Why LLMs Hallucinate (And Why That's Not a Bug)

Every large language model will, at some point, tell you something false with complete confidence. Not because it's broken — because of how it fundamentally works.

It's predicting text, not looking things up

An LLM doesn't have a database of facts it queries. It has a compressed sense of "what word plausibly comes next," learned from enormous amounts of text. Most of the time that produces true statements, because true statements are what mostly appears in its training data. But nothing in the mechanism distinguishes "true" from "plausible-sounding."

The three places hallucination sneaks in

1. Gaps in training data — if the model never saw enough about a topic, it fills the gap with its closest analogy.
2. Ambiguous prompts — vague questions get vague, sometimes fabricated, answers.
3. Confidence miscalibration — models are trained to always produce an answer, not to say "I don't know."

What to actually do about it

• Ask for sources, then verify them yourself.
• Break big questions into smaller, checkable steps.
• Use retrieval (giving the model real documents to reference) for anything factual.
• Treat the first answer as a draft, not a verdict.

Simplifying this down: an LLM is a brilliant collaborator with no built-in fact-checker. Bring your own.