Hallucination: Why AI Confidently Lies

What is it?

A hallucination is a false statement from an AI that still sounds confident and convincing. For example, the AI might invent a book, a law, or a function in a library that doesn't actually exist — and phrase it as plain fact.

Why does it happen?

An LLM predicts the most likely next word, one at a time (see LLM). It has no built-in "truth checkbox". If it doesn't actually know the real answer, it still fills the gap with plausible-sounding text — because sounding plausible is exactly what it was trained to do.

When does it happen most?

• With very specific facts (names, numbers, quotes)
• With questions about things after the training cutoff
• With niche topics that had little training data
• When asked about things that don't exist ("tell me about function X in library Y")

What helps against it?

• Give the AI real sources or documents as context instead of letting it guess
• Always double-check important facts, especially numbers, quotes, API names
• Ask the AI to state uncertainty instead of guessing