Hallucination

A hallucination is confident model output with nothing real behind it. Because a model produces text by next-token prediction, it will produce a plausible answer whether or not it has a basis for one; there is no built-in 'I don't know' unless the training or the prompt made room for it. In coding, that means a function that looks like it belongs to the library but doesn't, a flag that was never in the CLI, or a summary of a file that misstates what the file says.

It helps to separate two kinds. The first is missing knowledge: the model reaches for a fact it never learned, or learned before the knowledge cutoff, and fills the gap. The second is drift from the context it was given: the file is right there in the window, but the model paraphrases it wrong, usually late in a long session once attention degradation has set in. The first is fixed by loading a primary source; the second by shortening the session.

The working defence is verification the model does not control. Have the agent run the code, not describe it; make a passing test the definition of done; ask for line-numbered citations when it claims a file says something. And treat the confident tone as noise: fluency and correctness are produced by the same mechanism, so one tells you nothing about the other.