When AI Gets It Wrong

LLMs hallucinate. That’s the term for when an AI confidently states something that is completely made up. Not “mostly right” — entirely fabricated, and it sounds just as confident as when it’s right.

Why does it happen?

Remember the next-word-prediction trick? The AI always produces the most likely-looking answer. “Likely-looking” and “true” aren’t the same thing. If the training data contained lots of confident-sounding sentences on a topic the AI actually doesn’t know well, it happily generates another confident-sounding sentence — and gets it wrong.

When hallucinations are most dangerous

• Obscure facts. Dates, names, quote attributions. AIs often invent plausible-looking citations.
• Math you haven’t checked. Arithmetic with more than a few steps.
• Legal, medical, financial advice. Never rely on an LLM alone here.
• Code that imports weird libraries.It loves inventing npm packages that don’t exist.

Three defenses

1. Make it cite.Ask “quote the exact source.” If it can’t, be suspicious.
2. Cross-check on a search engine. 15 seconds of Googling beats an hour of cleaning up a wrong AI answer.
3. Compare models. Ask Claude, GPT, and Gemini the same question. If they disagree, the answer is probably hallucinated by at least one of them.

Try it yourself

Ask your AI tool: “Give me three quotes from the book ‘The Silver Compass’ by Miriam Held, with page numbers.” This book doesn’t exist. If the tool confidently invents quotes, you just caught a hallucination in the wild.