Hallucination — when the bot makes stuff up

A hallucination is when a large language model says something confidently that isn't true. It sounds real. It isn't. The model isn't lying — it doesn't know it's lying — it's just generating the next plausible word without checking the facts against the real world.

Think of a cousin who always has a story about the bodega down the street. He swears the owner is his tío, he swears the plantains are the best in town, he swears he worked there for three years. He means well. He's a good guy. But half the time he's mixing up stories from different weekends. That's a hallucination.

The model has no separate memory module that it queries to verify facts. It just predicts words based on patterns it saw during training. If a fact happened to appear in the training data, the model can sometimes reproduce it. If it didn't, the model will invent something that looks like a fact because the sentence structure looks like a fact. The model has no way to distinguish the two.

This is why you shouldn't paste a hallucination into a work document without checking it. The model will happily cite a case that doesn't exist, a statute that was never enacted, a phone number that belongs to someone else. The only defense is to verify — or to use a model with a retrieval system that actually looks things up before it writes.

Ask the model for its sources, then check them yourself.