All terms
Glossary
Hallucination
An answer that sounds plausible and is factually wrong. The model does not invent deliberately, it strings likely words together.
Language models predict the next token. Whether the result is true is not a category in that process. That is why wrong answers sound exactly as convincing as right ones.
It happens most where training data is thin: niche products, internal names, recent events, exact figures and citations.
How you notice it
- Answers sound confident but are wrong.
- Sources or figures are cited that do not exist.
- Departments lose trust in the system.
Frequently asked
Can this be eliminated entirely?
No, but it can be contained. Ground answers in supplied documents, return citations, and allow the model to say it does not know. Those three measures matter more in production than any fine tuning.