Glossary / AI
Hallucination
When an AI generates plausible-sounding content that is not grounded in fact or contradicts its source.
The term has long been used in language generation. A 2023 survey by Ziwei Ji and colleagues divides it into content that contradicts the source and content that cannot be verified from it[1].
Because a large language model picks a plausible next word, it can answer fluently about things it does not know. Interpretability research has found clues that models have a default "I don't know" state, and that a mechanism suppressing it sometimes misfires.
Related terms
- Large language modelAn AI trained on vast amounts of text that generates language by predicting the next word. The core of ChatGPT, Claude and similar systems.
- Mechanistic interpretabilityThe research field that tries to read an AI's internal computation part by part and explain, as a mechanism, why it produced an answer.