Why LLMs Hallucinate

My daughter asked ChatGPT how many legs a spider has. Correct answer. Then she asked it what her stuffed bunny’s name was. It gave her a name. Very confidently. Completely wrong.

“Daddy, why did it lie?” — “It didn’t lie. It guessed.” — “But it didn’t say it was guessing.”

And that’s the whole problem in one sentence.

Remember episode 1? An LLM works by guessing the next word, one word at a time, based on everything it has seen before. Here’s what nobody tells you: it never stops guessing. Not when it knows the answer, and not when it doesn’t. Both come out sounding exactly the same.

That’s a hallucination. Not a bug. Not lying. Just guessing when it should have said “I don’t know.”

The grown-up word is hallucination. Her word for it was “making stuff up with a straight face.” I think hers is better.

So what do you do about it? Same thing you’d teach a kid: check the answer before you trust it. Ask where it came from. That’s what the next episode is about.