Why It Sometimes Makes Things Up
A guessing machine doesn't have a built-in fact-checker. When it doesn't actually know something, it still has to guess, confidently and fluently.
A very confident storyteller
Picture someone who has read an extraordinary number of books and can talk fluently about almost anything, but who, when asked something they genuinely don't know, doesn't say "I'm not sure." Instead, without quite meaning to, they fill the gap with something that sounds exactly like the kind of thing that would be true, delivered with the same fluent confidence as everything else they say.
That's a close picture of what people call hallucination in a language model: a confident, fluent, entirely plausible-sounding answer that happens to be wrong, or entirely invented.
Why this happens at all
Go back to the core idea from the first chapter: the model's job, at every single step, is to guess the most plausible next token given everything so far. It has no separate step where it checks "is this claim actually true" against some internal database of facts. It was never built with a true/false switch. It was built to continue text in a way that fits the patterns it learned, and a fluent, specific-sounding, wrong answer often fits those patterns just as smoothly as a fluent, specific-sounding, correct one.
This treats the model like it has intentions and knows it's being dishonest. It doesn't have a concept of "lying" any more than a very good guesser has a concept of cheating at a guessing game. It produced its best guess.
There was no dedicated fact-check step to catch it, because that step doesn't exist inside the core guessing mechanism. Confidence in tone and accuracy of content are two entirely separate things for a model that only ever learned to sound fluent.
Where it happens most
Hallucination isn't random noise scattered evenly across every answer. It shows up more in specific, predictable situations:
What actually helps
The fix isn't making the model try harder to be honest; it's giving it something better than its own guess to rely on. That's exactly what the tools and data covered in The Connection Layer, the free guide to MCP, are for: instead of guessing what a document says, the model can be given the actual document to read from. Instead of guessing today's numbers, it can be given a live lookup. A guessing machine that's handed the real answer doesn't need to guess.
Treat specific facts, numbers, and citations from an AI response the way you'd treat a claim from a stranger who sounds very sure of themselves: plausible-sounding confidence is not the same evidence as being correct. Verify anything that actually matters.