What is a hallucination in the context of LLMs?
Options
- A fluent answer that is false or not supported by the sources (Correct option)
- The model refuses to answer a question
- A very slow response
- The model repeating the user's question
Answer
A fluent answer that is false or not supported by the sources
Why
An LLM produces the most likely continuation of the text. Nothing in that process checks the facts, so it can state something wrong in a confident, well-written way.
It is more likely when the question is about rare facts, recent events or things it was not given.
The usual defence is to give the model the source text and ask it to answer only from it.
Remember
Hallucination = confident but unsupported.