Before an AI chatbot ever talks to you, it reads. It reads a giant pile of text: websites, books, news articles, forum posts. That pile is called training data.
Every answer the AI gives is built from that pile. People sometimes imagine the AI looking things up or thinking hard about your question. Really, the training data is driving the bus.
The pile is huge, but it is lopsided. Popular topics appear millions of times. Rare topics barely appear at all. Picture the training data as a library: some shelves are stuffed, and some shelves hold one dusty pamphlet.
Meet Priya. She asks the AI a few questions. Tap each question and check the shelf before you read the answer.
When the shelf is empty, the AI does not stop. It reaches for the closest shelves it can find: other villages, other festivals, other big old trees. Every piece sounds real because every piece came from somewhere real. The finished answer is still made up.
This is a second cause of AI hallucination. The first lesson showed a fact falling out of a long chat. This time the fact was never in there to begin with. The AI never read about Meethapani, so it stitched together a "typical village" and handed it to Priya with a straight face.
There is no card catalogue the AI can check. It cannot feel a gap in its own knowledge. It answers a question about a village it never read about in the same warm, confident tone it uses for Paris.
Before you trust an answer, ask yourself how popular the topic is. World capitals sit on a packed shelf. Family history, small towns, rare hobbies, and local services sit on thin ones. Thin shelf means double-check.
A festival name, an exact date, a famous tree. When an answer about a rare topic arrives full of specific details, treat those details as decoration until you can verify them somewhere else.
Ask the AI about your own neighbourhood, your own job, or a hobby you know inside out. Watching it guess about things you know teaches you how it guesses about things you do not.
For a small town, a family question, or a local service, go past the chatbot. Try an official website, a library, or a person who was actually there. The AI is a starting point, and rare topics deserve a second stop.
An empty shelf makes the AI guess. The next lesson looks at a stranger problem: a shelf packed with books that were wrong before the AI ever read them. Continue to The wrong books.