Chapter 7 · Large language models
Hallucination
LLMs can produce fluent, confident text that is false, because they predict plausible words, not checked facts.
Chapter 7 · Large language models
LLMs can produce fluent, confident text that is false, because they predict plausible words, not checked facts.
The model is asked about a book. This book does not exist.
The model does not look facts up. It writes the words that fit best, and a made-up detail can fit as well as a true one.
You know how an LLM works: it predicts the next likely token. Likely is not the same as true. There is no fact-checker inside.
Tap the lines you think are invented or wrong, then press Check. Some answers have nothing wrong in them, so do not tap everything. All five answers are written for this lesson.
Who wrote the novel “The Seventeen Teacups of Wimbleford”?
This book does not exist. We made it up for the lesson.
Tap lines, then press Check.
Why does a model sometimes state false things?
A model sounds very sure of an answer. What does that tell you?
Which habit helps most with important facts?