AI models sometimes make things up — confidently stating 'facts' that are completely wrong. This is called hallucination. It happens because the model doesn't store facts like a database. It has compressed billions of pages into statistical patterns — what words tend to follow other words. When you ask a question, it generates the most plausible-sounding continuation, whether or not it's actually true.

This is the core problem: the model has no internal "confidence meter" connected to factual accuracy. Every token it generates is chosen for plausibility, not truth — so fabricated details can sound just as authoritative as real ones. Sounding right and being right are two different things.

Understanding where knowledge comes from helps you spot these errors. It lives in two places: parameters (facts baked into the model during training — like a textbook you've memorized) and the context window (information you provide in the current conversation — like an open notebook on your desk).

Modern systems mitigate hallucinations with web search, tool calling, and retrieval-augmented generation (RAG) — letting the model look things up instead of relying on memory alone. But they can't be fully eliminated. The model still generates its response one token at a time, choosing what sounds most likely — and sometimes "most likely" just isn't true.