist. It did not lie, it wrote what a function in that library would plausibly be called.
AttributeError: module 'datelib' has no attribute 'parse_flexible_date'Where the analogy breaks
-
It is not lying. A lie needs intent. The model is not trying to mislead you, and it writes false and true statements in much the same tone.
-
It does not only happen on obscure topics. It can happen on familiar ones too, especially with exact details such as numbers, names, quotes and references.
-
Telling it to be accurate does not fix it. What helps is giving it real sources to work from and checking the output against them.