Why do large language models sometimes 'hallucinate' information?

Large language models sometimes make up information because they're trying their best to guess what comes next.

Imagine you have a really smart friend who loves telling stories. They know almost everything, but when they get tired or confused, they might say something that doesn’t quite fit, like adding an extra character into the story or mixing up two different tales. That’s kind of what happens with large language models.

They look at lots of words and try to figure out what makes sense next, just like your friend trying to remember their favorite stories. But sometimes, they don’t have all the pieces of the puzzle, maybe they're working from a memory that's not perfect or they’re putting together parts from different stories.

`

Think about playing with building blocks. You know how each block has shapes and colors, and you stack them to make towers or houses. A large language model is like someone stacking blocks, they pick the right shape most of the time, but once in a while, they might use a block that doesn’t quite match, making their tower wobble. That’s when they "hallucinate", adding something that seems right but isn't actually there.